search

facebookresearch / ParlAI

A framework for training and evaluating AI models on a variety of openly available dialogue datasets.
https://parl.ai
MIT License
10.49k stars 2.1k forks source link

will you give the best scripts to train the typical model such as seq2seq or transformer on each task/dataset in the ParlAI? #1468

Closed zhaoguangxiang closed 5 years ago

zhaoguangxiang commented 5 years ago

In my view, this can facilitate the research of dialogue systems.

stephenroller commented 5 years ago

Hi @zhaoguangxiang. One goal we have in the next few weeks is to actively beef up our model zoo. These models will come with the hyperparameters used to train them.

jaseweston commented 5 years ago

Yes, the goal is to have in parlai/projects tracking of best efforts on these tasks. For convai2 for example we have this here: https://github.com/facebookresearch/ParlAI/tree/master/projects/convai2 As Stephen says we do have things we are going to release into the model zoo soon. Note that this is also a community effort, where others can check in their best models, as there are many tasks...

On Wed, Feb 20, 2019 at 9:24 AM Stephen Roller notifications@github.com wrote:

Hi @zhaoguangxiang https://github.com/zhaoguangxiang. One goal we have in the next few weeks is to actively beef up our model zoo. These models will come with the hyperparameters used to train them.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/facebookresearch/ParlAI/issues/1468#issuecomment-465596555, or mute the thread https://github.com/notifications/unsubscribe-auth/AKjk-L_3U6ADsUXVF15zUi4KihZ2yCKpks5vPVqxgaJpZM4bFHHC .

stephenroller commented 5 years ago

Hi @zhaoguangxiang. While the zoo is never fully complete, there are some improvements from last week. Most of these have the training scripts available.

stephenroller commented 5 years ago

(Feel free to reopen for specific requests or additional questions)

JohannesTK commented 5 years ago

Have you thought of having different hyperparameters sets like Tensor2Tensor? I.e. transformer_base, transformer_tiny, etc. Feels like this would vastly speed up the experimentation and training on new datasets because it would take out a lot of the guessing game.