Probability distributions over sequences in pytorch and cupy.
pip install seqdist
Comparison against builtin pytorch implementation of the standard CTC loss:
sample_inputs = logits, targets, input_lengths, target_lengths = ctc.generate_sample_inputs(T_min=450, T_max=500, N=128, C=20, L_min=80, L_max=100)
print('pytorch loss: {:.4f}'.format(ctc.loss_pytorch(*sample_inputs)))
print('seqdist loss: {:.4f}'.format(ctc.loss_cupy(*sample_inputs)))
pytorch loss: 12.8080
seqdist loss: 12.8080
Pytorch:
report(benchmark_fwd_bwd(ctc.loss_pytorch, *sample_inputs))
fwd: 4.79ms (4.17-5.33ms)
bwd: 9.69ms (8.33-10.88ms)
tot: 14.47ms (12.67-16.20ms)
Seqdist:
report(benchmark_fwd_bwd(ctc.loss_cupy, *sample_inputs))
fwd: 7.22ms (6.78-7.85ms)
bwd: 6.21ms (5.82-8.57ms)
tot: 13.43ms (12.63-16.41ms)
betas = [0.1, 1.0, 10.]
alignments = {'beta={:.1f}'.format(beta): to_np(ctc.soft_alignments(*sample_inputs, beta=beta)) for beta in betas}
alignments['viterbi'] = to_np(ctc.viterbi_alignments(*sample_inputs))
fig, axs = plt.subplots(2, 2, figsize=(15, 8))
for (ax, (title, data)) in zip(np.array(axs).flatten(), alignments.items()):
ax.imshow(data[:, 0].T, vmax=0.05);
ax.set_title(title)