Info on clock time per epoch (approximately)

[mrquincle]

Hi, thanks so much for this code!

It would be great to have some expected time each epoch should take.

Currently I have "ported" it to python3 at https://github.com/mrquincle/latent_3d_points. This is just a few changes: print() statements, making a few structs explicit, xrange -> range and basestring -> str. I did this because in the Colab notebook !find / -name cuda_config.h only shows the cuda files to be part of the python3 installation. Hence, it couldn't find the libraries when running using the python2 runtime.

Currently it runs (using the GPU which is selected) at a bit under a minute per epoch. Is this correct?

Epoch: 0001 training time (minutes)= 0.8922 loss= 0.092229053
INFO:tensorflow:data/single_class_ae/models.ckpt-1 is not in all_model_checkpoint_paths. Manually adding it.
Epoch: 0002 training time (minutes)= 0.8371 loss= 0.072633142
Epoch: 0003 training time (minutes)= 0.8422 loss= 0.068490269
Epoch: 0004 training time (minutes)= 0.8398 loss= 0.066228254
Epoch: 0005 training time (minutes)= 0.8405 loss= 0.065178742
Epoch: 0006 training time (minutes)= 0.8357 loss= 0.064000976
Epoch: 0007 training time (minutes)= 0.8380 loss= 0.062835155
Epoch: 0008 training time (minutes)= 0.8403 loss= 0.061971494
Epoch: 0009 training time (minutes)= 0.8401 loss= 0.060611139
Epoch: 0010 training time (minutes)= 0.8413 loss= 0.059806108

For 500 epochs this would be 0.85 * 500 minutes, which would be over 7 hours. Is that correct indeed?