multiprocess training and training details

[zhyunlong]

How can i use multiprocess training? Can i use stable-baselines‘ multiprocessing interface？ And I also wonder training details to repeat the results of squence tagging. I run the script, and didn't get same good performance. Thanks a lot.

[rajcscw]

Hey, for multi-process training, refer to stable-baselines interface since Nlp-gym does not provide implementations of RL algorithms. Also, with respect to hyperparameter settings for DQN and PPO, please refer to our paper https://arxiv.org/pdf/2011.08272.pdf

[zhyunlong]

Thanks for your reply, I see hyperparameters in training scripts(train_seq_tagging.py), does it mean i can just run the script to get same result? How many steps does model need to learn? Very apprecaite for you reply.

[zhyunlong]

Thanks for your reply, I see hyperparameters in training scripts(train_seq_tagging.py), does it mean i can just run the script to get same result? How many steps does model need to learn? Very apprecaite for you reply.

[rajcscw]

Hey, you can train the agent for 1e+6 steps, you can do this as follows:

for i in range(int(1e+2)):
    model.learn(total_timesteps=int(1e+4), reset_num_timesteps=False)
    eval_model(model, env)

Also, make sure to use the PPO algorithm, which gave best results...

[zhyunlong]

Thanks a lot!

[zhyunlong]

Another question, in nlp-gym/nlp_gym/data_pools/custom_seq_tagging_pools.py, class 'CONLLNerTaggingPool' has no attribute '_get_dataset_from_corpus'. Do you forget to involve this code?

[xkianteb]

Below is the method that is missing:

`

@staticmethod
def _get_dataset_from_corpus(corpus, split):
    # from flair.datasets import CONLL_03
    corpus = datasets.CONLL_03()
    if split == 'train':
        return corpus.train
    elif split == 'val':
        return corpus.dev
    elif split == 'test':
        return corpus.test

`

[rajcscw]

Yes @zhyunlong, you are right, missed that function during refactoring. @xkianteb Thanks for the snippet, that is the missing implementation 👍

[xkianteb]

@rajcscw Would you be open to me being a contributor to the repo? I would like to add a few more tasks.

[rajcscw]

Sure @xkianteb, sounds like a good idea. What tasks do you have in mind? If you are on discord/twitter, feel free to reach me with rajkumar_rrk, we can have a quick chat..