pytorch-ctc

Implementation of CTC (Connectionist Temporal Classification) beam search decoding with PyTorch bindings. C++ code borrowed liberally from TensorFlow with some improvements to increase flexibility.

Installation

The library is largely self-contained and requires only PyTorch and CFFI. Building the C++ library requires at least GCC-5. If gcc-5 or later is not your default compiler, you may specify the path via environment variables.

# get the code
git clone https://github.com/ryanleary/pytorch-ctc.git
cd pytorch-ctc

# install dependencies (PyTorch and CFFI)
pip install -r requirements.txt

# build the extension and install python package
# python setup.py install
CC=/path/to/gcc-5 CXX=/path/to/g++-5 python setup.py install

API

pytorch-ctc currently supports a single method:

output, score, out_seq_len = beam_decode(probs, seq_len=None, top_paths=1, beam_width=10,
                                         blank_index=0, merge_repeated=True)

where:

• probs is a FloatTensor of log-probabilities with shape (seq_len, batch_size, num_classes)
• seq_len is an optional IntTensor of integer sequence lengths with shape (batch_size)
• top_paths is used to specify how many hypotheses to return
• beam_width is the number of beams to evaluate in a given step
• blank_index is used to specify which position in the output distribution represents the blank class
• merge_repeated if True will collapse repeated characters

and returns:

• output is an IntTensor of character classes of shape (top_paths, batch_size, seq_len)
• score is a FloatTensor of log-probabilities representing the likelihood of the transcription with shape (top_paths, batch_size)
• out_seq_len is an IntTensor containing the length of the output sequence with shape (top_paths, batch_size)