how disable qa, lm and discriminator rewards?

[w82318029]

Thanks for sharing your code! I try to run your code with only adversarial discriminator. Would you please info me how to disable qa, lm and discriminator rewards? Thank you !

[tomhosking]

Have a look at these flags - if you set --lm_weight 0 --qa_weight 0 --disc_weight 1.0 --disc_train that should do what you're looking for.

[w82318029]

Thank you for your quick response.

[w82318029]

One more question. I try to run your code with my own data. so i need to train the discriminator from scratch. how to get word_embedding, char_embedding, word_dictionary and char_dictionary ? Thank you !

[tomhosking]

To be honest, I can't remember - but the discriminator was based on this QANet implementation. I think you'd need to follow the instructions on there. Or just use the embeddings/vocab I provide.

[w82318029]

thank you!

[w82318029]

sorry to bother you again. I set --lm_weight 0 --qa_weight 0 --disc_weight 1.0 --disc_train. I got error message after go through 215 step. I try with setting --lm_weight 0.25 --qa_weight 0.5 --disc_weight 1.0 --disc_train, I got same error message at 215 steps. Have you encountered same error?

[w82318029]

Model type is RL-S2S Loaded SQUAD with 75722 triples Modifying seq2seq model to incorporate rl rewards Building and loading lm Building and loading qa model Total number of trainable parameters: 788673 Total number of trainable parameters: 33135137 Total number of trainable parameters: 806785 loading wangyanmeng ./models/saved/discriminator Loading discriminator from ./models/saved/discriminator/model.checkpoint-4000 Traceback (most recent call last): File "/〇pt/conda/lib/python3.6/site-packages/tensorf1ow/python/cTient/session.py", line 1334, in _do_call return fn(*args) File "/〇pt/conda/lib/python3.6/site-packages/tensorf1ow/python/cTient/session.py", line 1319, in _run_fn options, feed_dict, fetch_list, target_Tist, run_metadata) File "/〇pt/conda/lib/python3.6/site-packages/tensorf1ow/python/cTient/session.py", line 1407, in _can_tf_sessionrun run_metadata) tensorflow.python.framework.errors_impl.invalidArgumentError: Reshape cannot infer the missing input size for an empty tensor unless all specified input sizes are non-zero

[w82318029]

It looks like error occur at the moment of loading discriminator. but it actually occurs when running to 215th step

[tomhosking]

Does it work with the following flags set? --lm_weight 0.25 --qa_weight 0.5 --disc_weight 1.0 --nodisc_train --lm_weight 0.25 --qa_weight 0.5 --disc_weight 0.0 --nodisc_train

The error mentions empty tensors, which sounds like there might be an issue with your dataset?

[w82318029]

Thank you for your suggestions. On squad1.1 datasets, it didnot work yesterday. Then, I tried --lm_weight 0.25 --qa_weight 0.5 --disc_weight 1.0 --nodisc_train it worked fine. And tried --lm_weight 0.25 --qa_weight 0.5 --disc_weight 1.0 --disc_train it also worked fine. maybe becasue “--nodisc_train“ changed something? saved a disc model ?

[tomhosking]

I'm not sure why that would help - but glad it's working now! The discriminator code was fairly late in the project so it's definitely not as robust as the rest.

[w82318029]

Thank you for your adivces. Would you please tell me how to fine tuning the a pretrained seq2seq network with only adversarial discriminator ? set flag restore=True ? anything else? Thanks a lot !

[tomhosking]

Have a look at these flags - if you set --lm_weight 0 --qa_weight 0 --disc_weight 1.0 --disc_train that should do what you're looking for.

This set of flags will disable the LM/QA rewards, just use the discriminator, and allow the discriminator to be trainable (so that it becomes adversarial).

[w82318029]

Sorry to bother you again. I got this error again "Reshape cannot infer the missing input size for an empty tensor unless all specified input sizes are non-zero". it seems like it related to low gpu memory. I use a v100 gpu with 16G memory. the code takes almost all of it. how much memory it takes when you run this code ? thanks a lot. any advice ?

[tomhosking]

Ah yes good point - you may need to reduce the batch size with --batch_size 8 or similar to get everything to fit in memory, the default batch size is for without the extra models.

[w82318029]

Thank you for the suggestion ! It works fine now.