The Deep Neural Nets (DNN) library is a deep learning framework designed to be small in size, computationally efficient and portable.
We started the project as a fork of the popular OpenCV library, while removing some components that is not tightly related to the deep learning framework. Comparing to Caffe and many other implements, DNN is relatively independent to third-party libraries, (Yes, we don't require Boost and Database systems to be install before crafting your own network models) and it can be more easily portable to mobile systems, like iOS, Android and RaspberryPi etc. And more importantly, DNN is powerful! It supports both convolutional networks and recurrent networks, as well as combinations of the two.
The following features have been implemented:
- Mini-batch based learning, with OpenMP support
- YAML based network definition
- Gradient checking for all implemented layers
The following modules are implemented in current version:
| Module Name | Description |
|---|---|
InputLayer |
for storing original input images |
ConvolutionLayer |
performs 2d convolution upon images |
MaxPoolingLayer |
performs max-pooling operation |
DenseLayer |
fully connected Layer (optionally, perform activation and dropout) |
SimpleRNNLayer |
for processing sequence data |
MergeLayer |
for combining output results from multiple different layers |
More modules will be available online !
| Layer Type | Attributes |
|---|---|
Input |
name,n_input_planes,input_height,input_width,seq_length |
Convolution |
name,visualize,n_output_planes,ksize |
MaxPooling |
name,visualize,ksize |
SpatialTransform |
name,input_layer,n_output_planes,output_height,output_width |
Dense |
name,input_layer(optional),visualize,n_output_planes,activation |
TimeDistributed |
name,n_output_planes,output_height,output_width,seq_length,time_index |
SimpleRNN |
name,n_output_planes,seq_length,time_index,activation |
Merge |
name,input_layers,visualize,n_output_planes |
With the above parameters given in YAML format, one can simply define a network. For instance, a lenet model can be defined as:
%YAML:1.0
layers:
- {type: Input, name: input1, n_input_planes: 1, input_height: 28, input_width: 28, seq_length: 1}
- {type: Convolution, name: conv1, visualize: 0, n_output_planes: 6, ksize: 5, stride: 1}
- {type: MaxPooling, name: pool1, visualize: 0, ksize: 2, stride: 2}
- {type: Convolution, name: conv2, visualize: 0, n_output_planes: 16, ksize: 5, stride: 1}
- {type: MaxPooling, name: pool2, visualize: 0, ksize: 2, stride: 2}
- {type: Dense, name: fc1, visualize: 0, n_output_planes: 10, activation: softmax}Then, by ruuning network training program:
$ network train --solver data/mnist/lenet_solver.xmlone can start to train a simple network right away. And this is the way the source code and data models are tested in Travis-Ci. (See .travis.yml in the root directory)
CMake is required for successfully compiling the project.
Under root directory of the project:
$ cd $DNN_ROOT
$ mkdir build
$ cmake ..
$ make -j4MIT