Improve the e2tomoseg_convnet.py and make training faster

[ZhenHuangLab]

I found out that the e2tomoseg_convnet.py could be improved by using model.fit. Training will be six times faster.

For example, when I ran the following command:

time e2tomoseg_convnet_test.py --trainset=particles/GCB_001_bin6_SIRT_preproc__good_2_trainset.hdf --nettag=convnet_iter100 --learnrate=0.0001 --niter=100 --ncopy=1 --batch=16 --nkernel=40,40,1 --ksize=15,15,15 --poolsz=2,1,1 --trainout --training --device=gpu

I could get the output as follows:

...
iteration 97, cost -5.999
iteration 98, cost -6.013
iteration 99, cost -6.013
Writting network output of training set to neuralnets/trainout_nnet_save__convnet_iter100.hdf...
Saving the trained net to neuralnets/nnet_save__convnet_iter100.hdf...
Done
Total time: 2589.5 s

real	43m12.468s
user	34m6.747s
sys	12m38.237s

When using the model.fit, the training will be much faster, and I can get almost the same results (like the loss, the network...)

Epoch 97/100
392/392 [==============================] - 4s 10ms/step - loss: -6.0035
Epoch 98/100
392/392 [==============================] - 4s 10ms/step - loss: -6.0049
Epoch 99/100
392/392 [==============================] - 4s 10ms/step - loss: -5.9938
Epoch 100/100
392/392 [==============================] - 4s 10ms/step - loss: -6.0119
Writting network output of training set to neuralnets/trainout_nnet_save__convnet_iter100.hdf...
Saving the trained net to neuralnets/nnet_save__convnet_iter100.hdf...
Done
Total time: 450.5 s

real	7m32.213s
user	5m22.877s
sys	1m22.681s

I can even set the callback parameter in model.fit() to set up some learning rate scheduler to improve the training process. So I think it is better to use model.fit instead of the cycle of model.train_on_batch.