Verification noise (versus epoch) when using batch normalization
I am using the following model in Keras:
Input / conv1 / conv2 / maxpool / conv3 / conv4 / maxpool / conv5 / conv6 / maxpool / FC1 / FC2 / FC3 / softmax (2 nodes).
When I use Batch Normalization after each activation (Wx) and before ReLu nonlinearity (Wx), the loss and validation accuracy is noisy (Red = Training_set / Blue = validation_set):
If I remove the BN layers, then the validation loss is as smooth as the learning loss .
I tried the following (but didn't work):
1. Increase batch size from 64 to 256 2. Reduce learning rate 3. Add L2-reg and / or dropout of different amplitude 4. Train split / validation ratio: 20%, 30%. FYI, dataset are images of cats and images of dogs.
+3
source to share
No one has answered this question yet
Check out similar questions: