mjc92 / gettothepoint Goto Github PK

the copydecoder not exists in the project

You are using softmax + NLLLoss
You should either use log_softmax + NLLLoss or softmax + CrossEntropyLoss
https://discuss.pytorch.org/t/is-log-softmax-nllloss-crossentropyloss/9352

Hi,

Thanks for the pytorch implementation. I just wanted to know how do I go about running this?
Considering that the numpy arrays are already saved, I just want to know what are the steps that have to be done next?

Hi,

Thanks for your work on implementing this model in PyTorch. Just wanted to know if this implementation managed to reproduce the results mentioned in the paper?

I am trying to train pointer generator network . After training for 10k iterations it starts overfitting.
Any suggestions on why this might be happening. coverage_loss is not used.
Note- 1==50 iterations. coverage_loss is not used.

when I use the package spacy to split words, I find the result has lot of error. maybe it will affect the final result.

Why it's multiplied by attn_mask instead of mask?

code :
while (epoch < args.epochs):
epoch += 1
random.shuffle(file_list)
print("length of file list is:",len(file_list))
for file in file_list:
opt.zero_grad()
with open(os.path.join(args.data_dir, file)) as f:
minibatch = f.read()
stories, summaries = batch.process_minibatch(minibatch, vocab)
out_list, cov_loss = model(stories, summaries, batch, vocab, True)
target = num_to_var(summaries[:, 2]).reshape(-1, 1)
target = pack(target, batch.output_lens.tolist(), batch_first=True)[0]
pad_out = pack(out_list, batch.output_lens.tolist(), batch_first=True)[0]

    pad_out = torch.log(pad_out)
    print("pad_out and target shapes:",pad_out.shape,target.shape)
   
    loss = criterion(pad_out, target) + cov_loss
    print('loss:',loss)

    loss.backward()
    opt.step()

beam search was not implemented in the code so i took one of the possible output just to train it.
forward pass it working and loss is calculating then during loss.backward() it shows the below error.

Traceback (most recent call last):
File "/home/m9506574/Downloads/GetToThePoint-master/shubh3_train.py", line 142, in
loss.backward()
File "/home/m9506574/anaconda3/envs/my_env/lib/python3.6/site-packages/torch/tensor.py", line 102, in backward
torch.autograd.backward(self, gradient, retain_graph, create_graph)
File "/home/m9506574/anaconda3/envs/my_env/lib/python3.6/site-packages/torch/autograd/init.py", line 90, in backward
allow_unreachable=True) # allow_unreachable flag
RuntimeError: select(): index 1 out of range for tensor of size [1, 100, 50020] at dimension 0

-> pack padded sequence i think it needs to be done before passing into the model and after getting the output from the model we have to do pad padded sequence. can someone explain it to me that why they are not doing pack padded sequence before passing it into the model and pad padded sequence was not implemented.