search

berleon / seqgan

Apache License 2.0
1 stars 1 forks source link

Questions about seqgan implementation #1

Open AlPros opened 6 years ago

AlPros commented 6 years ago

Good time of a day, Sir!

These days I am working with Generative Adversarial Networks, specifically - Sequential GAN. Your implementation is the only one using Keras libraries, which I could find as an example of application. As I understood, this particular seqGAN is used for text sequence generation. In my case, I would like to predict a numerical sequence.

I have some questions which might help me better understand the flow of the code: 1) Which data set has been utilised for testing this seqGAN? It would be very helpful for me to understand the internal dimensions of Generator and Discriminator networks. 1.1) If it is not possible to recover that testing data set, can You please provide some information about these lines of code from 'test_seqgan.py': output_size = 64 input_size = 64 z_size = 48 nb_chars = 32

Is nb_chars the number of possible labels? What is the input_size and output_size? Are these variables describe the length of sequences which are pushed into the input of generator and received after generation of data? Why does z_size is different from input or output size? As I know, z_size is usually a random noise which is used by generator for data generation?

2) In code I could not find any references which are addressed to file 'seq2seq'. How is that file involved in process of prediction? Moreover, it seems that these strings call for elements, which are not included in seq2seq library (It has been automatically installed from another github resource, mentioned in 'requirements.txt'): from seq2seq.layers.encoders import LSTMEncoder from seq2seq.layers.state_transfer_rnn import StateTransferLSTM

3) By this moment, I could find only one paper, related to sequential GAN, which is "seqGAN: Sequence Generative Adversarial Nets with Policy Gradient" by Lantao Yu et al. This implementation looks different, from the one which has been described in that paper. May You let me know what is the reference or paper for this implementation has been used so I can read and better understand the internal processes.

Thank You in advance, I am really appreciating Your sharing this implementation and Your help.

LuCeHe commented 6 years ago

Ive started to play with this code @AlPros, did you make any progress?

Arkamnd commented 5 years ago

Hi, could you please add a readme file?

Arkamnd commented 5 years ago

@AlPros @LuCeHe were you able to run this implementation? @berleon could you please add a readme file?

AlPros commented 5 years ago

@LuCeHe @Arkamnd Unfortunately, I could not make a significant progress with this code. Probably, You may use this code as some skeleton to get the overall idea. However, it seems that some of the modules are missing.