Problems about training related.

[xudif]

Dear authors:
Sorry to bother you again.
I had converted the data format and was ready to start training, I have some questions to ask.

1. What should the training process look like? Is it to fineturn all Pasta and verbs in one training, like Module_trained:['foot','leg','hip','hand', arm','head','verb'] ?
2. The “Partrelevation Predictor”mentioned in the paper I did not find in the model-related code. Can you tell me exactly where that is?

[Hongwei Fan]

Training procedure

First, we train the PaSta feature by finetuning all of the PaSta classifiers i.e. ['foot', ..., 'head'] in MODULE_TRAINED (you can choose to freeze the backbone or not). Then, we lock the weights before PaSta feature and only finetune the verb classifier i.e. ['verb']. With this two-stage strategy, we are able to infer the verb compositionally.

"Part Relevation Predictor"

We have simplified the A2V model in this version to show the capability of our PaSta-based Activity Reasoning module. The implementation of "Part Relevation Predictor" can be found here.

[xudif]

Thanks！