Hi, @karpathy
Maybe it is not appropriate to add the feature embedding layer in misc/net_utils.lua( Line 38 ). If the CNN part is not fine-tuned while training, the parameters of feature embedding layer are not updated. What's more the parameters here are initialized randomly.
One more question, How does neuraltalk2 initialize parameters of RNN part ? if I want to initialize them using uniform distribution, can I use code like this: params:uniform(-0.08, 0.08) ?
I am also confused about this. Maybe we should firstly train a model, then we initialize the parameters of feature embedding layer using the pre-trained model and only train the language model.
Hi, @karpathy Maybe it is not appropriate to add the feature embedding layer in
misc/net_utils.lua
( Line 38 ). If the CNN part is not fine-tuned while training, the parameters of feature embedding layer are not updated. What's more the parameters here are initialized randomly.One more question, How does neuraltalk2 initialize parameters of RNN part ? if I want to initialize them using uniform distribution, can I use code like this:
params:uniform(-0.08, 0.08)
?