Skip to content

monitor

Author: Heli Qi Affiliation: NAIST Date: 2022.07

Monitor

Bases: ABC

Abstract Base Class (ABC) for a Monitor that keeps track of various aspects of training and evaluatio.

Source code in speechain/monitor.py 
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
class Monitor(ABC):
    """Abstract Base Class (ABC) for a Monitor that keeps track of various aspects of
    training and evaluatio."""

    def __init__(
        self, logger, args: argparse.Namespace, result_path: str = None, **kwargs
    ) -> None:
        """Initializes the monitor class.

        Args:
            logger: A logger object for logging messages.
            args: Command line arguments passed as a Namespace object.
            result_path: Path to the directory where results should be saved.
            **kwargs: Additional keyword arguments.
        """  # Initialize shared members
        self.logger = logger
        self.result_path = (
            args.train_result_path
            if result_path is None
            else parse_path_args(result_path)
        )
        self.gpus = args.gpus if isinstance(args.gpus, List) else [args.gpus]

        # Initialize shared record information
        self.epoch_records = dict(
            consumed_time=dict(data_load_time=[], model_forward_time=[]),
            consumed_memory=dict(),
            criteria=dict(),
        )
        # Add GPU ranks to consumed_memory if they don't exist
        for rank, gpu in enumerate(self.gpus):
            if f"Rank{rank}" not in self.epoch_records["consumed_memory"].keys():
                self.epoch_records["consumed_memory"][f"Rank{rank}"] = []

        # Initialize step-level records
        self.step_records = dict(
            consumed_time=dict(data_load_time=[], model_forward_time=[]),
            criteria=dict(),
        )
        self.mode = None
        self.monitor_init(args, **kwargs)

        # Initialize snapshooter for log snapshotting
        self.logs_queue = Queue()
        snapshot_conf = dict(
            result_path=self.result_path,
            snap_mode=self.mode,
            **args.monitor_snapshot_conf,
        )
        # Initialize multiprocessing event to enable communication with the snapshooter process
        self.event = Event()
        self.event.clear()
        Process(
            target=snapshot_logs,
            args=(self.logs_queue, self.event, snapshot_conf),
            daemon=True,
        ).start()

    def enqueue(self, logs: Dict or List[Dict]) -> None:
        """Enqueues logs into the logs_queue.

        Args:
            logs:
                A dictionary or list of dictionaries containing logs.
        """
        if isinstance(logs, Dict):
            self.logs_queue.put(logs)
        elif isinstance(logs, List):
            for log in logs:
                self.logs_queue.put(log)
        else:
            raise RuntimeError("Expected logs to be a Dict or a List[Dict].")

    def empty_queue(self) -> None:
        """Checks if the logs queue is empty.

        Returns:
            Boolean value indicating whether the logs queue is empty or not.
        """
        self.logs_queue.qsize()
        return self.logs_queue.empty()

    @contextmanager
    def measure_time(self, names: str or List[str]) -> None:
        """Context manager for measuring time of execution.

        Args:
            names (str or List[str]):
                Name or list of names for the operations to be timed.
        """
        start = time.perf_counter()
        yield
        t = time.perf_counter() - start

        names = (
            ["consumed_time", names]
            if isinstance(names, str)
            else ["consumed_time"] + names
        )
        dict_pointer = self.step_records
        for i, name in enumerate(names):
            if name not in dict_pointer.keys():
                dict_pointer[name] = [] if i == len(names) - 1 else dict()
            dict_pointer = dict_pointer[name]
        dict_pointer.append(t)

    def refresh_step_records(self, records: Dict = None) -> None:
        """Refreshes the step records by resetting the values.

        Args:
            records (Dict, optional):
                A dictionary containing the records to be refreshed.
        """
        if records is None:
            records = self.step_records
        if isinstance(records, Dict):
            for key in records.keys():
                if isinstance(records[key], Dict):
                    self.refresh_step_records(records[key])
                elif isinstance(records[key], List):
                    records[key] = []
                else:
                    raise RuntimeError(
                        f"Unexpected type in records: {type(records[key])}."
                    )
        else:
            raise RuntimeError("Expected records to be of type Dict.")

    def record_step_info(self, key: str, step_info: Dict) -> None:
        """Records information at each step during training or evaluation.

        Args:
            key (str):
                Key under which information should be recorded.
            step_info (Dict):
                Dictionary containing the information to be recorded.
        """
        for name, info in step_info.items():
            if name not in self.step_records[key].keys():
                self.step_records[key][name] = []
            # result is in the form of torch.Tensor, so it needs to be transformed by .item()
            if isinstance(info, (torch.Tensor, np.ndarray)):
                if len(info.shape) == 1:
                    info = info[0]
                info = info.item()
            elif isinstance(info, List):
                info = info[0]
            self.step_records[key][name].append(info)

    def record_consumed_time(self, epoch_message: str) -> str:
        """Records time consumed in each epoch during training or evaluation.

        Args:
            epoch_message (str):
                String to be included in the epoch message.

        Returns:
            The updated epoch message.
        """
        epoch_message += " -- Consumed Time -- \n"

        # Record data loading time
        _total_time = sum(self.step_records["consumed_time"]["data_load_time"])
        epoch_message += f"Total data load time: {_total_time:.2f}s -- "
        self.epoch_records["consumed_time"]["data_load_time"].append(_total_time)

        # Record model forward time
        _total_time = sum(self.step_records["consumed_time"]["model_forward_time"])
        epoch_message += f"Total model forward time: {_total_time:.2f}s -- "
        self.epoch_records["consumed_time"]["model_forward_time"].append(_total_time)
        epoch_message += "\n"

        return epoch_message

    def record_consumed_memory(self, epoch_message: str) -> str:
        """Records memory consumed in each epoch during training or evaluation.

        Args:
            epoch_message (str):
                String to be included in the epoch message.

        Returns:
            The updated epoch message.
        """
        epoch_message += " -- Consumed Memory -- \n"
        gpus = GPUtil.getGPUs()
        if len(gpus) == 0:
            self.logger.warn(
                f"GPUtil.getGPUs() returns nothing at the {self.mode} part of epoch no.{self.epoch}. "
            )

        # Record consumed memory for each GPU
        for rank, gpu in enumerate(self.gpus):
            # recover the GPU number from 'CUDA_VISIBLE_DEVICES'
            if "CUDA_VISIBLE_DEVICES" in os.environ.keys():
                gpu = os.environ["CUDA_VISIBLE_DEVICES"].split(",")[gpu]
                # if gpu is in the form of the local ID, it will be converted into an integer for List indexing
                if gpu.isdigit():
                    gpu = int(gpu)
                # if gpu is in the form of the UUID, it will be kept as a string

            # --- torch.cuda is only able to report the GPU used in the current rank --- #
            # --- but torch.cuda can report the precise allocated and reserved memory information of the model --- #
            # turn bytes into MB，
            # memory_allocated = int(torch.cuda.memory_allocated(gpu) * (10 ** (-6)))
            # memory_reserved = int(torch.cuda.memory_reserved(gpu) * (10 ** (-6)))
            # epoch_message += f"GPU rank no.{rank} (cuda:{gpu}): " \
            #                  f"allocated memory {memory_allocated} MB, " \
            #                  f"reserved memory {memory_reserved} MB -- "
            # self.epoch_records['consumed_memory'][f'Rank{rank}']['memory_allocated'].append(memory_allocated)
            # self.epoch_records['consumed_memory'][f'Rank{rank}']['memory_reserved'].append(memory_reserved)

            # --- GPUtil can load the information of all the GPUs no matter the rank of the current process --- #
            # --- but it sometimes fails to report anything because of some IO errors --- #
            # --- and the returned used memory is the overall memory consumption of the GPU, --- #
            # --- which means that if there are more than one jobs running on the same GPU, --- #
            # --- the used memory is not precise for the current job. --- #
            memory_used = 0
            # for local GPU ID
            if isinstance(gpu, int) and gpu < len(gpus):
                memory_used = gpus[gpu].memoryUsed
                gpu_name = f"cuda:{gpu}"
            # for GPU UUID
            elif isinstance(gpu, str):
                curr_gpu = None
                for g in gpus:
                    if g.uuid == gpu:
                        curr_gpu = g
                if curr_gpu is not None:
                    memory_used = curr_gpu.memoryUsed
                gpu_name = gpu
            else:
                raise RuntimeError(
                    f"Unexpected errors happen when retrieving GPU IDs. (got {gpu})"
                )
            # if some unexpected errors happen above, memory_used will remain 0
            epoch_message += f"GPU rank no.{rank} ({gpu_name}): {memory_used} MB -- "

            if f"Rank{rank}" not in self.epoch_records["consumed_memory"].keys():
                self.epoch_records["consumed_memory"][f"Rank{rank}"] = []
            self.epoch_records["consumed_memory"][f"Rank{rank}"].append(memory_used)
        epoch_message += "\n"

        return epoch_message

    def record_criteria(self, epoch_message: str) -> str:
        """Records criteria in each epoch during training or evaluation.

        Args:
            epoch_message (str):
                String to be included in the epoch message.

        Returns:
            The updated epoch message.
        """
        epoch_message += " -- Criteria information -- \n"

        # Loop through all training criteria and record average and standard deviation
        for name, results in self.step_records["criteria"].items():
            if name not in self.epoch_records["criteria"].keys():
                self.epoch_records["criteria"][name] = []

            # calculate the average criterion value
            aver_result = np.mean(results).item()
            std_result = np.std(results).item()
            epoch_message += f"Average {name}: {aver_result:.2e} ± {std_result:.2f}\n"
            # record the average criterion value
            self.epoch_records["criteria"][name].append(aver_result)
        epoch_message += "\n"

        return epoch_message

    @abstractmethod
    def monitor_init(self, args: argparse.Namespace, **kwargs) -> None:
        """Abstract method for initializing the monitor.

        Args:
            args (str):
                Command line arguments passed as a Namespace object.
            kwargs:
                Additional keyword arguments.
        """
        raise NotImplementedError

    @abstractmethod
    def start_epoch(self, **kwargs) -> None:
        """Abstract method to be called at the start of each epoch."""
        raise NotImplementedError

    @abstractmethod
    def step(self, **kwargs) -> None:
        """Abstract method to be called at each step in an epoch."""
        raise NotImplementedError

    @abstractmethod
    def finish_epoch(self, **kwargs):
        """Abstract method to be called at the end of each epoch."""
        raise NotImplementedError

    @abstractmethod
    def state_dict(self) -> Dict:
        """This method currently performs no operation and always returns None.

        The intention is for subclasses to override this method to return a dictionary
        containing the state of the Monitor. However, in the base Monitor class,
        it does not have any state to save, so it returns None.

        Returns:
            None
        """
        return None

    def load_state_dict(self, state_dict: Dict = None) -> None:
        """Loads the Monitor state.

        Args:
            state_dict (Dict):
                Dictionary containing a whole state of the Monitor.
        """
        if state_dict is not None:
            for key, value in state_dict.items():
                self.__setattr__(key, value)

__init__(logger, args, result_path=None, **kwargs)

Initializes the monitor class.

Parameters:

Name Type Description Default
logger

A logger object for logging messages.

 required
args Namespace

Command line arguments passed as a Namespace object.

 required
result_path str

Path to the directory where results should be saved.

 None
**kwargs

Additional keyword arguments.

 {}
Source code in speechain/monitor.py 
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
def __init__(
    self, logger, args: argparse.Namespace, result_path: str = None, **kwargs
) -> None:
    """Initializes the monitor class.

    Args:
        logger: A logger object for logging messages.
        args: Command line arguments passed as a Namespace object.
        result_path: Path to the directory where results should be saved.
        **kwargs: Additional keyword arguments.
    """  # Initialize shared members
    self.logger = logger
    self.result_path = (
        args.train_result_path
        if result_path is None
        else parse_path_args(result_path)
    )
    self.gpus = args.gpus if isinstance(args.gpus, List) else [args.gpus]

    # Initialize shared record information
    self.epoch_records = dict(
        consumed_time=dict(data_load_time=[], model_forward_time=[]),
        consumed_memory=dict(),
        criteria=dict(),
    )
    # Add GPU ranks to consumed_memory if they don't exist
    for rank, gpu in enumerate(self.gpus):
        if f"Rank{rank}" not in self.epoch_records["consumed_memory"].keys():
            self.epoch_records["consumed_memory"][f"Rank{rank}"] = []

    # Initialize step-level records
    self.step_records = dict(
        consumed_time=dict(data_load_time=[], model_forward_time=[]),
        criteria=dict(),
    )
    self.mode = None
    self.monitor_init(args, **kwargs)

    # Initialize snapshooter for log snapshotting
    self.logs_queue = Queue()
    snapshot_conf = dict(
        result_path=self.result_path,
        snap_mode=self.mode,
        **args.monitor_snapshot_conf,
    )
    # Initialize multiprocessing event to enable communication with the snapshooter process
    self.event = Event()
    self.event.clear()
    Process(
        target=snapshot_logs,
        args=(self.logs_queue, self.event, snapshot_conf),
        daemon=True,
    ).start()

empty_queue()

Checks if the logs queue is empty.

Returns:

Type Description
None

Boolean value indicating whether the logs queue is empty or not.
Source code in speechain/monitor.py 
117
118
119
120
121
122
123
124
def empty_queue(self) -> None:
    """Checks if the logs queue is empty.

    Returns:
        Boolean value indicating whether the logs queue is empty or not.
    """
    self.logs_queue.qsize()
    return self.logs_queue.empty()

enqueue(logs)

Enqueues logs into the logs_queue.

Parameters:

Name Type Description Default
logs Dict or List[Dict]

A dictionary or list of dictionaries containing logs.

 required
Source code in speechain/monitor.py 
102
103
104
105
106
107
108
109
110
111
112
113
114
115
def enqueue(self, logs: Dict or List[Dict]) -> None:
    """Enqueues logs into the logs_queue.

    Args:
        logs:
            A dictionary or list of dictionaries containing logs.
    """
    if isinstance(logs, Dict):
        self.logs_queue.put(logs)
    elif isinstance(logs, List):
        for log in logs:
            self.logs_queue.put(log)
    else:
        raise RuntimeError("Expected logs to be a Dict or a List[Dict].")

finish_epoch(**kwargs) abstractmethod

Abstract method to be called at the end of each epoch.

Source code in speechain/monitor.py 
338
339
340
341
@abstractmethod
def finish_epoch(self, **kwargs):
    """Abstract method to be called at the end of each epoch."""
    raise NotImplementedError

load_state_dict(state_dict=None)

Loads the Monitor state.

Parameters:

Name Type Description Default
state_dict Dict

Dictionary containing a whole state of the Monitor.

 None
Source code in speechain/monitor.py 
356
357
358
359
360
361
362
363
364
365
def load_state_dict(self, state_dict: Dict = None) -> None:
    """Loads the Monitor state.

    Args:
        state_dict (Dict):
            Dictionary containing a whole state of the Monitor.
    """
    if state_dict is not None:
        for key, value in state_dict.items():
            self.__setattr__(key, value)

measure_time(names)

Context manager for measuring time of execution.

Parameters:

Name Type Description Default
names str or List[str]

Name or list of names for the operations to be timed.

 required
Source code in speechain/monitor.py 
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
@contextmanager
def measure_time(self, names: str or List[str]) -> None:
    """Context manager for measuring time of execution.

    Args:
        names (str or List[str]):
            Name or list of names for the operations to be timed.
    """
    start = time.perf_counter()
    yield
    t = time.perf_counter() - start

    names = (
        ["consumed_time", names]
        if isinstance(names, str)
        else ["consumed_time"] + names
    )
    dict_pointer = self.step_records
    for i, name in enumerate(names):
        if name not in dict_pointer.keys():
            dict_pointer[name] = [] if i == len(names) - 1 else dict()
        dict_pointer = dict_pointer[name]
    dict_pointer.append(t)

monitor_init(args, **kwargs) abstractmethod

Abstract method for initializing the monitor.

Parameters:

Name Type Description Default
args str

Command line arguments passed as a Namespace object.

 required
kwargs

Additional keyword arguments.

 {}
Source code in speechain/monitor.py 
316
317
318
319
320
321
322
323
324
325
326
@abstractmethod
def monitor_init(self, args: argparse.Namespace, **kwargs) -> None:
    """Abstract method for initializing the monitor.

    Args:
        args (str):
            Command line arguments passed as a Namespace object.
        kwargs:
            Additional keyword arguments.
    """
    raise NotImplementedError

record_consumed_memory(epoch_message)

Records memory consumed in each epoch during training or evaluation.

Parameters:

Name Type Description Default
epoch_message str

String to be included in the epoch message.

 required

Returns:

Type Description
str

The updated epoch message.
Source code in speechain/monitor.py 
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
def record_consumed_memory(self, epoch_message: str) -> str:
    """Records memory consumed in each epoch during training or evaluation.

    Args:
        epoch_message (str):
            String to be included in the epoch message.

    Returns:
        The updated epoch message.
    """
    epoch_message += " -- Consumed Memory -- \n"
    gpus = GPUtil.getGPUs()
    if len(gpus) == 0:
        self.logger.warn(
            f"GPUtil.getGPUs() returns nothing at the {self.mode} part of epoch no.{self.epoch}. "
        )

    # Record consumed memory for each GPU
    for rank, gpu in enumerate(self.gpus):
        # recover the GPU number from 'CUDA_VISIBLE_DEVICES'
        if "CUDA_VISIBLE_DEVICES" in os.environ.keys():
            gpu = os.environ["CUDA_VISIBLE_DEVICES"].split(",")[gpu]
            # if gpu is in the form of the local ID, it will be converted into an integer for List indexing
            if gpu.isdigit():
                gpu = int(gpu)
            # if gpu is in the form of the UUID, it will be kept as a string

        # --- torch.cuda is only able to report the GPU used in the current rank --- #
        # --- but torch.cuda can report the precise allocated and reserved memory information of the model --- #
        # turn bytes into MB，
        # memory_allocated = int(torch.cuda.memory_allocated(gpu) * (10 ** (-6)))
        # memory_reserved = int(torch.cuda.memory_reserved(gpu) * (10 ** (-6)))
        # epoch_message += f"GPU rank no.{rank} (cuda:{gpu}): " \
        #                  f"allocated memory {memory_allocated} MB, " \
        #                  f"reserved memory {memory_reserved} MB -- "
        # self.epoch_records['consumed_memory'][f'Rank{rank}']['memory_allocated'].append(memory_allocated)
        # self.epoch_records['consumed_memory'][f'Rank{rank}']['memory_reserved'].append(memory_reserved)

        # --- GPUtil can load the information of all the GPUs no matter the rank of the current process --- #
        # --- but it sometimes fails to report anything because of some IO errors --- #
        # --- and the returned used memory is the overall memory consumption of the GPU, --- #
        # --- which means that if there are more than one jobs running on the same GPU, --- #
        # --- the used memory is not precise for the current job. --- #
        memory_used = 0
        # for local GPU ID
        if isinstance(gpu, int) and gpu < len(gpus):
            memory_used = gpus[gpu].memoryUsed
            gpu_name = f"cuda:{gpu}"
        # for GPU UUID
        elif isinstance(gpu, str):
            curr_gpu = None
            for g in gpus:
                if g.uuid == gpu:
                    curr_gpu = g
            if curr_gpu is not None:
                memory_used = curr_gpu.memoryUsed
            gpu_name = gpu
        else:
            raise RuntimeError(
                f"Unexpected errors happen when retrieving GPU IDs. (got {gpu})"
            )
        # if some unexpected errors happen above, memory_used will remain 0
        epoch_message += f"GPU rank no.{rank} ({gpu_name}): {memory_used} MB -- "

        if f"Rank{rank}" not in self.epoch_records["consumed_memory"].keys():
            self.epoch_records["consumed_memory"][f"Rank{rank}"] = []
        self.epoch_records["consumed_memory"][f"Rank{rank}"].append(memory_used)
    epoch_message += "\n"

    return epoch_message

record_consumed_time(epoch_message)

Records time consumed in each epoch during training or evaluation.

Parameters:

Name Type Description Default
epoch_message str

String to be included in the epoch message.

 required

Returns:

Type Description
str

The updated epoch message.
Source code in speechain/monitor.py 
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
def record_consumed_time(self, epoch_message: str) -> str:
    """Records time consumed in each epoch during training or evaluation.

    Args:
        epoch_message (str):
            String to be included in the epoch message.

    Returns:
        The updated epoch message.
    """
    epoch_message += " -- Consumed Time -- \n"

    # Record data loading time
    _total_time = sum(self.step_records["consumed_time"]["data_load_time"])
    epoch_message += f"Total data load time: {_total_time:.2f}s -- "
    self.epoch_records["consumed_time"]["data_load_time"].append(_total_time)

    # Record model forward time
    _total_time = sum(self.step_records["consumed_time"]["model_forward_time"])
    epoch_message += f"Total model forward time: {_total_time:.2f}s -- "
    self.epoch_records["consumed_time"]["model_forward_time"].append(_total_time)
    epoch_message += "\n"

    return epoch_message

record_criteria(epoch_message)

Records criteria in each epoch during training or evaluation.

Parameters:

Name Type Description Default
epoch_message str

String to be included in the epoch message.

 required

Returns:

Type Description
str

The updated epoch message.
Source code in speechain/monitor.py 
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
def record_criteria(self, epoch_message: str) -> str:
    """Records criteria in each epoch during training or evaluation.

    Args:
        epoch_message (str):
            String to be included in the epoch message.

    Returns:
        The updated epoch message.
    """
    epoch_message += " -- Criteria information -- \n"

    # Loop through all training criteria and record average and standard deviation
    for name, results in self.step_records["criteria"].items():
        if name not in self.epoch_records["criteria"].keys():
            self.epoch_records["criteria"][name] = []

        # calculate the average criterion value
        aver_result = np.mean(results).item()
        std_result = np.std(results).item()
        epoch_message += f"Average {name}: {aver_result:.2e} ± {std_result:.2f}\n"
        # record the average criterion value
        self.epoch_records["criteria"][name].append(aver_result)
    epoch_message += "\n"

    return epoch_message

record_step_info(key, step_info)

Records information at each step during training or evaluation.

Parameters:

Name Type Description Default
key str

Key under which information should be recorded.

 required
step_info Dict

Dictionary containing the information to be recorded.

 required
Source code in speechain/monitor.py 
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
def record_step_info(self, key: str, step_info: Dict) -> None:
    """Records information at each step during training or evaluation.

    Args:
        key (str):
            Key under which information should be recorded.
        step_info (Dict):
            Dictionary containing the information to be recorded.
    """
    for name, info in step_info.items():
        if name not in self.step_records[key].keys():
            self.step_records[key][name] = []
        # result is in the form of torch.Tensor, so it needs to be transformed by .item()
        if isinstance(info, (torch.Tensor, np.ndarray)):
            if len(info.shape) == 1:
                info = info[0]
            info = info.item()
        elif isinstance(info, List):
            info = info[0]
        self.step_records[key][name].append(info)

refresh_step_records(records=None)

Refreshes the step records by resetting the values.

Parameters:

Name Type Description Default
records Dict

A dictionary containing the records to be refreshed.

 None
Source code in speechain/monitor.py 
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
def refresh_step_records(self, records: Dict = None) -> None:
    """Refreshes the step records by resetting the values.

    Args:
        records (Dict, optional):
            A dictionary containing the records to be refreshed.
    """
    if records is None:
        records = self.step_records
    if isinstance(records, Dict):
        for key in records.keys():
            if isinstance(records[key], Dict):
                self.refresh_step_records(records[key])
            elif isinstance(records[key], List):
                records[key] = []
            else:
                raise RuntimeError(
                    f"Unexpected type in records: {type(records[key])}."
                )
    else:
        raise RuntimeError("Expected records to be of type Dict.")

start_epoch(**kwargs) abstractmethod

Abstract method to be called at the start of each epoch.

Source code in speechain/monitor.py 
328
329
330
331
@abstractmethod
def start_epoch(self, **kwargs) -> None:
    """Abstract method to be called at the start of each epoch."""
    raise NotImplementedError

state_dict() abstractmethod

This method currently performs no operation and always returns None.

The intention is for subclasses to override this method to return a dictionary containing the state of the Monitor. However, in the base Monitor class, it does not have any state to save, so it returns None.

Returns:

Type Description
Dict

None
Source code in speechain/monitor.py 
343
344
345
346
347
348
349
350
351
352
353
354
@abstractmethod
def state_dict(self) -> Dict:
    """This method currently performs no operation and always returns None.

    The intention is for subclasses to override this method to return a dictionary
    containing the state of the Monitor. However, in the base Monitor class,
    it does not have any state to save, so it returns None.

    Returns:
        None
    """
    return None

step(**kwargs) abstractmethod

Abstract method to be called at each step in an epoch.

Source code in speechain/monitor.py 
333
334
335
336
@abstractmethod
def step(self, **kwargs) -> None:
    """Abstract method to be called at each step in an epoch."""
    raise NotImplementedError

TestMonitor

Bases: Monitor

Class responsible for monitoring the testing process and logging real- time information. It extends the base Monitor class.

Attributes:

Name Type Description
distributed bool

If True, indicates distributed training.
report_per_steps int

Frequency of reporting in steps.
bad_cases_selection List[List]

Criteria for selecting bad cases, if any.
data_saving_logs_queue Queue

Queue object for storing logs to be saved.
prev_test_time float

Time of the previous test, used to calculate elapsed time.
total_step_num int

Total number of steps in the test.
step_info dict

Information about the current test step.
finished_group_num int

Number of groups that have finished testing.
Source code in speechain/monitor.py 
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
class TestMonitor(Monitor):
    """Class responsible for monitoring the testing process and logging real- time
    information. It extends the base Monitor class.

    Attributes:
        distributed (bool):
            If True, indicates distributed training.
        report_per_steps (int):
            Frequency of reporting in steps.
        bad_cases_selection (List[List]):
            Criteria for selecting bad cases, if any.
        data_saving_logs_queue (Queue):
            Queue object for storing logs to be saved.
        prev_test_time (float):
            Time of the previous test, used to calculate elapsed time.
        total_step_num (int):
            Total number of steps in the test.
        step_info (dict):
            Information about the current test step.
        finished_group_num (int):
            Number of groups that have finished testing.
    """

    def monitor_init(self, args: argparse.Namespace, **kwargs):
        """Initializes the TestMonitor with the given arguments.

        Args:
            args (argparse.Namespace):
                Parsed command line arguments.
        """
        self.distributed = args.distributed
        self.report_per_steps = args.report_per_steps
        self.bad_cases_selection = args.bad_cases_selection
        if self.bad_cases_selection is None:
            self.bad_cases_selection = []
        elif not isinstance(self.bad_cases_selection[0], List):
            self.bad_cases_selection = [self.bad_cases_selection]

        # initialize the snapshooter of the monitor
        self.data_saving_logs_queue = Queue()
        # create daemon processes for data saving
        assert (
            args.saving_proc_num >= 1
        ), "'saving_proc_num' should be an integer larger than 1!"
        self.saving_proc_num = args.saving_proc_num
        for proc_id in range(self.saving_proc_num):
            Process(
                target=data_saving_logs,
                args=(proc_id, self.data_saving_logs_queue),
                daemon=True,
            ).start()

    def start_epoch(self, total_step_num: int):
        """Starts a new testing epoch.

        Args:
            total_step_num (int): Total number of steps in the epoch.
        """
        # para init
        self.prev_test_time = time.time()
        self.total_step_num = total_step_num

        if not hasattr(self, "step_info"):
            self.step_info = dict(group_time=[], total_time=0)
        if not hasattr(self, "finished_group_num"):
            self.finished_group_num = 0

    def step(self, step_num: int, test_results: Dict[str, Dict], test_index: List[str]):
        """Executes a single test step and logs the results.

        Args:
            step_num (int): Current step number.
            test_results (Dict[str, Dict]): Results of the test step.
            test_index (List[str]): Indexes of the tests.
        """
        # --- Write the testing results of the current step to the testing files --- #
        # loop each result list in the returned Dict
        for name, result in test_results.items():
            # for .txt file, register the result contents into self.step_info. text data doesn't occupy too much memory
            if result["format"].lower() == "txt":
                if name not in self.step_info.keys():
                    self.step_info[name] = dict()

                if not isinstance(result["content"], List):
                    result["content"] = [result["content"]]

                for index, content in zip(test_index, result["content"]):
                    # for the List-type element, turn it into its string format
                    if isinstance(content, List):
                        content = str(content)
                    self.step_info[name][index] = content

            # for other files, data needs to be saved to the disk in real time to reduce the memory burden
            else:
                # we enqueue a whole batch of data to be saved in order to reduce the number of query operations of each daemon process
                self.data_saving_logs_queue.put(
                    dict(
                        file_format=result["format"].lower(),
                        save_path=os.path.join(self.result_path, name),
                        file_name_list=test_index,
                        file_content_list=result["content"],
                        group_ids=(
                            result["group_ids"]
                            if "group_ids" in result.keys()
                            else None
                        ),
                        sample_rate=(
                            result["sample_rate"]
                            if "sample_rate" in result.keys()
                            else None
                        ),
                    )
                )

        # monitor the approximate size of the queue to be more memory-friendly
        while self.data_saving_logs_queue.qsize() > 5 * self.saving_proc_num:
            self.logger.warning(
                f"There has been more than {3 * self.saving_proc_num} batches in data_saving_logs_queue, "
                f"so the data generation is paused for 30 seconds."
            )
            time.sleep(30)

        # --- Report the testing midway information to users --- #
        test_step_message = None
        # record the tesing time of the current step
        curr_test_time = time.time()
        self.step_info["group_time"].append(curr_test_time - self.prev_test_time)
        self.prev_test_time = curr_test_time

        # meet the reporting interval
        if step_num % self.report_per_steps == 0:
            curr_group_time = sum(self.step_info["group_time"])

            # the first testing step
            if self.finished_group_num == 0:
                prev_group_time = curr_group_time
            # other testing steps
            else:
                # calculate the average time of all the previous groups
                prev_group_time = self.step_info["total_time"] / self.finished_group_num

            # calculate the number of remaining steps
            self.finished_group_num += 1
            finish_step_num = int(self.finished_group_num * self.report_per_steps)
            remaining_step_num = self.total_step_num - finish_step_num
            # take the weighted average consuming time of each group
            aver_group_time = (prev_group_time + curr_group_time) / 2
            remaining_time = aver_group_time * (
                remaining_step_num / self.report_per_steps
            )

            # update the time records
            self.step_info["total_time"] += curr_group_time
            self.step_info["group_time"] = []

            test_step_message = (
                f"Testing Midway Report -- "
                f"testing time for the recent {self.report_per_steps} steps: {curr_group_time:.2f}s -- "
                f"finished step number: {finish_step_num} -- "
                f"remaining step number: {remaining_step_num} -- "
                f"expected remaining time: "
            )

            remaining_days, remaining_time = int(
                remaining_time // (3600 * 24)
            ), remaining_time % (3600 * 24)
            if remaining_days > 0:
                test_step_message += f"{remaining_days:d}d "

            remaining_hours, remaining_time = (
                int(remaining_time // 3600),
                remaining_time % 3600,
            )
            if remaining_hours > 0:
                test_step_message += f"{remaining_hours:d}h "

            remaining_minutes, remaining_time = (
                int(remaining_time // 60),
                remaining_time % 60,
            )
            if remaining_minutes > 0:
                test_step_message += f"{remaining_minutes:d}m "

            remaining_seconds = remaining_time
            test_step_message += f"{remaining_seconds:.2f}s"

        if test_step_message is not None:
            self.logger.info(test_step_message)

    def wait_empty_queues(self, sleep_time: int = 60):
        """Waits until the log queue is empty before continuing.

        Args:
            sleep_time (int, optional): Time to wait in seconds when the queue is not empty. Defaults to 60.
        """
        while True:
            if self.data_saving_logs_queue.empty():
                # wait for one more time when the queue becomes empty to left enough time for data saving
                time.sleep(sleep_time)
                break
            else:
                self.logger.info(
                    f"The data saving process is still working. Waiting for {sleep_time} seconds."
                )
                time.sleep(sleep_time)

        # If using distributed training, synchronize all processes before continuing
        if self.distributed:
            torch.distributed.barrier()

    def finish_epoch(self, meta_info: Dict = None):
        """This method completes various tasks at the end of each epoch, such
        as:

        - Gathering checkpoint information from all GPU processes
        - Generating data path files
        - Producing evaluation reports for overall and group-level performance
        - Presenting top-N bad cases (if 'instance_reports.md' is in self.step_info)
        - Plotting histograms for numerical metrics in step_info

        Arguments:
            meta_info (Dict, optional):
                Meta information about testing samples. Used for group-level evaluation.

        Note:
            The method is designed for a distributed setting, where results from different processes need to be gathered.
            If the program is not running in a distributed setting, some steps (like gathering checkpoint information) will be skipped.
        """
        # group all the testing samples by their meta info
        group_meta_info = None
        if meta_info is not None:
            group_meta_info = dict()
            # loop each type of meta data
            for meta_type, meta_dict in meta_info.items():
                if meta_type not in group_meta_info.keys():
                    group_meta_info[meta_type] = dict()

                # loop each group of samples
                for index, group in meta_info[meta_type].items():
                    if group not in group_meta_info[meta_type].keys():
                        group_meta_info[meta_type][group] = []
                    group_meta_info[meta_type][group].append(index)

        # --- Gather the checkpoint information of all the processes --- #
        # load the checkpoint of rank0 and delete testing time information
        self.load_state_dict(
            torch.load(os.path.join(self.result_path, "rank0_tmp", "checkpoint.pth"))[
                "monitor"
            ]
        )
        self.step_info.pop("group_time")
        self.step_info.pop("total_time")

        if self.distributed:
            for rank in range(1, torch.distributed.get_world_size()):
                _tmp_dict = torch.load(
                    os.path.join(self.result_path, f"rank{rank}_tmp", "checkpoint.pth")
                )["monitor"]["step_info"]
                for key in self.step_info.keys():
                    self.step_info[key].update(_tmp_dict[key])

        # make sure that all the samples are sorted by their indices
        for key in self.step_info.keys():
            self.step_info[key] = dict(
                sorted(self.step_info[key].items(), key=lambda x: x[0])
            )
            # .md files remain their original names
            if key.endswith(".md"):
                np.savetxt(
                    os.path.join(self.result_path, key),
                    list(self.step_info[key].items()),
                    fmt="%s",
                )
            # normal .txt files have the prefix 'idx2' attached at the beginning of their names
            else:
                np.savetxt(
                    os.path.join(self.result_path, f"idx2{key}"),
                    list(self.step_info[key].items()),
                    fmt="%s",
                )

        # --- Gather all the save-during-testing files & Generate their path files --- #
        # generate the data path files
        for file_name in os.listdir(self.result_path):
            # only consider folders
            if not os.path.isdir(os.path.join(self.result_path, file_name)):
                continue
            # only consider the folders not named as 'figures' and 'rank_tmp'
            if (
                file_name.startswith("rank")
                or file_name == "figures"
                or "=" in file_name
            ):
                continue
            idx2path = {}
            for data_file in search_file_in_subfolder(
                os.path.join(self.result_path, file_name),
                tgt_match_fn=lambda x: len(x.split(".")) > 1,
            ):
                data_index = ".".join(os.path.basename(data_file).split(".")[:-1])

                # add new file into the Dict
                if data_index not in idx2path.keys():
                    idx2path[data_index] = data_file
                # for multiple files with the same index (probably because of the non-reproducible resuming issue)
                else:
                    # same the latest file
                    if get_file_birthtime(data_file) > get_file_birthtime(
                        idx2path[data_index]
                    ):
                        while os.path.exists(idx2path[data_index]):
                            os.remove(idx2path[data_index])
                        idx2path[data_index] = data_file

            idx2path = list(sorted(idx2path.items(), key=lambda x: x[0]))
            np.savetxt(
                os.path.join(self.result_path, f"idx2{file_name}"), idx2path, fmt="%s"
            )

        # --- Group-level Evaluation Report Production --- #
        result_path = os.path.join(self.result_path, "overall_results.md")
        result_string = ""

        # The overall evaluation performance
        result_string += "# Overall Evaluation (mean ± std):\n"
        content_dict = dict()
        # loop each metric and record the overall model performance
        for metric, result_dict in self.step_info.items():
            result_list = list(result_dict.values())
            # only average the numerical results
            if not isinstance(result_list[0], (int, float)):
                continue

            content_dict[metric] = (
                f"{np.mean(result_list):.4f} ± {np.std(result_list):.4f}"
            )
        result_string += get_list_strings(content_dict=content_dict)

        # record the group-level model performance
        if group_meta_info is not None:
            for meta_name, group_dict in group_meta_info.items():
                result_string += (
                    f"# {meta_name}-wise Evaluation:\n"
                    f"(***bold&italic*** numbers represent the maximal ones in all groups while "
                    f"**bold** numbers represent the minimal ones.)\n\n"
                )
                table_headers, table_contents = [meta_name], dict()
                # loop each group and calculate the group-specific performance
                for group_name, group_list in group_dict.items():
                    # loop each metric
                    for metric, result_dict in self.step_info.items():
                        result_list = [
                            result_dict[index]
                            for index in group_list
                            if index in result_dict.keys()
                        ]
                        # skip the non-numerical results
                        if len(result_list) > 0 and not isinstance(
                            result_list[0], (int, float)
                        ):
                            continue
                        # average the numerical results and record them
                        if len(result_list) != 0:
                            # create group item lazily
                            if group_name not in table_contents.keys():
                                table_contents[group_name] = []

                            if metric not in table_headers:
                                table_headers.append(metric)
                            table_contents[group_name].append(np.mean(result_list))

                # get the max and min group value for each numerical metric
                for i in range(len(table_headers) - 1):
                    metric_value_list = [value[i] for value in table_contents.values()]
                    max_value, min_value = max(metric_value_list), min(
                        metric_value_list
                    )
                    # loop each group
                    for group in table_contents.keys():
                        # turn the max number into a bold&italic string
                        if table_contents[group][i] == max_value:
                            table_contents[group][
                                i
                            ] = f"***{table_contents[group][i]:.4f}***"
                        # turn the min number into a bold string
                        elif table_contents[group][i] == min_value:
                            table_contents[group][
                                i
                            ] = f"**{table_contents[group][i]:.4f}**"
                        # turn other numbers into pure strings
                        else:
                            table_contents[group][i] = f"{table_contents[group][i]:.4f}"

                # attach the list of the current group into the result string
                result_string += get_table_strings(
                    contents=list(table_contents.values()),
                    first_col=list(table_contents.keys()),
                    headers=table_headers,
                )
        np.savetxt(result_path, [result_string], fmt="%s")

        # --- Top-N Bad Cases Presentation --- #
        # only present topn bad cases if instance_reports.md is given
        if "instance_reports.md" in self.step_info.keys():
            # loop each tri-tuple
            for metric, mode, num in self.bad_cases_selection:
                result_path = os.path.join(
                    self.result_path, f"top{num}_{mode}_{metric}.md"
                )

                if metric in self.step_info.keys():
                    # get the indices of the topn samples
                    selected_samples = sorted(
                        self.step_info[metric].items(),
                        key=lambda x: x[1],
                        reverse=True if mode.lower() == "max" else False,
                    )[:num]
                    selected_samples = [s[0] for s in selected_samples]

                    # make the .md string for all the top-n bad samples
                    sample_reports = ""
                    for s_index in selected_samples:
                        sample_reports += (
                            f"**{s_index}**"
                            + self.step_info["instance_reports.md"][s_index]
                        )
                    np.savetxt(result_path, [sample_reports], fmt="%s")

        # --- Histograms Plotting --- #
        # remove the old figures if have
        if os.path.exists(os.path.join(self.result_path, "figures")):
            shutil.rmtree(os.path.join(self.result_path, "figures"))

        # loop each metric and plot the histogram figure
        for metric, result_dict in self.step_info.items():
            result_list = list(result_dict.values())
            # only consider the numerical results
            if not isinstance(result_list[0], (int, float)):
                continue

            self.enqueue(
                dict(materials=copy.deepcopy({metric: result_list}), plot_type="hist")
            )

        if not self.empty_queue():
            for i in range(60):
                if not self.empty_queue():
                    self.logger.info(
                        "The snapshooter is still snapshotting. Waiting for 10 seconds."
                    )
                    time.sleep(10)

    def state_dict(self):
        """Returns the state of the TestMonitor instance. This is useful for
        checkpointing or logging purposes.

        Returns:
        - A dictionary that includes:
            * step_info: A dictionary storing information about each step.
            * finished_group_num: The number of groups that have finished processing.
        """
        return dict(
            step_info=self.step_info, finished_group_num=self.finished_group_num
        )

finish_epoch(meta_info=None)

This method completes various tasks at the end of each epoch, such as:

  • Gathering checkpoint information from all GPU processes
  • Generating data path files
  • Producing evaluation reports for overall and group-level performance
  • Presenting top-N bad cases (if 'instance_reports.md' is in self.step_info)
  • Plotting histograms for numerical metrics in step_info

Parameters:

Name Type Description Default
meta_info Dict

Meta information about testing samples. Used for group-level evaluation.

 None
Note

The method is designed for a distributed setting, where results from different processes need to be gathered. If the program is not running in a distributed setting, some steps (like gathering checkpoint information) will be skipped.

Source code in speechain/monitor.py 
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
1683
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
1749
1750
1751
1752
1753
1754
1755
1756
1757
1758
1759
1760
1761
1762
1763
1764
1765
1766
1767
1768
1769
1770
1771
1772
1773
1774
1775
1776
1777
1778
1779
1780
1781
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
1803
1804
1805
1806
1807
1808
1809
1810
1811
1812
1813
1814
1815
1816
1817
1818
1819
1820
1821
1822
1823
1824
1825
1826
1827
1828
1829
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
1840
1841
1842
1843
1844
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
1859
1860
1861
def finish_epoch(self, meta_info: Dict = None):
    """This method completes various tasks at the end of each epoch, such
    as:

    - Gathering checkpoint information from all GPU processes
    - Generating data path files
    - Producing evaluation reports for overall and group-level performance
    - Presenting top-N bad cases (if 'instance_reports.md' is in self.step_info)
    - Plotting histograms for numerical metrics in step_info

    Arguments:
        meta_info (Dict, optional):
            Meta information about testing samples. Used for group-level evaluation.

    Note:
        The method is designed for a distributed setting, where results from different processes need to be gathered.
        If the program is not running in a distributed setting, some steps (like gathering checkpoint information) will be skipped.
    """
    # group all the testing samples by their meta info
    group_meta_info = None
    if meta_info is not None:
        group_meta_info = dict()
        # loop each type of meta data
        for meta_type, meta_dict in meta_info.items():
            if meta_type not in group_meta_info.keys():
                group_meta_info[meta_type] = dict()

            # loop each group of samples
            for index, group in meta_info[meta_type].items():
                if group not in group_meta_info[meta_type].keys():
                    group_meta_info[meta_type][group] = []
                group_meta_info[meta_type][group].append(index)

    # --- Gather the checkpoint information of all the processes --- #
    # load the checkpoint of rank0 and delete testing time information
    self.load_state_dict(
        torch.load(os.path.join(self.result_path, "rank0_tmp", "checkpoint.pth"))[
            "monitor"
        ]
    )
    self.step_info.pop("group_time")
    self.step_info.pop("total_time")

    if self.distributed:
        for rank in range(1, torch.distributed.get_world_size()):
            _tmp_dict = torch.load(
                os.path.join(self.result_path, f"rank{rank}_tmp", "checkpoint.pth")
            )["monitor"]["step_info"]
            for key in self.step_info.keys():
                self.step_info[key].update(_tmp_dict[key])

    # make sure that all the samples are sorted by their indices
    for key in self.step_info.keys():
        self.step_info[key] = dict(
            sorted(self.step_info[key].items(), key=lambda x: x[0])
        )
        # .md files remain their original names
        if key.endswith(".md"):
            np.savetxt(
                os.path.join(self.result_path, key),
                list(self.step_info[key].items()),
                fmt="%s",
            )
        # normal .txt files have the prefix 'idx2' attached at the beginning of their names
        else:
            np.savetxt(
                os.path.join(self.result_path, f"idx2{key}"),
                list(self.step_info[key].items()),
                fmt="%s",
            )

    # --- Gather all the save-during-testing files & Generate their path files --- #
    # generate the data path files
    for file_name in os.listdir(self.result_path):
        # only consider folders
        if not os.path.isdir(os.path.join(self.result_path, file_name)):
            continue
        # only consider the folders not named as 'figures' and 'rank_tmp'
        if (
            file_name.startswith("rank")
            or file_name == "figures"
            or "=" in file_name
        ):
            continue
        idx2path = {}
        for data_file in search_file_in_subfolder(
            os.path.join(self.result_path, file_name),
            tgt_match_fn=lambda x: len(x.split(".")) > 1,
        ):
            data_index = ".".join(os.path.basename(data_file).split(".")[:-1])

            # add new file into the Dict
            if data_index not in idx2path.keys():
                idx2path[data_index] = data_file
            # for multiple files with the same index (probably because of the non-reproducible resuming issue)
            else:
                # same the latest file
                if get_file_birthtime(data_file) > get_file_birthtime(
                    idx2path[data_index]
                ):
                    while os.path.exists(idx2path[data_index]):
                        os.remove(idx2path[data_index])
                    idx2path[data_index] = data_file

        idx2path = list(sorted(idx2path.items(), key=lambda x: x[0]))
        np.savetxt(
            os.path.join(self.result_path, f"idx2{file_name}"), idx2path, fmt="%s"
        )

    # --- Group-level Evaluation Report Production --- #
    result_path = os.path.join(self.result_path, "overall_results.md")
    result_string = ""

    # The overall evaluation performance
    result_string += "# Overall Evaluation (mean ± std):\n"
    content_dict = dict()
    # loop each metric and record the overall model performance
    for metric, result_dict in self.step_info.items():
        result_list = list(result_dict.values())
        # only average the numerical results
        if not isinstance(result_list[0], (int, float)):
            continue

        content_dict[metric] = (
            f"{np.mean(result_list):.4f} ± {np.std(result_list):.4f}"
        )
    result_string += get_list_strings(content_dict=content_dict)

    # record the group-level model performance
    if group_meta_info is not None:
        for meta_name, group_dict in group_meta_info.items():
            result_string += (
                f"# {meta_name}-wise Evaluation:\n"
                f"(***bold&italic*** numbers represent the maximal ones in all groups while "
                f"**bold** numbers represent the minimal ones.)\n\n"
            )
            table_headers, table_contents = [meta_name], dict()
            # loop each group and calculate the group-specific performance
            for group_name, group_list in group_dict.items():
                # loop each metric
                for metric, result_dict in self.step_info.items():
                    result_list = [
                        result_dict[index]
                        for index in group_list
                        if index in result_dict.keys()
                    ]
                    # skip the non-numerical results
                    if len(result_list) > 0 and not isinstance(
                        result_list[0], (int, float)
                    ):
                        continue
                    # average the numerical results and record them
                    if len(result_list) != 0:
                        # create group item lazily
                        if group_name not in table_contents.keys():
                            table_contents[group_name] = []

                        if metric not in table_headers:
                            table_headers.append(metric)
                        table_contents[group_name].append(np.mean(result_list))

            # get the max and min group value for each numerical metric
            for i in range(len(table_headers) - 1):
                metric_value_list = [value[i] for value in table_contents.values()]
                max_value, min_value = max(metric_value_list), min(
                    metric_value_list
                )
                # loop each group
                for group in table_contents.keys():
                    # turn the max number into a bold&italic string
                    if table_contents[group][i] == max_value:
                        table_contents[group][
                            i
                        ] = f"***{table_contents[group][i]:.4f}***"
                    # turn the min number into a bold string
                    elif table_contents[group][i] == min_value:
                        table_contents[group][
                            i
                        ] = f"**{table_contents[group][i]:.4f}**"
                    # turn other numbers into pure strings
                    else:
                        table_contents[group][i] = f"{table_contents[group][i]:.4f}"

            # attach the list of the current group into the result string
            result_string += get_table_strings(
                contents=list(table_contents.values()),
                first_col=list(table_contents.keys()),
                headers=table_headers,
            )
    np.savetxt(result_path, [result_string], fmt="%s")

    # --- Top-N Bad Cases Presentation --- #
    # only present topn bad cases if instance_reports.md is given
    if "instance_reports.md" in self.step_info.keys():
        # loop each tri-tuple
        for metric, mode, num in self.bad_cases_selection:
            result_path = os.path.join(
                self.result_path, f"top{num}_{mode}_{metric}.md"
            )

            if metric in self.step_info.keys():
                # get the indices of the topn samples
                selected_samples = sorted(
                    self.step_info[metric].items(),
                    key=lambda x: x[1],
                    reverse=True if mode.lower() == "max" else False,
                )[:num]
                selected_samples = [s[0] for s in selected_samples]

                # make the .md string for all the top-n bad samples
                sample_reports = ""
                for s_index in selected_samples:
                    sample_reports += (
                        f"**{s_index}**"
                        + self.step_info["instance_reports.md"][s_index]
                    )
                np.savetxt(result_path, [sample_reports], fmt="%s")

    # --- Histograms Plotting --- #
    # remove the old figures if have
    if os.path.exists(os.path.join(self.result_path, "figures")):
        shutil.rmtree(os.path.join(self.result_path, "figures"))

    # loop each metric and plot the histogram figure
    for metric, result_dict in self.step_info.items():
        result_list = list(result_dict.values())
        # only consider the numerical results
        if not isinstance(result_list[0], (int, float)):
            continue

        self.enqueue(
            dict(materials=copy.deepcopy({metric: result_list}), plot_type="hist")
        )

    if not self.empty_queue():
        for i in range(60):
            if not self.empty_queue():
                self.logger.info(
                    "The snapshooter is still snapshotting. Waiting for 10 seconds."
                )
                time.sleep(10)

monitor_init(args, **kwargs)

Initializes the TestMonitor with the given arguments.

Parameters:

Name Type Description Default
args Namespace

Parsed command line arguments.

 required
Source code in speechain/monitor.py 
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
def monitor_init(self, args: argparse.Namespace, **kwargs):
    """Initializes the TestMonitor with the given arguments.

    Args:
        args (argparse.Namespace):
            Parsed command line arguments.
    """
    self.distributed = args.distributed
    self.report_per_steps = args.report_per_steps
    self.bad_cases_selection = args.bad_cases_selection
    if self.bad_cases_selection is None:
        self.bad_cases_selection = []
    elif not isinstance(self.bad_cases_selection[0], List):
        self.bad_cases_selection = [self.bad_cases_selection]

    # initialize the snapshooter of the monitor
    self.data_saving_logs_queue = Queue()
    # create daemon processes for data saving
    assert (
        args.saving_proc_num >= 1
    ), "'saving_proc_num' should be an integer larger than 1!"
    self.saving_proc_num = args.saving_proc_num
    for proc_id in range(self.saving_proc_num):
        Process(
            target=data_saving_logs,
            args=(proc_id, self.data_saving_logs_queue),
            daemon=True,
        ).start()

start_epoch(total_step_num)

Starts a new testing epoch.

Parameters:

Name Type Description Default
total_step_num int

Total number of steps in the epoch.

 required
Source code in speechain/monitor.py 
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
def start_epoch(self, total_step_num: int):
    """Starts a new testing epoch.

    Args:
        total_step_num (int): Total number of steps in the epoch.
    """
    # para init
    self.prev_test_time = time.time()
    self.total_step_num = total_step_num

    if not hasattr(self, "step_info"):
        self.step_info = dict(group_time=[], total_time=0)
    if not hasattr(self, "finished_group_num"):
        self.finished_group_num = 0

state_dict()

Returns the state of the TestMonitor instance. This is useful for checkpointing or logging purposes.

  • A dictionary that includes:
    • step_info: A dictionary storing information about each step.
    • finished_group_num: The number of groups that have finished processing.
Source code in speechain/monitor.py 
1863
1864
1865
1866
1867
1868
1869
1870
1871
1872
1873
1874
def state_dict(self):
    """Returns the state of the TestMonitor instance. This is useful for
    checkpointing or logging purposes.

    Returns:
    - A dictionary that includes:
        * step_info: A dictionary storing information about each step.
        * finished_group_num: The number of groups that have finished processing.
    """
    return dict(
        step_info=self.step_info, finished_group_num=self.finished_group_num
    )

step(step_num, test_results, test_index)

Executes a single test step and logs the results.

Parameters:

Name Type Description Default
step_num int

Current step number.

 required
test_results Dict[str, Dict]

Results of the test step.

 required
test_index List[str]

Indexes of the tests.

 required
Source code in speechain/monitor.py 
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
def step(self, step_num: int, test_results: Dict[str, Dict], test_index: List[str]):
    """Executes a single test step and logs the results.

    Args:
        step_num (int): Current step number.
        test_results (Dict[str, Dict]): Results of the test step.
        test_index (List[str]): Indexes of the tests.
    """
    # --- Write the testing results of the current step to the testing files --- #
    # loop each result list in the returned Dict
    for name, result in test_results.items():
        # for .txt file, register the result contents into self.step_info. text data doesn't occupy too much memory
        if result["format"].lower() == "txt":
            if name not in self.step_info.keys():
                self.step_info[name] = dict()

            if not isinstance(result["content"], List):
                result["content"] = [result["content"]]

            for index, content in zip(test_index, result["content"]):
                # for the List-type element, turn it into its string format
                if isinstance(content, List):
                    content = str(content)
                self.step_info[name][index] = content

        # for other files, data needs to be saved to the disk in real time to reduce the memory burden
        else:
            # we enqueue a whole batch of data to be saved in order to reduce the number of query operations of each daemon process
            self.data_saving_logs_queue.put(
                dict(
                    file_format=result["format"].lower(),
                    save_path=os.path.join(self.result_path, name),
                    file_name_list=test_index,
                    file_content_list=result["content"],
                    group_ids=(
                        result["group_ids"]
                        if "group_ids" in result.keys()
                        else None
                    ),
                    sample_rate=(
                        result["sample_rate"]
                        if "sample_rate" in result.keys()
                        else None
                    ),
                )
            )

    # monitor the approximate size of the queue to be more memory-friendly
    while self.data_saving_logs_queue.qsize() > 5 * self.saving_proc_num:
        self.logger.warning(
            f"There has been more than {3 * self.saving_proc_num} batches in data_saving_logs_queue, "
            f"so the data generation is paused for 30 seconds."
        )
        time.sleep(30)

    # --- Report the testing midway information to users --- #
    test_step_message = None
    # record the tesing time of the current step
    curr_test_time = time.time()
    self.step_info["group_time"].append(curr_test_time - self.prev_test_time)
    self.prev_test_time = curr_test_time

    # meet the reporting interval
    if step_num % self.report_per_steps == 0:
        curr_group_time = sum(self.step_info["group_time"])

        # the first testing step
        if self.finished_group_num == 0:
            prev_group_time = curr_group_time
        # other testing steps
        else:
            # calculate the average time of all the previous groups
            prev_group_time = self.step_info["total_time"] / self.finished_group_num

        # calculate the number of remaining steps
        self.finished_group_num += 1
        finish_step_num = int(self.finished_group_num * self.report_per_steps)
        remaining_step_num = self.total_step_num - finish_step_num
        # take the weighted average consuming time of each group
        aver_group_time = (prev_group_time + curr_group_time) / 2
        remaining_time = aver_group_time * (
            remaining_step_num / self.report_per_steps
        )

        # update the time records
        self.step_info["total_time"] += curr_group_time
        self.step_info["group_time"] = []

        test_step_message = (
            f"Testing Midway Report -- "
            f"testing time for the recent {self.report_per_steps} steps: {curr_group_time:.2f}s -- "
            f"finished step number: {finish_step_num} -- "
            f"remaining step number: {remaining_step_num} -- "
            f"expected remaining time: "
        )

        remaining_days, remaining_time = int(
            remaining_time // (3600 * 24)
        ), remaining_time % (3600 * 24)
        if remaining_days > 0:
            test_step_message += f"{remaining_days:d}d "

        remaining_hours, remaining_time = (
            int(remaining_time // 3600),
            remaining_time % 3600,
        )
        if remaining_hours > 0:
            test_step_message += f"{remaining_hours:d}h "

        remaining_minutes, remaining_time = (
            int(remaining_time // 60),
            remaining_time % 60,
        )
        if remaining_minutes > 0:
            test_step_message += f"{remaining_minutes:d}m "

        remaining_seconds = remaining_time
        test_step_message += f"{remaining_seconds:.2f}s"

    if test_step_message is not None:
        self.logger.info(test_step_message)

wait_empty_queues(sleep_time=60)

Waits until the log queue is empty before continuing.

Parameters:

Name Type Description Default
sleep_time int

Time to wait in seconds when the queue is not empty. Defaults to 60.

 60
Source code in speechain/monitor.py 
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
1616
1617
1618
1619
def wait_empty_queues(self, sleep_time: int = 60):
    """Waits until the log queue is empty before continuing.

    Args:
        sleep_time (int, optional): Time to wait in seconds when the queue is not empty. Defaults to 60.
    """
    while True:
        if self.data_saving_logs_queue.empty():
            # wait for one more time when the queue becomes empty to left enough time for data saving
            time.sleep(sleep_time)
            break
        else:
            self.logger.info(
                f"The data saving process is still working. Waiting for {sleep_time} seconds."
            )
            time.sleep(sleep_time)

    # If using distributed training, synchronize all processes before continuing
    if self.distributed:
        torch.distributed.barrier()

TrainMonitor

Bases: Monitor

The TrainMonitor class extends the Monitor class by adding functionality to track training progress and performance.

It provides methods to initialize monitoring, record metrics at each step and epoch, and store the state of the training process.

Source code in speechain/monitor.py 
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
class TrainMonitor(Monitor):
    """The TrainMonitor class extends the Monitor class by adding functionality to track
    training progress and performance.

    It provides methods to initialize monitoring, record metrics at each step and epoch,
    and store the state of the training process.
    """

    def monitor_init(self, args: argparse.Namespace, **kwargs):
        """Initializes the training monitor with the given arguments. This method is
        responsible for setting up the general members and tracking optimizer
        information.

        Args:
            args (argparse.Namespace):
                Arguments provided for monitoring.
        """
        # general members
        self.report_per_steps = args.report_per_steps
        self.dry_run = args.dry_run
        self.no_optim = args.no_optim
        self.mode = "train"

        # training monitor needs to additionally track optimizer information
        # update epoch-level records
        self.epoch_records["consumed_time"].update(
            loss_backward_time=dict(), optim_time=dict()
        )
        self.epoch_records.update(optim_lr=dict())
        # update step-level records
        self.step_records["consumed_time"].update(
            loss_backward_time=dict(), optim_time=dict()
        )
        self.step_records.update(optim_lr=dict())

    def start_epoch(self, epoch: int):
        """Prepares the monitor for a new epoch of training.

        Args:
            epoch (int):
                The current epoch number.
        """
        # epoch-level information
        self.epoch = epoch
        self.epoch_start_time = time.time()

        # refresh the step-level records at the beginning of each epoch
        self.refresh_step_records()

        # logging the beginning information
        self.logger.info(f"The training part of epoch no.{epoch} starts.")

    def step(
        self,
        step_num: int,
        optim_lr: Dict[str, float],
        train_metrics: Dict[str, torch.Tensor],
    ):
        """Records information for a single training step.

        Args:
            step_num (int):
                The current step number.
            optim_lr (Dict[str, float]):
                The learning rates for each optimizer.
            train_metrics (Dict[str, torch.Tensor]):
                The training metrics for the current step.
        """
        # accumulate the values of training criteria
        if train_metrics is not None:
            self.record_step_info("criteria", train_metrics)

        # accumulate the optimization times and learning rates of each OptimScheduler
        if optim_lr is not None:
            self.record_step_info("optim_lr", optim_lr)

        # report all the information for every 'report_per_steps' training steps
        if step_num % self.report_per_steps == 0:
            # calculate the accumulated time for reporting
            _data_load_time = sum(
                self.step_records["consumed_time"]["data_load_time"][
                    -self.report_per_steps :
                ]
            )
            _model_forward_time = sum(
                self.step_records["consumed_time"]["model_forward_time"][
                    -self.report_per_steps :
                ]
            )
            # initialize the returned message of the current step
            step_message = (
                f"Training step no.{step_num - self.report_per_steps + 1:d}-{step_num:d} -- "
                f"data loading time: {_data_load_time:.2f}s -- "
                f"model forward time: {_model_forward_time:.2f}s -- "
            )

            if not self.dry_run:
                # report the values of criteria in each training step
                step_message += "Training Criteria: "
                for name, result in self.step_records["criteria"].items():
                    _tmp_criteria = result[-self.report_per_steps :]
                    step_message += f"{name}: {np.mean(_tmp_criteria):.2e} -- "

                if not self.no_optim:
                    # report the information of optimizers after each training step
                    step_message += "OptimSchedulers: "
                    for optim_name in self.step_records["optim_lr"].keys():
                        # accumulate the backward and optimization times
                        _loss_backward_time = sum(
                            self.step_records["consumed_time"]["loss_backward_time"][
                                optim_name
                            ][-self.report_per_steps :]
                        )
                        _optim_time = sum(
                            self.step_records["consumed_time"]["optim_time"][
                                optim_name
                            ][-self.report_per_steps :]
                        )
                        # average the learning rate
                        _lr = (
                            sum(
                                self.step_records["optim_lr"][optim_name][
                                    -self.report_per_steps :
                                ]
                            )
                            / self.report_per_steps
                        )

                        # accumulated optimization time and averaged learning_rates are reported
                        step_message += (
                            f"{optim_name}: "
                            f"loss backward time= {_loss_backward_time:.2f}s, "
                            f"optimization time={_optim_time:.2f}s, "
                            f"learning rate={_lr:.2e} -- "
                        )

            # logging the information of the current step
            self.logger.info(step_message)

    def finish_epoch(self):
        """Finishes monitoring for the current epoch, logging information and preparing
        for the next epoch."""
        # ---- The Information Logging Part ---- #
        # report the overall consuming time of the current epoch
        epoch_message = (
            f"The training part of epoch no.{self.epoch} is finished in {time.time() - self.epoch_start_time:.2f}s.\n"
            f"Summary of all training steps:\n"
        )

        # report the information of the consumed calculation time
        epoch_message = self.record_consumed_time(epoch_message)

        # report the information of the consumed GPU memory
        epoch_message = self.record_consumed_memory(epoch_message)

        # data loading only
        if not self.dry_run:
            # report the information of all the training criteria
            epoch_message = self.record_criteria(epoch_message)

            # no optimization
            if not self.no_optim:
                # report the information of all the OptimSchedulers
                epoch_message += " -- OptimScheduler information -- \n"
                # record the optimization information of the current epoch
                for optim_name in self.step_records["optim_lr"].keys():
                    if optim_name not in self.epoch_records["optim_lr"].keys():
                        self.epoch_records["optim_lr"][optim_name] = []
                    if (
                        optim_name
                        not in self.epoch_records["consumed_time"][
                            "loss_backward_time"
                        ].keys()
                    ):
                        self.epoch_records["consumed_time"]["loss_backward_time"][
                            optim_name
                        ] = []
                    if (
                        optim_name
                        not in self.epoch_records["consumed_time"]["optim_time"].keys()
                    ):
                        self.epoch_records["consumed_time"]["optim_time"][
                            optim_name
                        ] = []

                    epoch_message += f"{optim_name} -- "
                    # accumulate the loss backward time
                    _total_time = sum(
                        self.step_records["consumed_time"]["loss_backward_time"][
                            optim_name
                        ]
                    )
                    self.epoch_records["consumed_time"]["loss_backward_time"][
                        optim_name
                    ].append(_total_time)
                    epoch_message += f"Total loss backward time: {_total_time:.2f}s, "

                    # accumulate the optimization time
                    _total_time = sum(
                        self.step_records["consumed_time"]["optim_time"][optim_name]
                    )
                    self.epoch_records["consumed_time"]["optim_time"][
                        optim_name
                    ].append(_total_time)
                    epoch_message += f"Total optimization time: {_total_time:.2f}s, "

                    # average the learning rate
                    aver_lr = np.mean(self.step_records["optim_lr"][optim_name]).item()
                    self.epoch_records["optim_lr"][optim_name].append(aver_lr)
                    epoch_message += f"Average learning rate: {aver_lr:.2e}\n"
                epoch_message += "\n"

        # logging the information for the current epoch
        self.logger.info(epoch_message)

        # ---- The SnapShotting Part ---- #
        for key in self.epoch_records.keys():
            # only snapshot the time records in the dry running mode
            if self.dry_run and key != "consumed_time":
                continue
            # skip the learning rate records in the no optimization mode
            if self.no_optim and key == "optim_lr":
                continue
            # snapshot the epoch records so for to a curve figure
            self.enqueue(
                dict(
                    materials=copy.deepcopy(self.epoch_records[key]),
                    plot_type="curve",
                    epoch=self.epoch,
                    xlabel="epoch",
                    sep_save=False,
                    subfolder_names=key,
                )
            )

        # notify the snapshooter process of the new queue elements
        self.event.set()

    def state_dict(self):
        """Returns the current state of the monitor as a dictionary.

        Returns:
            dict: The current state of the monitor.
        """
        return dict(epoch_records=self.epoch_records)

finish_epoch()

Finishes monitoring for the current epoch, logging information and preparing for the next epoch.

Source code in speechain/monitor.py 
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
def finish_epoch(self):
    """Finishes monitoring for the current epoch, logging information and preparing
    for the next epoch."""
    # ---- The Information Logging Part ---- #
    # report the overall consuming time of the current epoch
    epoch_message = (
        f"The training part of epoch no.{self.epoch} is finished in {time.time() - self.epoch_start_time:.2f}s.\n"
        f"Summary of all training steps:\n"
    )

    # report the information of the consumed calculation time
    epoch_message = self.record_consumed_time(epoch_message)

    # report the information of the consumed GPU memory
    epoch_message = self.record_consumed_memory(epoch_message)

    # data loading only
    if not self.dry_run:
        # report the information of all the training criteria
        epoch_message = self.record_criteria(epoch_message)

        # no optimization
        if not self.no_optim:
            # report the information of all the OptimSchedulers
            epoch_message += " -- OptimScheduler information -- \n"
            # record the optimization information of the current epoch
            for optim_name in self.step_records["optim_lr"].keys():
                if optim_name not in self.epoch_records["optim_lr"].keys():
                    self.epoch_records["optim_lr"][optim_name] = []
                if (
                    optim_name
                    not in self.epoch_records["consumed_time"][
                        "loss_backward_time"
                    ].keys()
                ):
                    self.epoch_records["consumed_time"]["loss_backward_time"][
                        optim_name
                    ] = []
                if (
                    optim_name
                    not in self.epoch_records["consumed_time"]["optim_time"].keys()
                ):
                    self.epoch_records["consumed_time"]["optim_time"][
                        optim_name
                    ] = []

                epoch_message += f"{optim_name} -- "
                # accumulate the loss backward time
                _total_time = sum(
                    self.step_records["consumed_time"]["loss_backward_time"][
                        optim_name
                    ]
                )
                self.epoch_records["consumed_time"]["loss_backward_time"][
                    optim_name
                ].append(_total_time)
                epoch_message += f"Total loss backward time: {_total_time:.2f}s, "

                # accumulate the optimization time
                _total_time = sum(
                    self.step_records["consumed_time"]["optim_time"][optim_name]
                )
                self.epoch_records["consumed_time"]["optim_time"][
                    optim_name
                ].append(_total_time)
                epoch_message += f"Total optimization time: {_total_time:.2f}s, "

                # average the learning rate
                aver_lr = np.mean(self.step_records["optim_lr"][optim_name]).item()
                self.epoch_records["optim_lr"][optim_name].append(aver_lr)
                epoch_message += f"Average learning rate: {aver_lr:.2e}\n"
            epoch_message += "\n"

    # logging the information for the current epoch
    self.logger.info(epoch_message)

    # ---- The SnapShotting Part ---- #
    for key in self.epoch_records.keys():
        # only snapshot the time records in the dry running mode
        if self.dry_run and key != "consumed_time":
            continue
        # skip the learning rate records in the no optimization mode
        if self.no_optim and key == "optim_lr":
            continue
        # snapshot the epoch records so for to a curve figure
        self.enqueue(
            dict(
                materials=copy.deepcopy(self.epoch_records[key]),
                plot_type="curve",
                epoch=self.epoch,
                xlabel="epoch",
                sep_save=False,
                subfolder_names=key,
            )
        )

    # notify the snapshooter process of the new queue elements
    self.event.set()

monitor_init(args, **kwargs)

Initializes the training monitor with the given arguments. This method is responsible for setting up the general members and tracking optimizer information.

Parameters:

Name Type Description Default
args Namespace

Arguments provided for monitoring.

 required
Source code in speechain/monitor.py 
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
def monitor_init(self, args: argparse.Namespace, **kwargs):
    """Initializes the training monitor with the given arguments. This method is
    responsible for setting up the general members and tracking optimizer
    information.

    Args:
        args (argparse.Namespace):
            Arguments provided for monitoring.
    """
    # general members
    self.report_per_steps = args.report_per_steps
    self.dry_run = args.dry_run
    self.no_optim = args.no_optim
    self.mode = "train"

    # training monitor needs to additionally track optimizer information
    # update epoch-level records
    self.epoch_records["consumed_time"].update(
        loss_backward_time=dict(), optim_time=dict()
    )
    self.epoch_records.update(optim_lr=dict())
    # update step-level records
    self.step_records["consumed_time"].update(
        loss_backward_time=dict(), optim_time=dict()
    )
    self.step_records.update(optim_lr=dict())

start_epoch(epoch)

Prepares the monitor for a new epoch of training.

Parameters:

Name Type Description Default
epoch int

The current epoch number.

 required
Source code in speechain/monitor.py 
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
def start_epoch(self, epoch: int):
    """Prepares the monitor for a new epoch of training.

    Args:
        epoch (int):
            The current epoch number.
    """
    # epoch-level information
    self.epoch = epoch
    self.epoch_start_time = time.time()

    # refresh the step-level records at the beginning of each epoch
    self.refresh_step_records()

    # logging the beginning information
    self.logger.info(f"The training part of epoch no.{epoch} starts.")

state_dict()

Returns the current state of the monitor as a dictionary.

Returns:

Name Type Description
dict

The current state of the monitor.
Source code in speechain/monitor.py 
606
607
608
609
610
611
612
def state_dict(self):
    """Returns the current state of the monitor as a dictionary.

    Returns:
        dict: The current state of the monitor.
    """
    return dict(epoch_records=self.epoch_records)

step(step_num, optim_lr, train_metrics)

Records information for a single training step.

Parameters:

Name Type Description Default
step_num int

The current step number.

 required
optim_lr Dict[str, float]

The learning rates for each optimizer.

 required
train_metrics Dict[str, Tensor]

The training metrics for the current step.

 required
Source code in speechain/monitor.py 
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
def step(
    self,
    step_num: int,
    optim_lr: Dict[str, float],
    train_metrics: Dict[str, torch.Tensor],
):
    """Records information for a single training step.

    Args:
        step_num (int):
            The current step number.
        optim_lr (Dict[str, float]):
            The learning rates for each optimizer.
        train_metrics (Dict[str, torch.Tensor]):
            The training metrics for the current step.
    """
    # accumulate the values of training criteria
    if train_metrics is not None:
        self.record_step_info("criteria", train_metrics)

    # accumulate the optimization times and learning rates of each OptimScheduler
    if optim_lr is not None:
        self.record_step_info("optim_lr", optim_lr)

    # report all the information for every 'report_per_steps' training steps
    if step_num % self.report_per_steps == 0:
        # calculate the accumulated time for reporting
        _data_load_time = sum(
            self.step_records["consumed_time"]["data_load_time"][
                -self.report_per_steps :
            ]
        )
        _model_forward_time = sum(
            self.step_records["consumed_time"]["model_forward_time"][
                -self.report_per_steps :
            ]
        )
        # initialize the returned message of the current step
        step_message = (
            f"Training step no.{step_num - self.report_per_steps + 1:d}-{step_num:d} -- "
            f"data loading time: {_data_load_time:.2f}s -- "
            f"model forward time: {_model_forward_time:.2f}s -- "
        )

        if not self.dry_run:
            # report the values of criteria in each training step
            step_message += "Training Criteria: "
            for name, result in self.step_records["criteria"].items():
                _tmp_criteria = result[-self.report_per_steps :]
                step_message += f"{name}: {np.mean(_tmp_criteria):.2e} -- "

            if not self.no_optim:
                # report the information of optimizers after each training step
                step_message += "OptimSchedulers: "
                for optim_name in self.step_records["optim_lr"].keys():
                    # accumulate the backward and optimization times
                    _loss_backward_time = sum(
                        self.step_records["consumed_time"]["loss_backward_time"][
                            optim_name
                        ][-self.report_per_steps :]
                    )
                    _optim_time = sum(
                        self.step_records["consumed_time"]["optim_time"][
                            optim_name
                        ][-self.report_per_steps :]
                    )
                    # average the learning rate
                    _lr = (
                        sum(
                            self.step_records["optim_lr"][optim_name][
                                -self.report_per_steps :
                            ]
                        )
                        / self.report_per_steps
                    )

                    # accumulated optimization time and averaged learning_rates are reported
                    step_message += (
                        f"{optim_name}: "
                        f"loss backward time= {_loss_backward_time:.2f}s, "
                        f"optimization time={_optim_time:.2f}s, "
                        f"learning rate={_lr:.2e} -- "
                    )

        # logging the information of the current step
        self.logger.info(step_message)

TrainValidMonitor

Bases: object

A wrapper class for TrainMonitor and ValidMonitor. The motivations of wrapping TrainMonitor and ValidMonitor together are two-folds: 1. enable multi-metric best model recording among training and validation metrics. 2. decouple TrainMonitor and ValidMonitor from Runner to improve cohesion and code readability.

Attributes:

Name Type Description
logger

Logger object for logging purposes.
train_monitor

An instance of the TrainMonitor class.
valid_monitor

An instance of the ValidMonitor class.
Source code in speechain/monitor.py 
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
class TrainValidMonitor(object):
    """
    A wrapper class for TrainMonitor and ValidMonitor.
    The motivations of wrapping TrainMonitor and ValidMonitor together are two-folds:
        1. enable multi-metric best model recording among training and validation metrics.
        2. decouple TrainMonitor and ValidMonitor from Runner to improve cohesion and code readability.

    Attributes:
        logger: Logger object for logging purposes.
        train_monitor: An instance of the TrainMonitor class.
        valid_monitor: An instance of the ValidMonitor class.
    """

    def __init__(self, logger, args: argparse.Namespace, model: Model):
        """Initializes the TrainValidMonitor with a logger, arguments, and a model.

        Args:
            logger:
                Logger object for logging purposes.
            args:
                argparse.Namespace object, contains command line arguments.
            model:
                Model object, the model to be trained and validated.
        """
        self.logger = logger

        self.train_monitor = TrainMonitor(logger=logger, args=args)
        self.valid_monitor = ValidMonitor(logger=logger, args=args, model=model)

    def start_train_epoch(self, epoch: int):
        """Starts a new training epoch.

        Args:
            epoch: The epoch number.
        """
        self.train_monitor.start_epoch(epoch)

    def train_step(
        self,
        step_num: int,
        optim_lr: Dict[str, float],
        train_metrics: Dict[str, torch.Tensor],
    ):
        """Executes a training step.

        Args:
            step_num: The step number.
            optim_lr: The learning rate(s) of the optimizer(s).
            train_metrics: The training metrics.
        """
        self.train_monitor.step(
            step_num=step_num, optim_lr=optim_lr, train_metrics=train_metrics
        )

    def finish_train_epoch(self):
        """Finishes a training epoch."""
        self.train_monitor.finish_epoch()

    def start_valid_epoch(self, epoch: int):
        """Starts a new validation epoch.

        Args:
            epoch: The epoch number.
        """
        self.valid_monitor.start_epoch(epoch)

    def valid_step(self, valid_metrics: Dict[str, torch.Tensor]):
        """Executes a validation step.

        Args:
            valid_metrics: The validation metrics.
        """
        self.valid_monitor.step(valid_metrics=valid_metrics)

    def valid_model_snapshot(
        self, epoch: int, domain: str, sample_index: str, used_sample: Dict
    ):
        """Creates a snapshot of the validation model.

        Args:
            epoch: The epoch number.
            domain: The domain of validation.
            sample_index: The sample index.
            used_sample: The sample used for validation.
        """
        self.valid_monitor.model_snapshot(
            epoch=epoch,
            domain=domain,
            sample_index=sample_index,
            used_sample=used_sample,
        )

    def finish_valid_epoch(self, valid_flag: bool, valid_per_epochs: int):
        """Finishes a validation epoch.

        Args:
            valid_flag: The validation flag.
            valid_per_epochs: The number of epochs per validation.

        Returns:
            The result from the finish_epoch method of the ValidMonitor object.
        """
        return self.valid_monitor.finish_epoch(
            train_records=self.train_monitor.epoch_records,
            valid_flag=valid_flag,
            valid_per_epochs=valid_per_epochs,
        )

    def wait_empty_queues(self, sleep_time: int = 10, max_wait_round: int = 60):
        """Check whether the snapshot creator processes of train_monitor and
        valid_monitor are still working. Wait until the material queues of the snapshot
        creators become empty.

        Args:
            sleep_time: The sleep time in seconds between each check. Default is 10 seconds.
            max_wait_round: The maximum number of waiting rounds. Default is 60 rounds.
        """
        if not self.train_monitor.empty_queue() or not self.valid_monitor.empty_queue():
            for _ in range(max_wait_round):
                message = ""
                if not self.train_monitor.empty_queue():
                    message += "The training snapshooter is still snapshotting. "
                if not self.valid_monitor.empty_queue():
                    message += "The validation snapshooter is still snapshotting. "
                self.logger.info(message + f"Waiting for {sleep_time} seconds......")
                time.sleep(sleep_time)
            self.logger.info(
                f"The maximal waiting time {max_wait_round * sleep_time} seconds is reached, "
                f"so the snapshooters will be shut down......"
            )

    def state_dict(self):
        """Retrieves the state of the TrainMonitor and ValidMonitor in dictionary form.

        Returns:
            Dict:
                The current state of the TrainMonitor and ValidMonitor, typically including model parameters and
                other related information.
        """
        return dict(
            train_monitor=self.train_monitor.state_dict(),
            valid_monitor=self.valid_monitor.state_dict(),
        )

    def load_state_dict(self, state_dict):
        """Loads the state of the TrainMonitor and ValidMonitor from a dictionary.

        Args:
            state_dict:
                The dictionary containing the state of the TrainMonitor and ValidMonitor.
        """
        self.train_monitor.load_state_dict(state_dict["train_monitor"])
        self.valid_monitor.load_state_dict(state_dict["valid_monitor"])

__init__(logger, args, model)

Initializes the TrainValidMonitor with a logger, arguments, and a model.

Parameters:

Name Type Description Default
logger

Logger object for logging purposes.

 required
args Namespace

argparse.Namespace object, contains command line arguments.

 required
model Model

Model object, the model to be trained and validated.

 required
Source code in speechain/monitor.py 
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
def __init__(self, logger, args: argparse.Namespace, model: Model):
    """Initializes the TrainValidMonitor with a logger, arguments, and a model.

    Args:
        logger:
            Logger object for logging purposes.
        args:
            argparse.Namespace object, contains command line arguments.
        model:
            Model object, the model to be trained and validated.
    """
    self.logger = logger

    self.train_monitor = TrainMonitor(logger=logger, args=args)
    self.valid_monitor = ValidMonitor(logger=logger, args=args, model=model)

finish_train_epoch()

Finishes a training epoch.

Source code in speechain/monitor.py 
1277
1278
1279
def finish_train_epoch(self):
    """Finishes a training epoch."""
    self.train_monitor.finish_epoch()

finish_valid_epoch(valid_flag, valid_per_epochs)

Finishes a validation epoch.

Parameters:

Name Type Description Default
valid_flag bool

The validation flag.

 required
valid_per_epochs int

The number of epochs per validation.

 required

Returns:

Type Description

The result from the finish_epoch method of the ValidMonitor object.
Source code in speechain/monitor.py 
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
def finish_valid_epoch(self, valid_flag: bool, valid_per_epochs: int):
    """Finishes a validation epoch.

    Args:
        valid_flag: The validation flag.
        valid_per_epochs: The number of epochs per validation.

    Returns:
        The result from the finish_epoch method of the ValidMonitor object.
    """
    return self.valid_monitor.finish_epoch(
        train_records=self.train_monitor.epoch_records,
        valid_flag=valid_flag,
        valid_per_epochs=valid_per_epochs,
    )

load_state_dict(state_dict)

Loads the state of the TrainMonitor and ValidMonitor from a dictionary.

Parameters:

Name Type Description Default
state_dict

The dictionary containing the state of the TrainMonitor and ValidMonitor.

 required
Source code in speechain/monitor.py 
1367
1368
1369
1370
1371
1372
1373
1374
1375
def load_state_dict(self, state_dict):
    """Loads the state of the TrainMonitor and ValidMonitor from a dictionary.

    Args:
        state_dict:
            The dictionary containing the state of the TrainMonitor and ValidMonitor.
    """
    self.train_monitor.load_state_dict(state_dict["train_monitor"])
    self.valid_monitor.load_state_dict(state_dict["valid_monitor"])

start_train_epoch(epoch)

Starts a new training epoch.

Parameters:

Name Type Description Default
epoch int

The epoch number.

 required
Source code in speechain/monitor.py 
1252
1253
1254
1255
1256
1257
1258
def start_train_epoch(self, epoch: int):
    """Starts a new training epoch.

    Args:
        epoch: The epoch number.
    """
    self.train_monitor.start_epoch(epoch)

start_valid_epoch(epoch)

Starts a new validation epoch.

Parameters:

Name Type Description Default
epoch int

The epoch number.

 required
Source code in speechain/monitor.py 
1281
1282
1283
1284
1285
1286
1287
def start_valid_epoch(self, epoch: int):
    """Starts a new validation epoch.

    Args:
        epoch: The epoch number.
    """
    self.valid_monitor.start_epoch(epoch)

state_dict()

Retrieves the state of the TrainMonitor and ValidMonitor in dictionary form.

Returns:

Name Type Description
Dict

The current state of the TrainMonitor and ValidMonitor, typically including model parameters and other related information.
Source code in speechain/monitor.py 
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
def state_dict(self):
    """Retrieves the state of the TrainMonitor and ValidMonitor in dictionary form.

    Returns:
        Dict:
            The current state of the TrainMonitor and ValidMonitor, typically including model parameters and
            other related information.
    """
    return dict(
        train_monitor=self.train_monitor.state_dict(),
        valid_monitor=self.valid_monitor.state_dict(),
    )

train_step(step_num, optim_lr, train_metrics)

Executes a training step.

Parameters:

Name Type Description Default
step_num int

The step number.

 required
optim_lr Dict[str, float]

The learning rate(s) of the optimizer(s).

 required
train_metrics Dict[str, Tensor]

The training metrics.

 required
Source code in speechain/monitor.py 
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
def train_step(
    self,
    step_num: int,
    optim_lr: Dict[str, float],
    train_metrics: Dict[str, torch.Tensor],
):
    """Executes a training step.

    Args:
        step_num: The step number.
        optim_lr: The learning rate(s) of the optimizer(s).
        train_metrics: The training metrics.
    """
    self.train_monitor.step(
        step_num=step_num, optim_lr=optim_lr, train_metrics=train_metrics
    )

valid_model_snapshot(epoch, domain, sample_index, used_sample)

Creates a snapshot of the validation model.

Parameters:

Name Type Description Default
epoch int

The epoch number.

 required
domain str

The domain of validation.

 required
sample_index str

The sample index.

 required
used_sample Dict

The sample used for validation.

 required
Source code in speechain/monitor.py 
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
def valid_model_snapshot(
    self, epoch: int, domain: str, sample_index: str, used_sample: Dict
):
    """Creates a snapshot of the validation model.

    Args:
        epoch: The epoch number.
        domain: The domain of validation.
        sample_index: The sample index.
        used_sample: The sample used for validation.
    """
    self.valid_monitor.model_snapshot(
        epoch=epoch,
        domain=domain,
        sample_index=sample_index,
        used_sample=used_sample,
    )

valid_step(valid_metrics)

Executes a validation step.

Parameters:

Name Type Description Default
valid_metrics Dict[str, Tensor]

The validation metrics.

 required
Source code in speechain/monitor.py 
1289
1290
1291
1292
1293
1294
1295
def valid_step(self, valid_metrics: Dict[str, torch.Tensor]):
    """Executes a validation step.

    Args:
        valid_metrics: The validation metrics.
    """
    self.valid_monitor.step(valid_metrics=valid_metrics)

wait_empty_queues(sleep_time=10, max_wait_round=60)

Check whether the snapshot creator processes of train_monitor and valid_monitor are still working. Wait until the material queues of the snapshot creators become empty.

Parameters:

Name Type Description Default
sleep_time int

The sleep time in seconds between each check. Default is 10 seconds.

 10
max_wait_round int

The maximum number of waiting rounds. Default is 60 rounds.

 60
Source code in speechain/monitor.py 
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
def wait_empty_queues(self, sleep_time: int = 10, max_wait_round: int = 60):
    """Check whether the snapshot creator processes of train_monitor and
    valid_monitor are still working. Wait until the material queues of the snapshot
    creators become empty.

    Args:
        sleep_time: The sleep time in seconds between each check. Default is 10 seconds.
        max_wait_round: The maximum number of waiting rounds. Default is 60 rounds.
    """
    if not self.train_monitor.empty_queue() or not self.valid_monitor.empty_queue():
        for _ in range(max_wait_round):
            message = ""
            if not self.train_monitor.empty_queue():
                message += "The training snapshooter is still snapshotting. "
            if not self.valid_monitor.empty_queue():
                message += "The validation snapshooter is still snapshotting. "
            self.logger.info(message + f"Waiting for {sleep_time} seconds......")
            time.sleep(sleep_time)
        self.logger.info(
            f"The maximal waiting time {max_wait_round * sleep_time} seconds is reached, "
            f"so the snapshooters will be shut down......"
        )

ValidMonitor

Bases: Monitor

The ValidMonitor class extends the Monitor class by adding functionality to track validation progress and performance.

It provides methods to initialize monitoring, record metrics at each validation step, and snapshot the model at each epoch.

Source code in speechain/monitor.py 
 615
 616
 617
 618
 619
 620
 621
 622
 623
 624
 625
 626
 627
 628
 629
 630
 631
 632
 633
 634
 635
 636
 637
 638
 639
 640
 641
 642
 643
 644
 645
 646
 647
 648
 649
 650
 651
 652
 653
 654
 655
 656
 657
 658
 659
 660
 661
 662
 663
 664
 665
 666
 667
 668
 669
 670
 671
 672
 673
 674
 675
 676
 677
 678
 679
 680
 681
 682
 683
 684
 685
 686
 687
 688
 689
 690
 691
 692
 693
 694
 695
 696
 697
 698
 699
 700
 701
 702
 703
 704
 705
 706
 707
 708
 709
 710
 711
 712
 713
 714
 715
 716
 717
 718
 719
 720
 721
 722
 723
 724
 725
 726
 727
 728
 729
 730
 731
 732
 733
 734
 735
 736
 737
 738
 739
 740
 741
 742
 743
 744
 745
 746
 747
 748
 749
 750
 751
 752
 753
 754
 755
 756
 757
 758
 759
 760
 761
 762
 763
 764
 765
 766
 767
 768
 769
 770
 771
 772
 773
 774
 775
 776
 777
 778
 779
 780
 781
 782
 783
 784
 785
 786
 787
 788
 789
 790
 791
 792
 793
 794
 795
 796
 797
 798
 799
 800
 801
 802
 803
 804
 805
 806
 807
 808
 809
 810
 811
 812
 813
 814
 815
 816
 817
 818
 819
 820
 821
 822
 823
 824
 825
 826
 827
 828
 829
 830
 831
 832
 833
 834
 835
 836
 837
 838
 839
 840
 841
 842
 843
 844
 845
 846
 847
 848
 849
 850
 851
 852
 853
 854
 855
 856
 857
 858
 859
 860
 861
 862
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
class ValidMonitor(Monitor):
    """The ValidMonitor class extends the Monitor class by adding functionality to track
    validation progress and performance.

    It provides methods to initialize monitoring, record metrics at each validation
    step, and snapshot the model at each epoch.
    """

    def monitor_init(self, args: argparse.Namespace, model: Model = None):
        """Initializes the validation monitor with the given arguments and the model.
        This method is responsible for setting up the general members, tracking best
        models, early stopping, and last models.

        Args:
            args (argparse.Namespace):
                Arguments provided for monitoring.
            model (Model, optional):
                The model being validated. This parameter must not be None.

        Raises:
            AssertionError: If the provided model is None.
        """
        assert model is not None, "Model must be provided and not None."
        # register a pointer of the model
        self.model = model

        # running mode
        self.dry_run = args.dry_run
        self.no_optim = args.no_optim
        self.mode = "valid"

        # best models-related members
        self.best_model_selection = args.best_model_selection
        # receive a single metric as a standalone list or tuple
        if isinstance(self.best_model_selection, (List, tuple)) and isinstance(
            self.best_model_selection[0], str
        ):
            self.best_model_selection = [self.best_model_selection]
        else:
            assert isinstance(self.best_model_selection, List), (
                f"best_model_selection must be given as a list, "
                f"but got type(best_model_selection)={self.best_model_selection}."
            )

        for i in range(len(self.best_model_selection)):
            # checking the argument types
            assert isinstance(self.best_model_selection[i], (List, tuple)), (
                "Each element of best_model_selection must be either a list or a tuple, "
                f"but got type={type(self.best_model_selection[i])}."
            )
            assert len(self.best_model_selection[i]) == 4, (
                f"Each element of best_model_selection must be a quad-tuple or qual-list, "
                f"but got length={len(self.best_model_selection[i])}."
            )

            if isinstance(self.best_model_selection[i], tuple):
                self.best_model_selection[i] = list(self.best_model_selection[i])
            self.best_model_selection[i][2] = self.best_model_selection[i][2].lower()
            assert self.best_model_selection[i][2] in [
                "max",
                "min",
            ], f"The best_model_mode must be either 'max' or 'min', but got {self.best_model_selection[i][2]}."

        # model saving-related members
        self.best_model_performance = dict()
        for metric in self.best_model_selection:
            self.best_model_performance["_".join(metric[:2])] = dict()
        self.saved_model_epoch = []
        self.model_save_path = os.path.join(self.result_path, "models")
        if not os.path.exists(self.model_save_path):
            os.makedirs(self.model_save_path, exist_ok=True)

        # early stopping-related members, the first metric in self.best_model_selection is used
        self.early_stopping_metric = "_".join(self.best_model_selection[0][:2])
        self.early_stopping_mode = self.best_model_selection[0][2]
        self.early_stopping_patience = args.early_stopping_patience
        self.early_stopping_threshold = args.early_stopping_threshold
        self.early_stopping_epochs = 0
        self.last_best_performance = (
            0.0 if self.early_stopping_mode == "max" else torch.inf
        )

        # last models-related members
        self.last_model_number = args.last_model_number
        if self.last_model_number < 1:
            raise ValueError(
                "last_model_number cannot be lower than 1, "
                "otherwise the training will not be able to resume."
                f"Got last_model_number={self.last_model_number}!"
            )

        # initialize the snapshooter of this validation monitor
        self.visual_snapshot_interval = args.visual_snapshot_interval

    def start_epoch(self, epoch: int):
        """Prepares the monitor for a new epoch of validation.

        Args:
            epoch (int): The current epoch number.
        """
        # epoch-level information
        if epoch in self.best_model_performance.keys():
            self.logger.warning(
                f"The record of epoch no.{epoch} has already existed in the monitor! "
                f"It will be overwritten by the new record obtained shortly thereafter."
            )
        self.epoch = epoch
        self.epoch_start_time = time.time()

        # refresh the step-level records at the beginning of each epoch
        self.refresh_step_records()

        # logging the beginning information
        self.logger.info(f"The validation part of epoch no.{epoch} starts.")

    def step(self, valid_metrics: Dict[str, torch.Tensor]):
        """Records information for a single validation step.

        Args:
            valid_metrics (Dict[str, torch.Tensor]):
                The validation metrics for the current step.
        """
        # accumulate the values of validation criteria
        if valid_metrics is not None:
            self.record_step_info("criteria", valid_metrics)

    def model_snapshot(
        self, epoch: int, domain: str, sample_index: str, used_sample: Dict
    ):
        """Takes a snapshot of the model at the given epoch for a given sample.

        Args:
            epoch (int):
                The current epoch number.
            domain (str):
                The domain of the current sample.
            sample_index (str):
                The index of the current sample.
            used_sample (Dict):
                The current sample being used for validation.
        """
        # initialize the sub-dict for each sample
        if sample_index not in self.epoch_records.keys():
            self.epoch_records[sample_index] = dict()

        # get the visualization logs for model snapshotting
        vis_logs = self.model(
            batch_data=used_sample,
            epoch=epoch,
            domain=domain,
            epoch_records=self.epoch_records,
            sample_index=sample_index,
            snapshot_interval=self.visual_snapshot_interval,
        )

        # put all the visualization logs into the queue
        self.enqueue(vis_logs)

    @staticmethod
    def is_better(
        query: int or float, target: int or float, mode: str, threshold: float = 0.0
    ):
        """Compares a query value with a target value under a specified mode, optionally
        considering a threshold.

        Parameters:
            query (int, float):
                The value to be compared against the target.
            target (int, float):
                The reference value for the comparison.
            mode (str):
                The comparison mode - 'max' implies the query is considered better if it's larger,
                'min' implies the query is considered better if it's smaller.
            threshold (float, optional):
                A value that adjusts the target value before comparison. Default is 0.0.

        Returns:
            bool: True if the query value is considered better than the target, False otherwise.
        """
        _target = target
        # relative threshold if the argument value is positive
        if threshold > 0:
            _target *= 1 + threshold if mode == "max" else 1 - threshold
        # absolute threshold if the argument value is negative
        elif threshold < 0:
            _target += -threshold if mode == "max" else threshold

        # the threshold is applied to the better comparison
        return query > _target if mode == "max" else query < _target

    def model_insert(self, train_records: Dict, valid_flag: bool):
        """Inserts the model into the ensemble if it's better than the existing models.
        The model is evaluated using the training records and the validation flag.

        Parameters:
            train_records (Dict):
                A dictionary containing the training records, presumably including model performance metrics.
            valid_flag (bool):
                A flag indicating whether the model has passed validation.

        Returns:
            None
        """
        # loop each metric for best model selection
        for metric in self.best_model_selection:
            if metric[0] == "valid" and not valid_flag:
                continue

            _metric_name, _metric_mode, _model_num = (
                "_".join(metric[:2]),
                metric[2],
                metric[3],
            )
            _criteria_dict = (
                train_records["criteria"]
                if metric[0] == "train"
                else self.epoch_records["criteria"]
            )
            curr_performance = _criteria_dict[metric[1]][-1]

            # controls whether to insert the current model into self.best_model_performance or not
            model_insert_flag = False

            # if there is no empty positions for the model of the current epoch
            if len(self.best_model_performance[_metric_name]) == _model_num:
                # as long as the current performance is better than one existing record, it should be inserted.
                for performance in self.best_model_performance[_metric_name].values():
                    if self.is_better(
                        query=curr_performance, target=performance, mode=_metric_mode
                    ):
                        model_insert_flag = True
                        break
            # True if there are some empty positions for the best models
            else:
                model_insert_flag = True

            # record the current model performance
            if model_insert_flag:
                # record the performance of the current epoch
                self.best_model_performance[_metric_name][self.epoch] = curr_performance

        # save the model of the latest epoch onto the disk
        torch.save(
            self.model.state_dict(),
            os.path.join(self.model_save_path, f"epoch_{self.epoch}.pth"),
        )
        self.saved_model_epoch.append(self.epoch)

    def update_best_and_pop_worst(self, epoch_message: str):
        """Updates the best model in the ensemble and removes the worst model. The best
        and worst are determined based on the performance metrics. The function also
        handles logging related to model performance.

        Parameters:
            epoch_message (str):
                A string message related to the current epoch.

        Returns:
            Tuple[str, bool, Dict[str, bool]]:
                Returns a tuple containing the updated epoch message, a flag indicating whether early stopping
                conditions are met, and flags related to model performance metrics.
        """

        def whether_remove(remove_epoch: int):
            # retain the last several models within self.last_model_number
            if self.epoch - remove_epoch < self.last_model_number:
                return False
            else:
                remove_flag = True
                # access metric_epoch_records from the outer scope
                for _epoch_record in metric_epoch_records.values():
                    if remove_epoch in _epoch_record["sorted_epochs"]:
                        remove_flag = False
                        break
                return remove_flag

        # --- Gather the epoch record information for each metric --- #
        metric_epoch_records = dict()
        # loop each metric for best model selection
        for metric in self.best_model_selection:
            _metric_name, _metric_mode, _model_num = (
                "_".join(metric[:2]),
                metric[2],
                metric[3],
            )

            # find the best epoch and worst epoch in self.best_model_performance
            sorted_epochs = dict(
                sorted(
                    self.best_model_performance[_metric_name].items(),
                    key=lambda x: x[1],
                    reverse=True if _metric_mode == "max" else False,
                )
            )
            metric_epoch_records[_metric_name] = dict(
                sorted_epochs=list(sorted_epochs.keys()),
                metric_mode=_metric_mode,
                model_num=_model_num,
            )

        # --- Pop out the worst model and Update the model symbol links --- #
        metric_pop_flags = dict()
        for metric_name, epoch_record in metric_epoch_records.items():
            # controls whether the worst model has been pooped out or not
            metric_pop_flags[metric_name] = False
            # pop out the worst model if there is a redundant one in self.best_model_performance
            if len(epoch_record["sorted_epochs"]) > epoch_record["model_num"]:
                # pick up the epoch number of the worst model
                worst_epoch = epoch_record["sorted_epochs"][-1]
                self.best_model_performance[metric_name].pop(worst_epoch)
                epoch_record["sorted_epochs"].remove(worst_epoch)
                metric_pop_flags[metric_name] = True

            # update the symbol links of all the best models so far
            for i, epoch in enumerate(epoch_record["sorted_epochs"]):
                _best_model_pointer = (
                    f"{metric_name}_best.pth"
                    if i == 0
                    else f"{metric_name}_best_{i + 1}.pth"
                )
                # create a soft link from the best model pointer to the model fi le of the current epoch
                symlink_dst = os.path.join(self.model_save_path, _best_model_pointer)
                if os.path.islink(symlink_dst) or os.path.exists(symlink_dst):
                    os.unlink(symlink_dst)
                os.symlink(
                    os.path.join(self.model_save_path, f"epoch_{epoch}.pth"),
                    symlink_dst,
                )

        # update the symbol links of the last several models
        for epoch in range(self.epoch, max(0, self.epoch - self.last_model_number), -1):
            _last_model_pointer = (
                "latest.pth"
                if epoch == self.epoch
                else f"last_{self.epoch - epoch + 1}.pth"
            )
            # create a soft link from the best model pointer to the model file of the current epoch
            symlink_dst = os.path.join(self.model_save_path, _last_model_pointer)
            if os.path.islink(symlink_dst) or os.path.exists(symlink_dst):
                os.unlink(symlink_dst)
            os.symlink(
                os.path.join(self.model_save_path, f"epoch_{epoch}.pth"), symlink_dst
            )

        # remove the redundant model files
        saved_epochs = self.saved_model_epoch.copy()
        for epoch in saved_epochs:
            epoch_model_path = os.path.join(self.model_save_path, f"epoch_{epoch}.pth")
            if whether_remove(epoch):
                # remove the record of epoch in the memory
                self.saved_model_epoch.remove(epoch)

                # remove the model file if it exists
                if os.path.exists(epoch_model_path):
                    # ensure that the model to be removed is successfully removed
                    while os.path.exists(epoch_model_path):
                        os.remove(epoch_model_path)

        # --- Early-Stopping epoch number checking for the early-stopping metric --- #
        if len(metric_epoch_records[self.early_stopping_metric]["sorted_epochs"]) != 0:
            best_epoch = metric_epoch_records[self.early_stopping_metric][
                "sorted_epochs"
            ][0]
            # refresh to 0 or add 1 depending on the comparison the best one and the second best one
            if best_epoch == self.epoch:
                epoch_message += f"{self.early_stopping_metric} of the current epoch no.{self.epoch} is the best so far.\n"

                # compare the current performance and the last best performance
                best_performance = self.best_model_performance[
                    self.early_stopping_metric
                ][best_epoch]
                if self.is_better(
                    best_performance,
                    self.last_best_performance,
                    mode=self.early_stopping_mode,
                    threshold=self.early_stopping_threshold,
                ):
                    epoch_message += (
                        f"The early-stopping threshold {self.early_stopping_threshold} is reached, "
                        "so the early-stopping epoch number is refreshed.\n"
                    )
                    self.early_stopping_epochs = 0
                    self.last_best_performance = best_performance
                else:
                    epoch_message += (
                        f"The early-stopping threshold {self.early_stopping_threshold} is not reached, "
                        "so the early-stopping epoch number keeps increasing.\n"
                    )
                    self.early_stopping_epochs += 1
            # directly add 1 if the current epoch is not the best
            else:
                epoch_message += f"No improvement of {self.early_stopping_metric} in the current epoch no.{self.epoch}.\n"
                self.early_stopping_epochs += 1

            # report the updated early-stopping epoch number
            epoch_message += f"The early-stopping epoch number has been updated to {self.early_stopping_epochs}.\n"

        # early-stopping check by the patience
        early_stopping_flag = False
        if (
            self.early_stopping_patience is not None
            and self.early_stopping_epochs > self.early_stopping_patience
        ):
            epoch_message += (
                f"The early-stopping patience {self.early_stopping_patience} is reached, "
                f"so the training process stops here.\n"
            )
            early_stopping_flag = True
        elif self.early_stopping_patience is None:
            epoch_message += (
                "The early-stopping patience is not set by your exp_cfg, "
                "so the training will continue until num_epochs is reached.\n"
            )

        return epoch_message, early_stopping_flag, metric_pop_flags

    def save_aver_model(self, epoch_message: str, metric_pop_flags: Dict[str, bool]):
        """Stores the average model in the ensemble and updates the epoch message based
        on the model performance metrics.

        Parameters:
            epoch_message (str):
                A string message related to the current epoch.
            metric_pop_flags (Dict[str, bool]):
                A dictionary containing flags related to model performance metrics.

        Returns:
            str: The updated epoch message.
        """

        def save_aver_models(
            aver_epoch_list: List, aver_num: int, aver_model_name: str
        ):
            # no average model is saved if there is only one candidate model
            if len(aver_epoch_list) == 1:
                return ""

            # sum up the parameters of all best models
            avg_model = None
            for epoch in aver_epoch_list:
                _tgt_model_path = os.path.join(
                    self.model_save_path, f"epoch_{epoch}.pth"
                )
                # skip if the model doesn't exist
                if not os.path.exists(_tgt_model_path):
                    continue

                _avg = None
                # access self.model_save_path from the outer scope
                if avg_model is not None:
                    _avg = torch.load(_tgt_model_path, map_location="cpu")
                else:
                    avg_model = torch.load(_tgt_model_path, map_location="cpu")

                if _avg is not None:
                    for key in avg_model.keys():
                        avg_model[key] += _avg[key]

            # if no average model, skip this function and return an empty string
            if avg_model is None:
                return ""
            else:
                # for the parameters whose dtype is int, averaging is not performed
                # reference: https://github.com/espnet/espnet/blob/5fa6dcc4e649dc66397c629d0030d09ecef36b80/espnet2/main_funcs/average_nbest_models.py#L90
                for key in avg_model.keys():
                    if not str(avg_model[key].dtype).startswith("torch.int"):
                        avg_model[key] /= aver_num

                # save the average model
                _aver_model_path = os.path.join(self.model_save_path, aver_model_name)
                torch.save(avg_model, _aver_model_path)

                return f"{aver_model_name} has been updated to the average of epochs {aver_epoch_list}.\n"

        # --- Save the average model for the best models of each metric --- #
        # loop each metric for best model selection
        for metric in self.best_model_selection:
            _metric_name, _metric_mode, _model_num = (
                "_".join(metric[:2]),
                metric[2],
                metric[3],
            )

            # average the recorded best models so far
            if (
                len(self.best_model_performance[_metric_name]) == _model_num
                and metric_pop_flags[_metric_name]
            ):
                epoch_message += save_aver_models(
                    aver_epoch_list=list(
                        self.best_model_performance[_metric_name].keys()
                    ),
                    aver_num=len(self.best_model_performance[_metric_name]),
                    aver_model_name=f"{_model_num}_{_metric_name}_average.pth",
                )

        # --- Save the average model of the last models --- #
        if self.epoch >= self.last_model_number:
            epoch_message += save_aver_models(
                aver_epoch_list=list(
                    range(self.epoch, self.epoch - self.last_model_number, -1)
                )[::-1],
                aver_num=self.last_model_number,
                aver_model_name=f"{self.last_model_number}_last_average.pth",
            )

        return epoch_message

    def finish_epoch(
        self, train_records: Dict, valid_flag: bool, valid_per_epochs: int
    ):
        """Completes the current epoch by evaluating the model on the training records
        and potentially performing validation based on the validation flag and number of
        validation epochs.

        Parameters:
            train_records (Dict):
                A dictionary containing the training records, presumably including model performance metrics.
            valid_flag (bool):
                A flag indicating whether the model has passed validation.
            valid_per_epochs (int):
                The number of epochs between validations.

        Returns:
            bool: A flag indicating whether early stopping conditions are met.
        """
        # ---- The Information Logging Part ---- #
        if valid_flag:
            # report the overall consuming time of the current validation epoch
            epoch_message = (
                f"The validation part of epoch no.{self.epoch} is finished in "
                f"{time.time() - self.epoch_start_time:.2f}s.\n"
                f"Summary of all validation steps:\n"
            )

            # report the information of the consumed calculation time
            epoch_message = self.record_consumed_time(epoch_message)

            # report the information of the consumed GPU memory
            epoch_message = self.record_consumed_memory(epoch_message)

            if not self.dry_run:
                # report the information of all the validation criteria
                epoch_message = self.record_criteria(epoch_message)

            # ---- The SnapShotting Part ---- #
            for key in self.epoch_records.keys():
                # only snapshot the time and memory info in the dry running mode
                if self.dry_run and key != "consumed_time":
                    continue
                # skip the model visualization records
                elif key in ["consumed_time", "consumed_memory", "criteria"]:
                    # snapshot the epoch records so for to a curve figure
                    self.enqueue(
                        dict(
                            materials=copy.deepcopy(self.epoch_records[key]),
                            plot_type="curve",
                            epoch=self.epoch,
                            xlabel="epoch",
                            sep_save=False,
                            subfolder_names=key,
                            x_stride=valid_per_epochs,
                        )
                    )

            # notify the snapshooter process of the new queue elements
            self.event.set()

        else:
            epoch_message = (
                f"The validation part of epoch no.{self.epoch} is skipped.\n"
            )

        # ---- The Model Saving and Early Stopping Part ---- #
        early_stopping_flag = False
        if not self.dry_run:
            # insert the current model into self.best_model_performance if needed
            self.model_insert(train_records, valid_flag)

            # After inserting, deal with the worst model performance so far and check the early-stopping
            (
                epoch_message,
                early_stopping_flag,
                metric_pop_flags,
            ) = self.update_best_and_pop_worst(epoch_message)

            # save the average models of the best models so far if needed
            epoch_message = self.save_aver_model(epoch_message, metric_pop_flags)

        # log the information of the current validation epoch
        self.logger.info(epoch_message)
        return early_stopping_flag

    def state_dict(self):
        """Retrieves the state of the ensemble in dictionary form.

        Returns:
            Dict: The current state of the ensemble, typically including model parameters and other related information.
        """
        return dict(
            epoch_records=self.epoch_records,
            saved_model_epoch=self.saved_model_epoch,
            best_model_performance=self.best_model_performance,
            early_stopping_epochs=self.early_stopping_epochs,
            last_best_performance=self.last_best_performance,
        )

finish_epoch(train_records, valid_flag, valid_per_epochs)

Completes the current epoch by evaluating the model on the training records and potentially performing validation based on the validation flag and number of validation epochs.

Parameters:

Name Type Description Default
train_records Dict

A dictionary containing the training records, presumably including model performance metrics.

 required
valid_flag bool

A flag indicating whether the model has passed validation.

 required
valid_per_epochs int

The number of epochs between validations.

 required

Returns:

Name Type Description
bool

A flag indicating whether early stopping conditions are met.
Source code in speechain/monitor.py 
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
def finish_epoch(
    self, train_records: Dict, valid_flag: bool, valid_per_epochs: int
):
    """Completes the current epoch by evaluating the model on the training records
    and potentially performing validation based on the validation flag and number of
    validation epochs.

    Parameters:
        train_records (Dict):
            A dictionary containing the training records, presumably including model performance metrics.
        valid_flag (bool):
            A flag indicating whether the model has passed validation.
        valid_per_epochs (int):
            The number of epochs between validations.

    Returns:
        bool: A flag indicating whether early stopping conditions are met.
    """
    # ---- The Information Logging Part ---- #
    if valid_flag:
        # report the overall consuming time of the current validation epoch
        epoch_message = (
            f"The validation part of epoch no.{self.epoch} is finished in "
            f"{time.time() - self.epoch_start_time:.2f}s.\n"
            f"Summary of all validation steps:\n"
        )

        # report the information of the consumed calculation time
        epoch_message = self.record_consumed_time(epoch_message)

        # report the information of the consumed GPU memory
        epoch_message = self.record_consumed_memory(epoch_message)

        if not self.dry_run:
            # report the information of all the validation criteria
            epoch_message = self.record_criteria(epoch_message)

        # ---- The SnapShotting Part ---- #
        for key in self.epoch_records.keys():
            # only snapshot the time and memory info in the dry running mode
            if self.dry_run and key != "consumed_time":
                continue
            # skip the model visualization records
            elif key in ["consumed_time", "consumed_memory", "criteria"]:
                # snapshot the epoch records so for to a curve figure
                self.enqueue(
                    dict(
                        materials=copy.deepcopy(self.epoch_records[key]),
                        plot_type="curve",
                        epoch=self.epoch,
                        xlabel="epoch",
                        sep_save=False,
                        subfolder_names=key,
                        x_stride=valid_per_epochs,
                    )
                )

        # notify the snapshooter process of the new queue elements
        self.event.set()

    else:
        epoch_message = (
            f"The validation part of epoch no.{self.epoch} is skipped.\n"
        )

    # ---- The Model Saving and Early Stopping Part ---- #
    early_stopping_flag = False
    if not self.dry_run:
        # insert the current model into self.best_model_performance if needed
        self.model_insert(train_records, valid_flag)

        # After inserting, deal with the worst model performance so far and check the early-stopping
        (
            epoch_message,
            early_stopping_flag,
            metric_pop_flags,
        ) = self.update_best_and_pop_worst(epoch_message)

        # save the average models of the best models so far if needed
        epoch_message = self.save_aver_model(epoch_message, metric_pop_flags)

    # log the information of the current validation epoch
    self.logger.info(epoch_message)
    return early_stopping_flag

is_better(query, target, mode, threshold=0.0) staticmethod

Compares a query value with a target value under a specified mode, optionally considering a threshold.

Parameters:

Name Type Description Default
query (int, float)

The value to be compared against the target.

 required
target (int, float)

The reference value for the comparison.

 required
mode str

The comparison mode - 'max' implies the query is considered better if it's larger, 'min' implies the query is considered better if it's smaller.

 required
threshold float

A value that adjusts the target value before comparison. Default is 0.0.

 0.0

Returns:

Name Type Description
bool

True if the query value is considered better than the target, False otherwise.
Source code in speechain/monitor.py 
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
@staticmethod
def is_better(
    query: int or float, target: int or float, mode: str, threshold: float = 0.0
):
    """Compares a query value with a target value under a specified mode, optionally
    considering a threshold.

    Parameters:
        query (int, float):
            The value to be compared against the target.
        target (int, float):
            The reference value for the comparison.
        mode (str):
            The comparison mode - 'max' implies the query is considered better if it's larger,
            'min' implies the query is considered better if it's smaller.
        threshold (float, optional):
            A value that adjusts the target value before comparison. Default is 0.0.

    Returns:
        bool: True if the query value is considered better than the target, False otherwise.
    """
    _target = target
    # relative threshold if the argument value is positive
    if threshold > 0:
        _target *= 1 + threshold if mode == "max" else 1 - threshold
    # absolute threshold if the argument value is negative
    elif threshold < 0:
        _target += -threshold if mode == "max" else threshold

    # the threshold is applied to the better comparison
    return query > _target if mode == "max" else query < _target

model_insert(train_records, valid_flag)

Inserts the model into the ensemble if it's better than the existing models. The model is evaluated using the training records and the validation flag.

Parameters:

Name Type Description Default
train_records Dict

A dictionary containing the training records, presumably including model performance metrics.

 required
valid_flag bool

A flag indicating whether the model has passed validation.

 required

Returns:

Type Description

None
Source code in speechain/monitor.py 
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
def model_insert(self, train_records: Dict, valid_flag: bool):
    """Inserts the model into the ensemble if it's better than the existing models.
    The model is evaluated using the training records and the validation flag.

    Parameters:
        train_records (Dict):
            A dictionary containing the training records, presumably including model performance metrics.
        valid_flag (bool):
            A flag indicating whether the model has passed validation.

    Returns:
        None
    """
    # loop each metric for best model selection
    for metric in self.best_model_selection:
        if metric[0] == "valid" and not valid_flag:
            continue

        _metric_name, _metric_mode, _model_num = (
            "_".join(metric[:2]),
            metric[2],
            metric[3],
        )
        _criteria_dict = (
            train_records["criteria"]
            if metric[0] == "train"
            else self.epoch_records["criteria"]
        )
        curr_performance = _criteria_dict[metric[1]][-1]

        # controls whether to insert the current model into self.best_model_performance or not
        model_insert_flag = False

        # if there is no empty positions for the model of the current epoch
        if len(self.best_model_performance[_metric_name]) == _model_num:
            # as long as the current performance is better than one existing record, it should be inserted.
            for performance in self.best_model_performance[_metric_name].values():
                if self.is_better(
                    query=curr_performance, target=performance, mode=_metric_mode
                ):
                    model_insert_flag = True
                    break
        # True if there are some empty positions for the best models
        else:
            model_insert_flag = True

        # record the current model performance
        if model_insert_flag:
            # record the performance of the current epoch
            self.best_model_performance[_metric_name][self.epoch] = curr_performance

    # save the model of the latest epoch onto the disk
    torch.save(
        self.model.state_dict(),
        os.path.join(self.model_save_path, f"epoch_{self.epoch}.pth"),
    )
    self.saved_model_epoch.append(self.epoch)

model_snapshot(epoch, domain, sample_index, used_sample)

Takes a snapshot of the model at the given epoch for a given sample.

Parameters:

Name Type Description Default
epoch int

The current epoch number.

 required
domain str

The domain of the current sample.

 required
sample_index str

The index of the current sample.

 required
used_sample Dict

The current sample being used for validation.

 required
Source code in speechain/monitor.py 
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
def model_snapshot(
    self, epoch: int, domain: str, sample_index: str, used_sample: Dict
):
    """Takes a snapshot of the model at the given epoch for a given sample.

    Args:
        epoch (int):
            The current epoch number.
        domain (str):
            The domain of the current sample.
        sample_index (str):
            The index of the current sample.
        used_sample (Dict):
            The current sample being used for validation.
    """
    # initialize the sub-dict for each sample
    if sample_index not in self.epoch_records.keys():
        self.epoch_records[sample_index] = dict()

    # get the visualization logs for model snapshotting
    vis_logs = self.model(
        batch_data=used_sample,
        epoch=epoch,
        domain=domain,
        epoch_records=self.epoch_records,
        sample_index=sample_index,
        snapshot_interval=self.visual_snapshot_interval,
    )

    # put all the visualization logs into the queue
    self.enqueue(vis_logs)

monitor_init(args, model=None)

Initializes the validation monitor with the given arguments and the model. This method is responsible for setting up the general members, tracking best models, early stopping, and last models.

Parameters:

Name Type Description Default
args Namespace

Arguments provided for monitoring.

 required
model Model

The model being validated. This parameter must not be None.

 None

Raises:

Type Description
AssertionError

If the provided model is None.
Source code in speechain/monitor.py 
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
def monitor_init(self, args: argparse.Namespace, model: Model = None):
    """Initializes the validation monitor with the given arguments and the model.
    This method is responsible for setting up the general members, tracking best
    models, early stopping, and last models.

    Args:
        args (argparse.Namespace):
            Arguments provided for monitoring.
        model (Model, optional):
            The model being validated. This parameter must not be None.

    Raises:
        AssertionError: If the provided model is None.
    """
    assert model is not None, "Model must be provided and not None."
    # register a pointer of the model
    self.model = model

    # running mode
    self.dry_run = args.dry_run
    self.no_optim = args.no_optim
    self.mode = "valid"

    # best models-related members
    self.best_model_selection = args.best_model_selection
    # receive a single metric as a standalone list or tuple
    if isinstance(self.best_model_selection, (List, tuple)) and isinstance(
        self.best_model_selection[0], str
    ):
        self.best_model_selection = [self.best_model_selection]
    else:
        assert isinstance(self.best_model_selection, List), (
            f"best_model_selection must be given as a list, "
            f"but got type(best_model_selection)={self.best_model_selection}."
        )

    for i in range(len(self.best_model_selection)):
        # checking the argument types
        assert isinstance(self.best_model_selection[i], (List, tuple)), (
            "Each element of best_model_selection must be either a list or a tuple, "
            f"but got type={type(self.best_model_selection[i])}."
        )
        assert len(self.best_model_selection[i]) == 4, (
            f"Each element of best_model_selection must be a quad-tuple or qual-list, "
            f"but got length={len(self.best_model_selection[i])}."
        )

        if isinstance(self.best_model_selection[i], tuple):
            self.best_model_selection[i] = list(self.best_model_selection[i])
        self.best_model_selection[i][2] = self.best_model_selection[i][2].lower()
        assert self.best_model_selection[i][2] in [
            "max",
            "min",
        ], f"The best_model_mode must be either 'max' or 'min', but got {self.best_model_selection[i][2]}."

    # model saving-related members
    self.best_model_performance = dict()
    for metric in self.best_model_selection:
        self.best_model_performance["_".join(metric[:2])] = dict()
    self.saved_model_epoch = []
    self.model_save_path = os.path.join(self.result_path, "models")
    if not os.path.exists(self.model_save_path):
        os.makedirs(self.model_save_path, exist_ok=True)

    # early stopping-related members, the first metric in self.best_model_selection is used
    self.early_stopping_metric = "_".join(self.best_model_selection[0][:2])
    self.early_stopping_mode = self.best_model_selection[0][2]
    self.early_stopping_patience = args.early_stopping_patience
    self.early_stopping_threshold = args.early_stopping_threshold
    self.early_stopping_epochs = 0
    self.last_best_performance = (
        0.0 if self.early_stopping_mode == "max" else torch.inf
    )

    # last models-related members
    self.last_model_number = args.last_model_number
    if self.last_model_number < 1:
        raise ValueError(
            "last_model_number cannot be lower than 1, "
            "otherwise the training will not be able to resume."
            f"Got last_model_number={self.last_model_number}!"
        )

    # initialize the snapshooter of this validation monitor
    self.visual_snapshot_interval = args.visual_snapshot_interval

save_aver_model(epoch_message, metric_pop_flags)

Stores the average model in the ensemble and updates the epoch message based on the model performance metrics.

Parameters:

Name Type Description Default
epoch_message str

A string message related to the current epoch.

 required
metric_pop_flags Dict[str, bool]

A dictionary containing flags related to model performance metrics.

 required

Returns:

Name Type Description
str

The updated epoch message.
Source code in speechain/monitor.py 
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
def save_aver_model(self, epoch_message: str, metric_pop_flags: Dict[str, bool]):
    """Stores the average model in the ensemble and updates the epoch message based
    on the model performance metrics.

    Parameters:
        epoch_message (str):
            A string message related to the current epoch.
        metric_pop_flags (Dict[str, bool]):
            A dictionary containing flags related to model performance metrics.

    Returns:
        str: The updated epoch message.
    """

    def save_aver_models(
        aver_epoch_list: List, aver_num: int, aver_model_name: str
    ):
        # no average model is saved if there is only one candidate model
        if len(aver_epoch_list) == 1:
            return ""

        # sum up the parameters of all best models
        avg_model = None
        for epoch in aver_epoch_list:
            _tgt_model_path = os.path.join(
                self.model_save_path, f"epoch_{epoch}.pth"
            )
            # skip if the model doesn't exist
            if not os.path.exists(_tgt_model_path):
                continue

            _avg = None
            # access self.model_save_path from the outer scope
            if avg_model is not None:
                _avg = torch.load(_tgt_model_path, map_location="cpu")
            else:
                avg_model = torch.load(_tgt_model_path, map_location="cpu")

            if _avg is not None:
                for key in avg_model.keys():
                    avg_model[key] += _avg[key]

        # if no average model, skip this function and return an empty string
        if avg_model is None:
            return ""
        else:
            # for the parameters whose dtype is int, averaging is not performed
            # reference: https://github.com/espnet/espnet/blob/5fa6dcc4e649dc66397c629d0030d09ecef36b80/espnet2/main_funcs/average_nbest_models.py#L90
            for key in avg_model.keys():
                if not str(avg_model[key].dtype).startswith("torch.int"):
                    avg_model[key] /= aver_num

            # save the average model
            _aver_model_path = os.path.join(self.model_save_path, aver_model_name)
            torch.save(avg_model, _aver_model_path)

            return f"{aver_model_name} has been updated to the average of epochs {aver_epoch_list}.\n"

    # --- Save the average model for the best models of each metric --- #
    # loop each metric for best model selection
    for metric in self.best_model_selection:
        _metric_name, _metric_mode, _model_num = (
            "_".join(metric[:2]),
            metric[2],
            metric[3],
        )

        # average the recorded best models so far
        if (
            len(self.best_model_performance[_metric_name]) == _model_num
            and metric_pop_flags[_metric_name]
        ):
            epoch_message += save_aver_models(
                aver_epoch_list=list(
                    self.best_model_performance[_metric_name].keys()
                ),
                aver_num=len(self.best_model_performance[_metric_name]),
                aver_model_name=f"{_model_num}_{_metric_name}_average.pth",
            )

    # --- Save the average model of the last models --- #
    if self.epoch >= self.last_model_number:
        epoch_message += save_aver_models(
            aver_epoch_list=list(
                range(self.epoch, self.epoch - self.last_model_number, -1)
            )[::-1],
            aver_num=self.last_model_number,
            aver_model_name=f"{self.last_model_number}_last_average.pth",
        )

    return epoch_message

start_epoch(epoch)

Prepares the monitor for a new epoch of validation.

Parameters:

Name Type Description Default
epoch int

The current epoch number.

 required
Source code in speechain/monitor.py 
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
def start_epoch(self, epoch: int):
    """Prepares the monitor for a new epoch of validation.

    Args:
        epoch (int): The current epoch number.
    """
    # epoch-level information
    if epoch in self.best_model_performance.keys():
        self.logger.warning(
            f"The record of epoch no.{epoch} has already existed in the monitor! "
            f"It will be overwritten by the new record obtained shortly thereafter."
        )
    self.epoch = epoch
    self.epoch_start_time = time.time()

    # refresh the step-level records at the beginning of each epoch
    self.refresh_step_records()

    # logging the beginning information
    self.logger.info(f"The validation part of epoch no.{epoch} starts.")

state_dict()

Retrieves the state of the ensemble in dictionary form.

Returns:

Name Type Description
Dict

The current state of the ensemble, typically including model parameters and other related information.
Source code in speechain/monitor.py 
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
def state_dict(self):
    """Retrieves the state of the ensemble in dictionary form.

    Returns:
        Dict: The current state of the ensemble, typically including model parameters and other related information.
    """
    return dict(
        epoch_records=self.epoch_records,
        saved_model_epoch=self.saved_model_epoch,
        best_model_performance=self.best_model_performance,
        early_stopping_epochs=self.early_stopping_epochs,
        last_best_performance=self.last_best_performance,
    )

step(valid_metrics)

Records information for a single validation step.

Parameters:

Name Type Description Default
valid_metrics Dict[str, Tensor]

The validation metrics for the current step.

 required
Source code in speechain/monitor.py 
730
731
732
733
734
735
736
737
738
739
def step(self, valid_metrics: Dict[str, torch.Tensor]):
    """Records information for a single validation step.

    Args:
        valid_metrics (Dict[str, torch.Tensor]):
            The validation metrics for the current step.
    """
    # accumulate the values of validation criteria
    if valid_metrics is not None:
        self.record_step_info("criteria", valid_metrics)

update_best_and_pop_worst(epoch_message)

Updates the best model in the ensemble and removes the worst model. The best and worst are determined based on the performance metrics. The function also handles logging related to model performance.

Parameters:

Name Type Description Default
epoch_message str

A string message related to the current epoch.

 required

Returns:

Type Description

Tuple[str, bool, Dict[str, bool]]: Returns a tuple containing the updated epoch message, a flag indicating whether early stopping conditions are met, and flags related to model performance metrics.
Source code in speechain/monitor.py 
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
def update_best_and_pop_worst(self, epoch_message: str):
    """Updates the best model in the ensemble and removes the worst model. The best
    and worst are determined based on the performance metrics. The function also
    handles logging related to model performance.

    Parameters:
        epoch_message (str):
            A string message related to the current epoch.

    Returns:
        Tuple[str, bool, Dict[str, bool]]:
            Returns a tuple containing the updated epoch message, a flag indicating whether early stopping
            conditions are met, and flags related to model performance metrics.
    """

    def whether_remove(remove_epoch: int):
        # retain the last several models within self.last_model_number
        if self.epoch - remove_epoch < self.last_model_number:
            return False
        else:
            remove_flag = True
            # access metric_epoch_records from the outer scope
            for _epoch_record in metric_epoch_records.values():
                if remove_epoch in _epoch_record["sorted_epochs"]:
                    remove_flag = False
                    break
            return remove_flag

    # --- Gather the epoch record information for each metric --- #
    metric_epoch_records = dict()
    # loop each metric for best model selection
    for metric in self.best_model_selection:
        _metric_name, _metric_mode, _model_num = (
            "_".join(metric[:2]),
            metric[2],
            metric[3],
        )

        # find the best epoch and worst epoch in self.best_model_performance
        sorted_epochs = dict(
            sorted(
                self.best_model_performance[_metric_name].items(),
                key=lambda x: x[1],
                reverse=True if _metric_mode == "max" else False,
            )
        )
        metric_epoch_records[_metric_name] = dict(
            sorted_epochs=list(sorted_epochs.keys()),
            metric_mode=_metric_mode,
            model_num=_model_num,
        )

    # --- Pop out the worst model and Update the model symbol links --- #
    metric_pop_flags = dict()
    for metric_name, epoch_record in metric_epoch_records.items():
        # controls whether the worst model has been pooped out or not
        metric_pop_flags[metric_name] = False
        # pop out the worst model if there is a redundant one in self.best_model_performance
        if len(epoch_record["sorted_epochs"]) > epoch_record["model_num"]:
            # pick up the epoch number of the worst model
            worst_epoch = epoch_record["sorted_epochs"][-1]
            self.best_model_performance[metric_name].pop(worst_epoch)
            epoch_record["sorted_epochs"].remove(worst_epoch)
            metric_pop_flags[metric_name] = True

        # update the symbol links of all the best models so far
        for i, epoch in enumerate(epoch_record["sorted_epochs"]):
            _best_model_pointer = (
                f"{metric_name}_best.pth"
                if i == 0
                else f"{metric_name}_best_{i + 1}.pth"
            )
            # create a soft link from the best model pointer to the model fi le of the current epoch
            symlink_dst = os.path.join(self.model_save_path, _best_model_pointer)
            if os.path.islink(symlink_dst) or os.path.exists(symlink_dst):
                os.unlink(symlink_dst)
            os.symlink(
                os.path.join(self.model_save_path, f"epoch_{epoch}.pth"),
                symlink_dst,
            )

    # update the symbol links of the last several models
    for epoch in range(self.epoch, max(0, self.epoch - self.last_model_number), -1):
        _last_model_pointer = (
            "latest.pth"
            if epoch == self.epoch
            else f"last_{self.epoch - epoch + 1}.pth"
        )
        # create a soft link from the best model pointer to the model file of the current epoch
        symlink_dst = os.path.join(self.model_save_path, _last_model_pointer)
        if os.path.islink(symlink_dst) or os.path.exists(symlink_dst):
            os.unlink(symlink_dst)
        os.symlink(
            os.path.join(self.model_save_path, f"epoch_{epoch}.pth"), symlink_dst
        )

    # remove the redundant model files
    saved_epochs = self.saved_model_epoch.copy()
    for epoch in saved_epochs:
        epoch_model_path = os.path.join(self.model_save_path, f"epoch_{epoch}.pth")
        if whether_remove(epoch):
            # remove the record of epoch in the memory
            self.saved_model_epoch.remove(epoch)

            # remove the model file if it exists
            if os.path.exists(epoch_model_path):
                # ensure that the model to be removed is successfully removed
                while os.path.exists(epoch_model_path):
                    os.remove(epoch_model_path)

    # --- Early-Stopping epoch number checking for the early-stopping metric --- #
    if len(metric_epoch_records[self.early_stopping_metric]["sorted_epochs"]) != 0:
        best_epoch = metric_epoch_records[self.early_stopping_metric][
            "sorted_epochs"
        ][0]
        # refresh to 0 or add 1 depending on the comparison the best one and the second best one
        if best_epoch == self.epoch:
            epoch_message += f"{self.early_stopping_metric} of the current epoch no.{self.epoch} is the best so far.\n"

            # compare the current performance and the last best performance
            best_performance = self.best_model_performance[
                self.early_stopping_metric
            ][best_epoch]
            if self.is_better(
                best_performance,
                self.last_best_performance,
                mode=self.early_stopping_mode,
                threshold=self.early_stopping_threshold,
            ):
                epoch_message += (
                    f"The early-stopping threshold {self.early_stopping_threshold} is reached, "
                    "so the early-stopping epoch number is refreshed.\n"
                )
                self.early_stopping_epochs = 0
                self.last_best_performance = best_performance
            else:
                epoch_message += (
                    f"The early-stopping threshold {self.early_stopping_threshold} is not reached, "
                    "so the early-stopping epoch number keeps increasing.\n"
                )
                self.early_stopping_epochs += 1
        # directly add 1 if the current epoch is not the best
        else:
            epoch_message += f"No improvement of {self.early_stopping_metric} in the current epoch no.{self.epoch}.\n"
            self.early_stopping_epochs += 1

        # report the updated early-stopping epoch number
        epoch_message += f"The early-stopping epoch number has been updated to {self.early_stopping_epochs}.\n"

    # early-stopping check by the patience
    early_stopping_flag = False
    if (
        self.early_stopping_patience is not None
        and self.early_stopping_epochs > self.early_stopping_patience
    ):
        epoch_message += (
            f"The early-stopping patience {self.early_stopping_patience} is reached, "
            f"so the training process stops here.\n"
        )
        early_stopping_flag = True
    elif self.early_stopping_patience is None:
        epoch_message += (
            "The early-stopping patience is not set by your exp_cfg, "
            "so the training will continue until num_epochs is reached.\n"
        )

    return epoch_message, early_stopping_flag, metric_pop_flags

data_saving_logs(proc_id, logs_queue, wait_time=1)

Continuously monitors a queue for log data to save. If logs are available, they are retrieved from the queue and saved in the appropriate format. If the queue is empty, the function pauses for a specified wait time before checking again. In case of any exceptions during this process, a warning is issued and the function continues.

Parameters:

Name Type Description Default
proc_id int

The identifier of the current process.

 required
logs_queue Queue

The queue containing log data to be saved.

 required
wait_time int

The number of seconds to wait before checking the queue again if it is empty. Defaults to 1.

 1

Returns:

Type Description

None
Source code in speechain/monitor.py 
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
def data_saving_logs(proc_id: int, logs_queue: Queue, wait_time: int = 1):
    """Continuously monitors a queue for log data to save. If logs are available, they
    are retrieved from the queue and saved in the appropriate format. If the queue is
    empty, the function pauses for a specified wait time before checking again. In case
    of any exceptions during this process, a warning is issued and the function
    continues.

    Args:
        proc_id (int):
            The identifier of the current process.
        logs_queue (Queue):
            The queue containing log data to be saved.
        wait_time (int, optional):
            The number of seconds to wait before checking the queue again if it is empty. Defaults to 1.

    Returns:
        None
    """
    while True:
        try:
            # Continually check if the logs_queue has items to be saved
            if not logs_queue.empty():
                log = logs_queue.get()
                save_data_by_format(**log)
            else:
                # If the queue is empty, pause for a specified wait time
                time.sleep(wait_time)
        except Exception as e:
            warnings.warn(
                f"Process {proc_id} encountered an exception in data_saving_logs(): {e}. Continuing operation."
            )