Bases: Criterion
This criterion calculates the accuracy rate (0.0~1.0) between model predictions
and target labels.
This criterion doesn't have initialization function.
Source code in speechain/criterion/accuracy.py
| class Accuracy(Criterion):
"""This criterion calculates the accuracy rate (0.0~1.0) between model predictions
and target labels.
This criterion doesn't have initialization function.
"""
def __call__(
self, logits: torch.Tensor, text: torch.Tensor, text_len: torch.Tensor
):
"""
Args:
logits: (batch, logits_maxlen, vocab_size)
The token predictions from the model
text: (batch, text_maxlen)
The ground-truth token sequences
text_len: (batch,)
The length for the token predictions.
Returns:
The accuracy of the token predictions.
"""
# For the text attached by a <sos/eos> at the beginning
if logits.size(1) == text.size(1) - 1:
# text_len must match the sequence dimension of text
assert text_len.max() == text.size(1), (
f"There is a mismatch of the sentence length between text and text_len. "
f"Expect text_len.max() is either equal to or 1 smaller than text.size(1), "
f"but got text_len.max()={text_len.max()} and text.size(1)={text.size(1)}."
)
# remove the <sos/eos> at the beginning
text = text[:, 1:].squeeze(dim=-1)
# don't use text_len -= 1 here because it will also change the text_len outside this function
text_len = text_len - 1
# Otherwise, text must not have a <sos/eos> at the beginning (equal in length with logits)
elif logits.size(1) != text.size(1):
raise RuntimeError
# mask generation for the input text length
text_mask = make_mask_from_len(text_len).squeeze(dim=1)
if text.is_cuda:
text_mask = text_mask.cuda(text.device)
# calculate the accuracy by the correct prediction
if logits.dim() == text.dim() + 1:
logits = logits.argmax(dim=-1)
elif logits.dim() != text.dim():
raise RuntimeError(
f"logits.shape={logits.shape} but text.shape={text.shape}!"
)
correct_num = logits.eq(text).masked_select(text_mask).sum()
total_num = text_len.sum()
return correct_num / total_num
|
__call__(logits, text, text_len)
Parameters:
| Name |
Type |
Description |
Default |
logits
|
Tensor
|
(batch, logits_maxlen, vocab_size)
The token predictions from the model
|
required
|
text
|
Tensor
|
(batch, text_maxlen)
The ground-truth token sequences
|
required
|
text_len
|
Tensor
|
(batch,)
The length for the token predictions.
|
required
|
Returns:
| Type |
Description |
|
|
The accuracy of the token predictions.
|
Source code in speechain/criterion/accuracy.py
| def __call__(
self, logits: torch.Tensor, text: torch.Tensor, text_len: torch.Tensor
):
"""
Args:
logits: (batch, logits_maxlen, vocab_size)
The token predictions from the model
text: (batch, text_maxlen)
The ground-truth token sequences
text_len: (batch,)
The length for the token predictions.
Returns:
The accuracy of the token predictions.
"""
# For the text attached by a <sos/eos> at the beginning
if logits.size(1) == text.size(1) - 1:
# text_len must match the sequence dimension of text
assert text_len.max() == text.size(1), (
f"There is a mismatch of the sentence length between text and text_len. "
f"Expect text_len.max() is either equal to or 1 smaller than text.size(1), "
f"but got text_len.max()={text_len.max()} and text.size(1)={text.size(1)}."
)
# remove the <sos/eos> at the beginning
text = text[:, 1:].squeeze(dim=-1)
# don't use text_len -= 1 here because it will also change the text_len outside this function
text_len = text_len - 1
# Otherwise, text must not have a <sos/eos> at the beginning (equal in length with logits)
elif logits.size(1) != text.size(1):
raise RuntimeError
# mask generation for the input text length
text_mask = make_mask_from_len(text_len).squeeze(dim=1)
if text.is_cuda:
text_mask = text_mask.cuda(text.device)
# calculate the accuracy by the correct prediction
if logits.dim() == text.dim() + 1:
logits = logits.argmax(dim=-1)
elif logits.dim() != text.dim():
raise RuntimeError(
f"logits.shape={logits.shape} but text.shape={text.shape}!"
)
correct_num = logits.eq(text).masked_select(text_mask).sum()
total_num = text_len.sum()
return correct_num / total_num
|