Created
July 31, 2019 10:13
-
-
Save tadas-subonis/795f20c5f2b4e549fa2aecc84d474db2 to your computer and use it in GitHub Desktop.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
from torch_baidu_ctc import ctc_loss | |
class CTCLossT(nn.Module): | |
def __init__(self): | |
super(CTCLossT, self).__init__() | |
self.blank = 0 | |
self.reduction = 'sum' | |
def forward(self, log_probs, targets): | |
#log_probs = log_probs.cpu() | |
#targets = targets.cpu() | |
#print(log_probs.shape, targets.shape) | |
#targets = targets.permute(1, 0) | |
batch_size = log_probs.size(1) | |
#print(log_probs) | |
#T = input_image_max_len | |
T = log_probs.size(0) | |
N = batch_size | |
D = targets.size(1) | |
input_lengths = torch.full(size=(N,), fill_value=D, dtype=torch.long) | |
target_lengths = torch.full(size=(N,), fill_value=D, dtype=torch.long) | |
with torch.backends.cudnn.flags(enabled=False): | |
loss = F.ctc_loss(log_probs, targets, input_lengths, target_lengths, self.blank, self.reduction, zero_infinity=True) | |
#print("loss", loss) | |
return loss |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment