How to use caffe Matlab Wrapper to finetune ConvNets

[Terry Chen]

Hi, I recently wanna to use caffe matlab wrapper to online finetune the alexnet or some others. Using new caffe Solver can directly train net but cannot finetune the net. Anyone know how to finetune Net in matlab? Thanks. 

[Devendra Mandan]

Maybe I don't entirely understand your problem, but you can follow the flickr style finetuning tutorial as it is, it will work, you don't need to use the matlab wrapper for this.

[Ronghang Hu]

You are able to use the matlab wrapper for fine-tuning. Here is the procedure:

solver = caffe.Solver('your_solver.prototxt');

solver.net.copy_from('your_pretrained_weights.caffemodel');

solver.solve();

[Terry Chen]

Thanks a lot. It works for me. 

[Terry Chen]

Yes,Ronghang's answer solve my problem. Thank you. 

[Terry Chen]

Hi, Ronghang, can I set data layer in train_val.prototxt as empty, just set as deploy file:

input: "data"

input_dim: 1

input_dim: 3

input_dim: 48

input_dim: 48

input: "label"

input_dim: 1

input_dim: 2

input_dim: 1

input_dim: 1

and use maltab to set data?  I have changed the train_val.prototxt like this, it make matlab crash.  How to do that? Thank you . 

On Sunday, August 2, 2015 at 10:26:23 AM UTC-7, Ronghang Hu wrote:

[Ronghang Hu]

No, I don't think you can do that right now in matlab wrapper. There is an interface to do that in python wrapper which allows you to use MemoryDataLayer and load data from numpy array, but this feature has not been implemented in matlab wrapper yet.

To overcome, you can resort to HDF5DataLayer (store your data as .h5 file), or use python wrapper instead.

[Devendra Mandan]

Sorry, but can you please answer my question also https://groups.google.com/forum/#!topic/caffe-users/e_xEICAwsdI

[Prophecies]

Can you tell me how to load pretrained weights when my new network contains pretrained network as only a part of whole thing?
e.g previous network say X, with pretrained weights X.caffemodel

My network

some layers -> X' -> some  more layers -> accuracy & loss where X' is X with input and output layers stripped.

In this case, how to selectively initialize the parts of network containing X with weights from X.caffemodel ?

Thanks in advance.

[Douglas Almonfrey]

Hi, is it possible to change data layer input size prior fini-tunning in matlab? I'm using windows data layer and I want to fine-tune the network for a 128x64 input size. Thanks!

[Royi]

Ronghang,

Do you use MATLAB on Windows?

Whenever I write solver = caffe.Solver('your_solver.prototxt'); MATLAB crashes.

Though for the same .prototxt the Command Line works perfectly.

Thank You.