Closed milancurcic closed 1 year ago
cnn_mnist
converges with or without this fix but to a relatively low accuracy (~93% in 10 epochs), whereas it should easily get to >98%. While the fix introduced in this PR is necessary for updating conv2d
layers to work, there is another bug elsewhere that is causing bias and weight gradients to remain zero during training.
Conv2d layers were previously not getting their parameters updated during training.
cnn_mnist
now converges.