You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I wonder whether it is possible to initialize the filter weights from outer source. For examples, a network N_1 of architecture arch_1 has been trained , and then initialize another network N_2 of architecture arch_2 with the weights learned in N_1.
Alternatively, a python wrapper (or matlab) for saving the network state might allow to do it offline.
The text was updated successfully, but these errors were encountered:
sijinli
changed the title
Initialize the weight from outer source
Initialize the weights from outer source
May 25, 2014
This is finetuning: take pretrained model weights and a new architecture, then train the new architecture with weights initialized to the layers in common with the pretrained model. Finetuning is key for transfer learning and training problems where there's insufficient data to train from scratch. It's a first-class command in Caffe: finetune_net.bin solver_proto_file pretrained_net
I wonder whether it is possible to initialize the filter weights from outer source. For examples, a network N_1 of architecture arch_1 has been trained , and then initialize another network N_2 of architecture arch_2 with the weights learned in N_1.
Alternatively, a python wrapper (or matlab) for saving the network state might allow to do it offline.
The text was updated successfully, but these errors were encountered: