Best Model for Eval.py

[mtkmark]

We got several out-come after training

1. pkl files contains variable for each layers/matrix , so we got best value after training. I understand.
2. for pth files, there are models description. why do we have "best" models . is it just one models ?

Thanks
Mark

-rw-rw-r-- 1 student38 student38 289623 Nov 16 01:19 infos_1-best.pkl
-rw-rw-r-- 1 student38 student38 170494134 Nov 16 01:19 model-cnn-best.pth
-rw-rw-r-- 1 student38 student38 78777561 Nov 16 01:19 model-best.pth
-rw-rw-r-- 1 student38 student38 908614 Nov 16 01:19 histories_1.pkl
-rw-rw-r-- 1 student38 student38 289623 Nov 16 01:19 infos_1.pkl
-rw-rw-r-- 1 student38 student38 338273633 Nov 16 01:19 optimizer-cnn.pth
-rw-rw-r-- 1 student38 student38 157522139 Nov 16 01:18 optimizer.pth
-rw-rw-r-- 1 student38 student38 170494134 Nov 16 01:18 model-cnn.pth
-rw-rw-r-- 1 student38 student38 78777561 Nov 16 01:18 model.pth

[connie980149]

pkl files only save the training information like iteration, epoch, best score...etc.
pth file is saving the parameters of models.
The parameters will be modified during training, the "best" model saves the weights when the validation score is the best.

[fansia]

Hi Connie,

Could we read training information like iteration/epoch/training loss etc from pkl? and How?

Thanks,
Eric

[connie980149]

After loading the pkl file (ex. infos = cPickle.load(f)), the infos is a dictionary of python, you can use, for example, infos['iter'] to get the iteration saved in pkl file.
Also you can get training loss from histories['loss_history'] in histories.pkl.