re-train network

[Phoebe-star]

for your model , if I load resnet50 to extract feature and let the feature as input to another CNN . when training , the CNN is trainable , but the resnet50 is also trainable?? or the resnet50 is keeping resnet50.npy weight??

[liukingly]

In the backpropagation process, how is the error of the last hidden layer connected by the Inception model calculated? Is the sum of the errors of the four branches?