Simple implementation of the Connectionist Temporal Classification loss

Implementation as described in the paper: "Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks", Graves et al.: https://dl.acm.org/doi/abs/10.1145/1143844.1143891

Should be used for experimentation puposes.

Install requirements

pip install setuptools
pip install -e .

Run CTC loss test:

python3 ctc/run_test.py

Inputs info:

Results summary:

Acoustic emissions matrix shape:
        torch.Size([125, 36])

Label to ID dictionary:
        {'<s>': 0, '<pad>': 1, '</s>': 2, '<unk>': 3, '|': 4, 'E': 5, 'A': 6, 'I': 7, 'N': 8, 'U': 9, 'T': 10, 'R': 11, 'C': 12, 'O': 13, 'Ă': 14, 'S': 15, 'M': 16, 'D': 17, 'L': 18, 'P': 19, 'V': 20, 'Ș': 21, 'Ț': 22, 'F': 23, 'Î': 24, 'B': 25, 'Z': 26, 'G': 27, 'Â': 28, 'H': 29, 'X': 30, 'K': 31, 'J': 32, 'Y': 33, 'W': 34, 'Q': 35}

Transcript text:
        ÎȚI MULȚUMESC MULT

Transcript labels:
        ['<s>', 'Î', '<s>', 'Ț', '<s>', 'I', '<s>', '|', '<s>', 'M', '<s>', 'U', '<s>', 'L', '<s>', 'Ț', '<s>', 'U', '<s>', 'M', '<s>', 'E', '<s>', 'S', '<s>', 'C', '<s>', '|', '<s>', 'M', '<s>', 'U', '<s>', 'L', '<s>', 'T', '<s>', '|', '<s>']
Viterbi (argmax) decoding labels:
        ['<s>', 'Î', '<s>', 'Ț', '<s>', 'I', '<s>', '|', '<s>', 'M', '<s>', 'U', '<s>', 'L', '<s>', 'Ț', '<s>', 'U', '<s>', 'M', '<s>', 'E', '<s>', 'S', 'C', '<s>', '|', '<s>', 'M', '<s>', 'U', '<s>', 'L', '<s>', 'T', '<s>', '|', '<s>']

Output results:

Case name	Forward Viterbi negative loglikelihood	Backwrd Viterbi negative loglikelihood	Torch CTC loss	Implemented CTC Loss	Gradient matrix norm	Joint probability norm	Emissions Gradient norm
Half of IDs are wrong	-1343.07	-1132.36	1006.6	1343.07	nan	1.41227	nan
No blank separators	-44.3633	-44.3633	44.3633	44.3633	2.54267	1.40262e+07	1.86154
Transcript IDs	22.9122	22.9123	-22.9123	-22.9122	1.52612	5.66188	1.5262
Transcript IDs with max probabilities == 1	110.704	110.704	-110.704	-110.704	36.2394	0.573886	36.2393
All IDs are the same	nan	nan	986.052	nan	nan	nan	nan

References:

• "Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks", Graves et al.: https://dl.acm.org/doi/abs/10.1145/1143844.1143891
• "Sequence Modeling With CTC", Hannun et al.: https://distill.pub/2017/ctc/
• CTC implementations:
  • https://github.com/amaas/stanford-ctc
  • https://github.com/vadimkantorov/ctc