Poor result with Keras inbuilt VGG16 model

Part 1 (2017)
1

Hi,

I tried to solve Lesson 1 with the inbuilt Keras VGG16 model, but the results are absolutely terrible!

I get 52% accuracy which is barely better than complete hazard.

Any idea why this would be the case? Notebook

Vlad

2

You are using tensorflow which makes it a bit harder to compare directly to the theano version. However, it looks like you only ran one epoch. Did you try changing the epochs to something more like nb_epoch=5? Additionally you can adjust the learning rate once you start making progress.

3

Hi,

It was actually the learning rate that was the problem! I get ~96% after an epoch now, thank you!

1 Like
4

@jimanvlad can you share what you changed it to that helped?