GPU training speed slower than CPU

[nebulasdice]

When i use tensorflow-gpu to train, the GPU utility is only 4%. and the training speed is slower than CPU.
How dose this happen? My CUDA and tensorflow-gpu haven been tested and they work just fine.

I did some search and one interpretation is that the data loading algorithm is not parallel so the GPU can not show its power. Is this the problem i'm facing?

[Kalelv45]

Since the main problem for me was the math error (out of reach). I decided to live everything as it was, and just change the sigmoid function. Now it goes way faster.

Don't know if it was caused by those imports (early stops and others), that weren't present in the siraj code.

Hope it helped.

[xtr33me]

It seems some logic I added for trying to incorporate Tensorboard into Keras via callbacks was slowing things down. I have gone about removing the code for now in the latest commit. Closing this.