cabbagito / convolutional-neural-networks-from-scratch Goto Github PK

During training the entire Validation Dataset is being shuffled and then the first couple of Data points are being chosen,
This is stupid.
The Validation Dataset should be shuffled beforehand and then random indices should be chosen.

The Convolution operation has been implemented naively using for loops, this is very inefficient and should be implemented in a vectorized form.

Stride doesn't work when calculating the gradient with respect to the input in the Convolution Operation.

Implement Dropout and Batch Norm.