This PR creates a powerful Matlab interface for Caffe. It has been an issue for a long time in Caffe that Matlab interface is not as powerful as Python interface. #501 (followed by #1913) attempted to resolve this issue by adding more functions. However features in #1913 are still limited (e.g. not allowing creating multiple nets, not allowing training a net), and development of #1913 seems to have been stalled. Since #1913 is already called MatCaffe2, I'll call this PR MatCaffe3 :)

This Matlab interface of Caffe implements more features when compared with #1913, with comparable or even less code, by creating wrappers for caffe::Solver, caffe::Net, caffe::Layer and caffe::Blob in Matlab interface. In this interface, almost everything that can be done in PyCaffe can also be done here (except for a MATLAB_LAYER, which can be a future functionality).

This PR is add-only and non-invasive, which means the old Matlab interface is not removed in case someone still wants to use it. Note: the old Matlab wrapper doesn't work right now due to #1970. This PR removes the old Matlab wrapper as suggested by @shelhamer , but keeps the HDF5 example and modified the image classification example using BVLC CaffeNet.

Update message:

• caffe/matlab/+caffe/imagenet/ilsvrc_2012_mean.mat has been updated in Update ilsvrc_2012_mean.mat to W x H x C, update demo and add comments #2527 to contain mean_data in Width x Height x Channels with BGR channel order and single precision, which makes it consistent with caffe-supported data format. Previously the image_mean in ilsvrc_2012_mean.mat was in Height x Width x Channels with BGR, which was neither Matlab's image format nor caffe's data format, but somewhere in the middle, and inconsistent with read_mean.
• In this new interface, I did not provide specific mean subtraction functions, image preparation functions or image classification functions, as it depends on which method you are using. For mean subtraction, some methods do image mean subtraction (e.g. CaffeNet) and some do channel mean subtraction (e.g. VGG Net). For taking crops, CaffeNet first resize to 256x256 and takes 10 crops, while some method resize image to have min(h, w)=256 and take 10 crops from 256x256 central region, and some first resize to min(h, w)=256 and takes 10 crops from 4 conor + center (and flips) of the resized image instead its 256x256 central region. Since approach to prepare input can be arbitrary, instead of providing a series of such image preparing functions, I decided to leave that to users and illustrate it in caffe/matlab/classification_demo.m for CaffeNet.

MATLAB

The MATLAB interface -- matcaffe -- is the caffe package in caffe/matlab in which you can integrate Caffe in your Matlab code.

In MatCaffe, you can

• Creating multiple Nets in Matlab
• Do forward and backward computation
• Access any layer within a network, and any parameter blob in a layer
• Get and set data or diff to any blob within a network, not restricting to input blobs or output blobs
• Save a network's parameters to file, and load parameters from file
• Reshape a blob and reshape a network
• Edit network parameter and do network surgery
• Create multiple Solvers in Matlab for training
• Resume training from solver snapshots
• Access train net and test nets in a solver
• Run for a certain number of iterations and give back control to Matlab
• Intermingle arbitrary Matlab code with gradient steps

An ILSVRC image classification demo is in caffe/matlab/demo/classification_demo.m (you need to download BVLC CaffeNet from Model Zoo to run it).

Build MatCaffe

Build MatCaffe with make all matcaffe. After that, you may test it using make mattest.

Common issue: if you run into error messages like libstdc++.so.6:version 'GLIBCXX_3.4.15' not found during make mattest, then it usually means that your Matlab's runtime libraries do not match your compile-time libraries. You may need to do the following before you start Matlab:

export LD_LIBRARY_PATH=/opt/intel/mkl/lib/intel64:/usr/local/cuda/lib64
export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libstdc++.so.6

Or the equivalent based on where things are installed on your system, and do make mattest again to see if the issue is fixed. Note: this issue is sometimes more complicated since during its startup Matlab may overwrite your LD_LIBRARY_PATH environment variable. You can run !ldd ./matlab/+caffe/private/caffe_.mexa64 (the mex extension may differ on your system) in Matlab to see its runtime libraries, and preload your compile-time libraries by exporting them to your LD_PRELOAD environment variable.

After successful building and testing, add this package to Matlab search PATH by starting matlab from caffe root folder and running the following commands in Matlab command window.

addpath ./matlab

You can save your Matlab search PATH by running savepath so that you don't have to run the command above again every time you use MatCaffe.

Use MatCaffe

MatCaffe is very similar to PyCaffe in usage.

Examples below shows detailed usages and assumes you have downloaded BVLC CaffeNet from Model Zoo and started matlab from caffe root folder.

model = './models/bvlc_reference_caffenet/deploy.prototxt';
weights = './models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel';

Set mode and device

Mode and device should always be set BEFORE you create a net or a solver.

Use CPU:

caffe.set_mode_cpu();

Use GPU and specify its gpu_id:

caffe.set_mode_gpu();
caffe.set_device(gpu_id);

Create a network and access its layers and blobs

Create a network:

net = caffe.Net(model, weights, 'test'); % create net and load weights

Or

net = caffe.Net(model, 'test'); % create net but not load weights
net.copy_from(weights); % load weights

which creates net object as

  Net with properties:

           layer_vec: [1x23 caffe.Layer]
            blob_vec: [1x15 caffe.Blob]
              inputs: {'data'}
             outputs: {'prob'}
    name2layer_index: [23x1 containers.Map]
     name2blob_index: [15x1 containers.Map]
         layer_names: {23x1 cell}
          blob_names: {15x1 cell}

The two containers.Map objects are useful to find the index of a layer or a blob by its name.

You have access to every blob in this network. To fill blob 'data' with all ones:

net.blobs('data').set_data(ones(net.blobs('data').shape));

To multiply all values in blob 'data' by 10:

net.blobs('data').set_data(net.blobs('data').get_data() * 10);

Be aware that since Matlab is 1-indexed and column-major, the usual 4 blob dimensions in Matlab are [width, height, channels, num], and width is the fastest dimension. Also be aware that images are in BGR channels. Also, Caffe uses single-precision float data. If your data is not single, set_data will automatically convert it to single.

You also have access to every layer, so you can do network surgery. For example, to multiply conv1 parameters by 10:

net.params('conv1', 1).set_data(net.params('conv1', 1).get_data() * 10); % set weights
net.params('conv1', 2).set_data(net.params('conv1', 2).get_data() * 10); % set bias

Alternatively, you can use

net.layers('conv1').params(1).set_data(net.layers('conv1').params(1).get_data() * 10);
net.layers('conv1').params(2).set_data(net.layers('conv1').params(2).get_data() * 10);

To save the network you just modified:

net.save('my_net.caffemodel');

To get a layer's type (string):

layer_type = net.layers('conv1').type;

Forward and backward

Forward pass can be done using net.forward or net.forward_prefilled. Function net.forward takes in a cell array of N-D arrays containing data of input blob(s) and outputs a cell array containing data from output blob(s). Function net.forward_prefilled uses existing data in input blob(s) during forward pass, takes no input and produces no output. After creating some data for input blobs like data = rand(net.blobs('data').shape); you can run

res = net.forward({data});
prob = res{1};

Or

net.blobs('data').set_data(data);
net.forward_prefilled();
prob = net.blobs('prob').get_data();

Backward is similar using net.backward or net.backward_prefilled and replacing get_data and set_data with get_diff and set_diff. After creating some gradients for output blobs like prob_diff = rand(net.blobs('prob').shape); you can run

res = net.backward({prob_diff});
data_diff = res{1};

Or

net.blobs('prob').set_diff(prob_diff);
net.backward_prefilled();
data_diff = net.blobs('data').get_diff();

However, the backward computation above doesn't get correct results, because Caffe decides that the network does not need backward computation. To get correct backward results, you need to set 'force_backward: true' in your network prototxt.

After performing forward or backward pass, you can also get the data or diff in internal blobs. For example, to extract pool5 features after forward pass:

pool5_feat = net.blobs('pool5').get_data();

Reshape

Assume you want to run 1 image at a time instead of 10:

net.blobs('data').reshape([227 227 3 1]); % reshape blob 'data'
net.reshape();

Then the whole network is reshaped, and now net.blobs('prob').shape should be [1000 1];

Training

Assume you have created training and validation lmdbs following our ImageNET Tutorial, to create a solver and train on ILSVRC 2012 classification dataset:

solver = caffe.Solver('./models/bvlc_reference_caffenet/solver.prototxt');

which creates solver object as

  Solver with properties:

          net: [1x1 caffe.Net]
    test_nets: [1x1 caffe.Net]

To train:

solver.solve();

Or train for only 1000 iterations (so that you can do something to its net before training more iterations)

solver.step(1000);

To get iteration number:

iter = solver.iter();

To get its network:

train_net = solver.net;
test_net = solver.test_nets(1);

To resume from a snapshot "your_snapshot.solverstate":

solver.restore('your_snapshot.solverstate');

Input and output

caffe.io class provides basic input functions load_image and read_mean. For example, to read ILSVRC 2012 mean file (assume you have downloaded imagenet example auxiliary files by running ./data/ilsvrc12/get_ilsvrc_aux.sh):

mean_data = caffe.io.read_mean('./data/ilsvrc12/imagenet_mean.binaryproto');

To read Caffe's example image and resize to [width, height] and suppose we want width = 256; height = 256;

im_data = caffe.io.load_image('./examples/images/cat.jpg');
im_data = imresize(im_data, [width, height]); % resize using Matlab's imresize

Keep in mind that width is the fastest dimension and channels are BGR, which is different from the usual way that Matlab stores an image. If you don't want to use caffe.io.load_image and prefer to load an image by yourself, you can do

im_data = imread('./examples/images/cat.jpg'); % read image
im_data = im_data(:, :, [3, 2, 1]); % convert from RGB to BGR
im_data = permute(im_data, [2, 1, 3]); % permute width and height
im_data = single(im_data); % convert to single precision

Also, you may take a look at caffe/matlab/demo/classification_demo.m to see how to prepare input by taking crops from an image.

We show in caffe/matlab/hdf5creation how to read and write HDF5 data with Matlab. We do not provide extra functions for data output as Matlab itself is already quite powerful in output.

Clear nets and solvers

Call caffe.reset_all() to clear all solvers and stand-alone nets you have created.