Training hyperparameters for train_penn_multimodel.py

Hi, thank you very much for releasing your excellent code. I just quit don't understand why you set the initial_epoch = 7 here.

deephar/exp/pennaction/train_penn_multimodel.py

Line 150 in 6870fa3

trainer.train(30, steps_per_epoch=steps_per_epoch, initial_epoch=7,

As I understand, the PE is trained on MPII for 120 epochs, then weights related to PE are frozen and the AR is trained on Penn for 7 epochs, then the model weights are saved as weights_mpii+penn_ar_007.hdf5.

Following that, full model is built, loaded weight, jointly trained start from initial_epoch = 7.

deephar/exp/pennaction/train_penn_multimodel.py

Line 84 in 6870fa3

full_model.load_weights(

Please confirm that I am wrong or right? Thanks.

Hi @edwardnguyen1705 ,
I'm sorry about the confusion here. This file was committed after a local modification to continue a training process.
If you notice, the weights loaded in l84 are referring to epoch 7:

deephar/exp/pennaction/train_penn_multimodel.py

Line 86 in 6870fa3

    
           'output/penn_multimodel_trial_15_only_mpii_pose_be215a3/weights_mpii+penn_ar_007.hdf5',

That's why the first stage action training is commented (with pose layers frozen)

deephar/exp/pennaction/train_penn_multimodel.py

Line 143 in 6870fa3

# trainer.train(2, steps_per_epoch=steps_per_epoch, initial_epoch=0,

and the final training starts at epoch 7.

As reported in the paper, the first stage training is performed for 2 epochs on PA.

Hi @dluvizon,
Thanks for your explanation. I have conducted some experiments, and the results are consistent to those stated in your paper.