zijianzhao / seqgan-pytorch

A implementation of SeqGAN in PyTorch, following the implementation in tensorflow.

SeqGAN-PyTorch

A implementation of SeqGAN in PyTorch, following the implementation in tensorflow.

Tested with:

• PyTorch v1 Stable
• Python 3.6
• CUDA at least 8.0 (For GPU)

Origin

The idea is from paper SeqGAN: Sequence Generative Adversarial Nets with Policy Gradient

The code is rewrited in PyTorch with the structure largely from Tensorflow Implementation

Runing

$ python main.py

After runing this file, the results will be printed on terminal. You can change the parameters in the main.py.

seqgan-pytorch's Issues

Why are you adding zero to your data and target?

Hello,
I have run your code, and I noticed that when you pretrain the generator, you prepend a zero matrix to the data and append a zero matrix to the target. I wonder what is the purpose of this operation, and how it affects the NLL loss calculation.
Also, in your adversarial training, when you use policy gradient for the generator, you still add zeros, but this time, you only replace the first element of the data (which is called ‘input’ in your function) with zero. Could you please explain these for me?
Thankyou. looking forward to your reply.

how to decode

Are the data in the eval.data the index of some dictionaries like "Oxford University Press" or something like that？

calculate perplexity and forward perplexity

How can I evaluate perplexity and forward perplexity of model?

Problem with running

Hi, thanks for the great code.

When I tried to run your vanilla example python main.py I got:

IndexError: invalid index of a 0-dim tensor. Use tensor.item() to convert a 0-dim tensor to a Python number

I'm running it under python 3.5.6 and pytorch 1.0.0

Can you get the results of the paper?

Can you get the results of the paper?There is a gap between the results I got and the results of the paper.Looking forward to your reply.

Bug on line 140 of main

arguments are incorrectly positioned. d_num_class is what you want to use for output dim ... I think, so you need to place that variable after num_filters and before dropout.

Right now you have it placed as the first input argument, which shifts all the following arguments into incorrect positions.

How to get real.data for my custom dataset?

I'd like to generate my custom data by SeqGAN, could you give more information about real.data file for that?

bug on 56 line of rollout.py

a subtle yet important mistake
the origin code:
'rewards[seq_len-1] += pred'
is likely to be changed as
'rewards[l-1] += pred'

loss.py

Hi, where is loss.py used?
I didn't see main.py import it .
you use nn.NLLLoss as loss function but loss.NLLLoss
thanks

MC rollout with variable sequences

Hi Zijian,
Thank you for the job. I have question that how to process variable sequences in MC rollout?
Thanks

Why do you use math.exp?

https://github.com/ZiJianZhao/SeqGAN-PyTorch/blob/master/main.py#L87

why?

In the original implementation by LantaoYu, exp is not used for loss values.

rewards

Hi,
I have a question that why does it perform the function torch.exe()
https://github.com/ZiJianZhao/SeqGAN-PyTorch/blob/master/main.py#L208

Thanks

in GAN training, rewards do not get reshaped correctly unless using cuda

rewards = rollout.get_reward(samples, 16, discriminator) rewards = Variable(torch.Tensor(rewards)) if args.cuda: rewards = torch.exp(rewards.cuda()).contiguous().view((-1,))
This is around lines 214-218, though maybe slightly different for you since I have edited the code.
I'm pretty sure that the reshaping of the rewards variable needs to happen regardless of whether or not cuda is being used.