search

jingyang2017 / AU-Net

Towards robust facial action units detection
MIT License
18 stars 3 forks source link

Questions about the pipeline of the essay #2

Open HZIH opened 3 months ago

HZIH commented 3 months ago

The author has done a really great job, I am wondering whether the sequence of the training phase is

  1. Train AUnet w/o VAE
  2. Train VAE
  3. Train AUnet+VAE
jingyang2017 commented 3 months ago

I think you are correct:

  1. Train AUnet w/o VAE
  2. Train VAE
  3. Train AUnet+VAE (load 1 or not)
  4. Train AUnet+VAE+TDN (load 3 or not)
HZIH commented 3 months ago

What's the meaning for (load 1 or not) and (load 3 or not). I guess you mean that load the weight in the first step and the weight in the third step. Moreover, If your pipeline is what we mentioned above, the code for function data_load is not correct, especially in the train_vae.py. The dataLoader there return data and label, but actually the data_load function should output a dict.

jingyang2017 commented 3 months ago

Yes. It is for load weights. The current data_load is for training AUnet+VAE+TDN. If you want to return data and label, you can rewrite the code.