search

bloomsburyai / question-generation

Neural text-to-text question generation
MIT License
217 stars 52 forks source link

Difference between MALUUBA and MALUUBA_RL #10

Closed rpoli40 closed 5 years ago

rpoli40 commented 5 years ago

Hi, I'm using commit be13417 and paid attention you made some changes later on. So what is the difference between MALUUBA and RL. Is it just a naming?

tomhosking commented 5 years ago

The difference is that the _RL model (now RL- models) include the placeholders and losses for policy gradient training. The difference between SEQ2SEQ and MALUUBA model types is the additional layer of condition encoding, which I replaced with the advanced_condition_encoding flag. If you want to fine tune the model then you need to use an RL model type (even for the first part of the training). Hope that helps.

rpoli40 commented 5 years ago

Once the latest commit is used when the training starts it prints that the model type is RL-S2S. Is it correct and then I'm using the approach described in the paper ("Machine comprehension by text-to-text neural question generation")

tomhosking commented 5 years ago

To replicate the approach used in the paper, run ./train.sh --advanced_condition_encoding --nocontext_as_set. You could also add the flag --model_type RL-MALUUBA if you wanted to give the run a more readable name, but this is optional.

rpoli40 commented 5 years ago

Thank you very much. Can you also list flags that should be used with other types of approaches you have.

On Fri, Dec 28, 2018 at 2:22 PM Tom Hosking notifications@github.com wrote:

To replicate the approach used in the paper, run ./train.sh --advanced_condition_encoding --nocontext_as_set. You could also add the flag --model_type RL-MALUUBA if you wanted to give the run a more readable name, but this is optional.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/bloomsburyai/question-generation/issues/10#issuecomment-450413440, or mute the thread https://github.com/notifications/unsubscribe-auth/Abba1G1scke94raoAmxznJqfYfv7Xz9Fks5u9m95gaJpZM4ZkD60 .

tomhosking commented 5 years ago

Check flags.py for a full list

rpoli40 commented 5 years ago

Hi Tom, Would you recommend retraining MPCM.py and language model if I would like to supplement Squad dataset with my own data. In case you recommend to retrain those are there any flags/arguments I should use? Thank you

On Fri, Dec 28, 2018 at 3:14 PM Tom Hosking notifications@github.com wrote:

Check flags.py for a full list

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/bloomsburyai/question-generation/issues/10#issuecomment-450420377, or mute the thread https://github.com/notifications/unsubscribe-auth/Abba1ChHCHutBhb1ReqIhsqunYtaiZr2ks5u9nu3gaJpZM4ZkD60 .