cambridgeltl / sw_study Goto Github PK

Hi, I have a problem when using this method.
I tried to train English embeddings on a downsampled corpus (100
M tokens) by using bpe method.
My problem is for the when the code reaches this part in the train function:
if pairs:
while pairs:
total_loss += train_batch(m, optimizer, pairs[:m.bs], len(pairs[:m.bs]), m.neg_idxs)
.
.
The problem is, it is very slow, and takes too much time to be trained.
I tried to enable the Cuda flag, but it did not change anything. It only uses my GPU memory but the GPU volatile (GPU utilization) remains zero.
Do you know what might be wrong here, or is this method very time-consuming?

Hi, I am having a hard time finding the script (if there is one) that generates the file en.sent.1m.5.sms. I understand that the en.sent.1m has to go through CHIPMUNK, but en.sent.1m.5.sms is a vocabulary using the CHIPMUNK output. How can I get that? Thanks.