Deep Convolution Neural Network

#1

Hi Everyone!!

I am using VGGNET16 architecture for one of my dataset in order to predict the 3 labels. The issue is, val. Accuracy is not changing and remains constant throughout all the epochs.
I am still getting good test accuracy however if u look at the classification report it doesn’t make any sense.The architecture is not able to classify the labels properly. While for the same dataset, LaNet 5 and AlexNet yielding good results.

For VGGNET 16:
Epochs - 10
Batch Size - 128
Optimizer - Adam
Loss - Categorical_crossentropy
Input shape - (50,50,18)

I also used dropout(0.2) after every 3 convolution layer.

Relu for hidden layers and softmax for last layer.

Let me know asap what can I go in order to increase my val accuracy and a better classification report.

Thanks

#2

How big are your training, validation and test sets?

Also, please check the labels in all the sets, are they balanced?

While for the same dataset, LaNet 5 and AlexNet yielding good results.

Please use the model which is yielding better result. The neural networks are not interpretable.