Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Added feature to allow different learning rates per layer in the NN #143

Open
wants to merge 4 commits into
base: master
Choose a base branch
from

Conversation

Tgaaly
Copy link

@Tgaaly Tgaaly commented May 17, 2015

Added feature to allow different learning rates per layer in the NN. This feature is useful for transfer learning where you pre-training parts of the NN and then fine-tuning additional layers on top. The learning rates should be higher for the new layers and lower for the pre-trained layers. This is similar to what was done here: http://caffe.berkeleyvision.org/gathered/examples/finetune_flickr_style.html

By default the nn.learningRatePerLayer=[] and this will not cause an error in the default case as I check to see if its empty. The changes are very simple and straight forward.

Tgaaly added 3 commits May 16, 2015 20:51
… useful for transfer learning, pre-training parts of the NN and fine-tuning other parts
… useful for transfer learning, pre-training parts of the NN and fine-tuning other parts - updated nntrain.m - where learningRatePerLayer is also scaled with the nn.scaling_learningRate
… useful for transfer learning, pre-training parts of the NN and fine-tuning other parts - error fix
@oxinabox
Copy link

This is really useful.

In my PR: #128
for DBNs it is also supported.
Because DBNs are RBMs and each RBM has a learning rate (alpha) parameter (in my PR, before I believe it was specced for DBNs in opts to the training function).

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

None yet

2 participants