PyTorch bindings for Warp-ctc

branch	status	pypi package
pytorch_bindings		-
pytorch-0.4		-
pytorch-1.0
pytorch-1.1

This is an extension onto the original repo found here.

Installation

Install PyTorch first.

for PyTorch 1.2

warp-ctc does not work with PyTorch 1.2. Use torch.nn.CTCLoss built in PyTorch 1.2.

for PyTorch 1.0 or 1.1

Install warpctc-pytorchXX-cudaYY wheel based on PyTorch and CUDA version of your environment.

# for PyTorch 1.1 and CUDA 10.1
$ pip install warpctc-pytorch11-cuda101

# for PyTorch 1.1 and CUDA 10.0
$ pip install warpctc-pytorch11-cuda100

# for PyTorch 1.1 and CUDA 9.2
$ pip install warpctc-pytorch11-cuda92

# for PyTorch 1.1 and CUDA 9.1
$ pip install warpctc-pytorch11-cuda91

# for PyTorch 1.1 and CUDA 9.0
$ pip install warpctc-pytorch11-cuda90

# for PyTorch 1.1 and CUDA 8.0
$ pip install warpctc-pytorch11-cuda80

# for PyTorch 1.0 and CUDA 10.1
$ pip install warpctc-pytorch10-cuda101

# for PyTorch 1.0 CUDA 10.0
$ pip install warpctc-pytorch10-cuda100

# for PyTorch 1.0 CUDA 9.2
$ pip install warpctc-pytorch10-cuda92

# for PyTorch 1.0 CUDA 9.1
$ pip install warpctc-pytorch10-cuda91

# for PyTorch 1.0 CUDA 9.0
$ pip install warpctc-pytorch10-cuda90

# for PyTorch 1.0 CUDA 8.0
$ pip install warpctc-pytorch10-cuda80

for PyTorch 0.4.1

Wheels for PyTorch 0.4.1 are not provided so users have to build from source manually.

WARP_CTC_PATH should be set to the location of a built WarpCTC (i.e. libwarpctc.so). This defaults to ../build, so from within a new warp-ctc clone you could build WarpCTC like this:

$ git clone https://github.com/espnet/warp-ctc.git
$ cd warp-ctc; git checkout -b pytorch-0.4 remotes/origin/pytorch-0.4
$ mkdir build; cd build
$ cmake ..
$ make

Now install the bindings:

$ cd ../pytorch_binding
$ pip install numpy cffi
$ python setup.py install

Example to use the bindings below.

import torch
from warpctc_pytorch import CTCLoss
ctc_loss = CTCLoss()
# expected shape of seqLength x batchSize x alphabet_size
probs = torch.FloatTensor([[[0.1, 0.6, 0.1, 0.1, 0.1], [0.1, 0.1, 0.6, 0.1, 0.1]]]).transpose(0, 1).contiguous()
labels = torch.IntTensor([1, 2])
label_sizes = torch.IntTensor([2])
probs_sizes = torch.IntTensor([2])
probs.requires_grad_(True)  # tells autograd to compute gradients for probs
cost = ctc_loss(probs, labels, probs_sizes, label_sizes)
cost.backward()

Documentation

CTCLoss(size_average=False, length_average=False, reduce=True)
    # size_average (bool): normalize the loss by the batch size (default: False)
    # length_average (bool): normalize the loss by the total number of frames in the batch. If True, supersedes size_average (default: False)
    # reduce (bool): average or sum over observation for each minibatch.
        If `False`, returns a loss per batch element instead and ignores `average` options.
        (default: `True`)

forward(acts, labels, act_lens, label_lens)
    # acts: Tensor of (seqLength x batch x outputDim) containing output activations from network (before softmax)
    # labels: 1 dimensional Tensor containing all the targets of the batch in one large sequence
    # act_lens: Tensor of size (batch) containing size of each output sequence from the network
    # label_lens: Tensor of (batch) containing label length of each example