Commits on Mar 20, 2019

1. Fix model export for DocClassificationModel with dense features (face… …

   …bookresearch#394)
   
   Summary:
   Pull Request resolved: facebookresearch#394
   
   DocClassificationTask supports dense features, but exporting a DocClassification model with dense features to caffe2 doesn't work. Also, exporting any model that uses both text features (word/char) and dense features together doesn't work. This diff fixes export for dense features.
   
   There are two things that needed fixing:
   
   - exporter puts model inputs in a list like this: `[features, feature lengths]`. However, dense features don't go through the representation layer - so they need to be hardcoded to be at the end of model inputs (ugh). The order of features needs to be `[features, feature lengths, dense features]`
   - dummy_model_input for all other fields has two entries. dummy_model_input for dense_features has one entry
   
   In the long run, exporter should be tightly coupled with the model itself - we should enforce that model inputs have the same order while training and exporting. But that requires a lot of refactoring.
   
   This is a blocking feature for [on-device M-suggestions](https://fb.workplace.com/groups/323789728341681/permalink/326686971385290/). A public test for that is scheduled next week.
   Personalization models in M-suggestions need dense features.
   
   Reviewed By: snisarg, gardenia22
   
   Differential Revision: D14473020
   
   fbshipit-source-id: 8d26ff3ffd28087d6172d66f39d1d3f5f8a0ee9b

   

   Kshitiz Malik authored and facebook-github-bot committed Mar 20, 2019
   Configuration menu

   • View commit details
   • Copy full SHA for db704e5
   • Browse repository at this point

   Copy the full SHA

   db704e5 View commit details

   Browse the repository at this point in the history