Input pipeline: batches created without 'replacing' about deep-fluids

[byungsook]

Hi, I agree that sampling without replacement is ideally better in the short run.
In our case, many epochs are required to train the network, hence the current code won't show much difference in the final performance.

[tudormot]

Hi @byungsook ,
many thanks for the response! Yes, you are right for long training it should not quite matter.Sorry a quick last question, unfortunately I am also not able to get as fast inference time as you mention in the paper. Essentially, I am using exactly your implementation for training and test, and to measure the time I do (in line 340 in trainer.py):

for b in range(niter): start=time.time() G_ = self.sess.run(self.G_, {self.z: z_c[self.test_b_num*b:self.test_b_num*(b+1),:]}) G_, _ = self.batch_manager.denorm(x=G_) end=time.time() print(end-start)

This gives me on average between 150 and 300 ms on different smoke datasets while you have around 1 ms. I am using a Tesla P100-PCIE GPU, which shouldn't be very different from what you have used..
Could you tell me what is the catch I am missing?

Thanks,
Tudor

[byungsook]

Please take a look at #9