Hello everyone. A bit of a noob question but a little advice would be great.
I was trying to implement this paper Link. But using the architecture they mentioned(trying to convert it to pytorch) I do not see how they got an accuracy of 92%.
They did say they trained 10 hours. I tried for over 3 with LAMB as well as AdamW optimizer and I am still stuck at around 50%. Even the lr_find() plots barely any curves.
If you look at the images… it seems highly doubtful to me to give 92 with such a simple net.
Any help would be greatly appreciated.
model.add(layers.Conv2D(initial_depth, (3, 3), activation='relu', input_shape=(gen_h, gen_w, channels)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(initial_depth*2, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(initial_depth*4, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(initial_depth*4, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Flatten())
model.add(layers.Dense(512, activation='relu'))
if output_cls < 2:
model.add(layers.Dense(1, activation='sigmoid'))
else:
model.add(layers.Dense(output_cls, activation='softmax'))
