PyTorch bindings for Warp-ctc
This is an extension onto the original repo found here.
Installation
Install PyTorch v0.4.
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/SeanNaren/warp-ctc.git
cd warp-ctc
mkdir build; cd build
cmake ..
makeNow install the bindings:
cd pytorch_binding
python setup.py installIf you try the above and get a dlopen error on OSX with anaconda3 (as recommended by pytorch):
cd ../pytorch_binding
python setup.py install
cd ../build
cp libwarpctc.dylib /Users/$WHOAMI/anaconda3/libThis will resolve the library not loaded error. This can be easily modified to work with other python installs if needed.
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)
# 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)
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