This is an extension onto the original repo found here.
Benchmarked on a dual-socket machine with two Intel E5-2660 v4 processors - warp-ctc used 10 threads to maximally take advantage of the CPU resources.
| T=150, L=40, A=28 | warp-ctc |
|---|---|
| N=1 | 1.89 ms |
| N=16 | 4.40 ms |
| N=32 | 6.39 ms |
| N=64 | 10.77 ms |
| N=128 | 19.69 ms |
| T=150, L=20, A=5000 | warp-ctc |
|---|---|
| N=1 | 10.22 ms |
| N=16 | 23.26 ms |
| N=32 | 44.70 ms |
| N=64 | 79.29 ms |
| N=128 | 146.83 ms |
Install PyTorch.
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 ..
makeOtherwise, set WARP_CTC_PATH to wherever you have libwarpctc.so
installed. If you have a GPU, you should also make sure that
CUDA_HOME is set to the home cuda directory (i.e. where
include/cuda.h and lib/libcudart.so live). For example:
export CUDA_HOME="/usr/local/cuda"
Now install the bindings:
cd pytorch_binding
python setup.py install
If 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/lib
This 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 torch.autograd import Variable
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 = Variable(torch.IntTensor([1, 2]))
label_sizes = Variable(torch.IntTensor([2]))
probs_sizes = Variable(torch.IntTensor([2]))
probs = Variable(probs, requires_grad=True) # tells autograd to compute gradients for probs
cost = ctc_loss(probs, labels, probs_sizes, label_sizes)
cost.backward()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