losses

loss functions associated with ctc and cif

(Note)： ctc decoder binding from wenet runtime

TODO:

• entmax and entmax losses
• kd ctc decodernbest strategey
• suport batch ctc decode not parallel
• suport batch ctc decode parallel
• suport chunk state ctc decode
• suport torch sparse tensor
• sequence focal loss
• cross entropy focal loss
• sigmod focal loss
• focal logits for mwer
• mwer loss ssupport

import torch
from torch.nn.utils.rnn import pad_sequence

from ctcdecoder import CTCDecoder
from edit_distance import edit_distance

inputs = torch.tensor(
        [[[0.25, 0.40, 0.35],
           [0.40, 0.35, 0.25],
           [0.10, 0.50, 0.40]]]);
inputs = inputs.log()
seq_len = torch.tensor([3])
decoder = CTCDecoder(3,3)
print(decoder.decode(inputs, seq_len))
# print(pad_sequence(decoder.decode(inputs, seq_len), batch_first=True, padding_value=-1))

#tensor([[ 2,  1],
#        [ 1,  2],
#        [ 1, -1]])
#

hyp = torch.tensor([[1,2,3], [1,2,3]])
hyp_lens = torch.tensor([3,3])
truth = torch.tensor([[4,5,6], [4, 5, 6]])
t_lens = torch.tensor([3,3])

print(edit_distance(hyp,hyp_lens,truth, t_lens)

mwer=CTCMWERLoss(8)
labels=torch.tensor([[1,0,2]])
labels_length = torch.tensor([3])
print(mwer.forward(inputs, labels, labels_length, torch.tensor(3)))
#tensor(0.0136)