用于深度学习网络的简单MATLAB工具箱

# SimpleDeepNetToolbox : Simple Deep Net Toolbox in MATLAB ---------- Authors: [Hiroyuki Kasai](http://kasai.kasailab.com/) Last page update: November, 14, 2018 Latest library version: 1.0.3 (see Release notes for more info.) <br /> Introduction ---------- The SimpleDeepNetToolbox is a **pure-MATLAB** and simple toolbox for deep learning. This toolbox was originally ported from [python library](https://github.com/oreilly-japan/deep-learning-from-scratch). However, major modification have been made for MATLAB implementation and its efficient implementation. There are much better other toolboxes available for deep learning, e.g. Theano, torch or tensorflow. I would definitely recommend you to use one of such tools for your problems at hand. The main purpose of this toolbox is to allows you, especially "MATLAB-lover" researchers, to undedrstand deep learning techniques using "non-black-box" simple implementations. <br /> ## <a name=""> List of available network architectures</a> - **Feedforward Backpropagation Neural Networks** - **Convolutional Neural Networks** ## <a name=""> List of available layers</a> - **Affine layer** - **Conolution layer** - **Pooling layer** - **Dropout layer** - **Batch normalization layer** (Under construction) - **ReLu (Rectified Linear Unit) layer** - **Sigmoid layer** - **Softmax layer** ## <a name=""> List of available optimization solvers</a> - **Vanila SGD** - **AdaGrad** - **Momentum SGD** <br /> Folders and files --------- <pre> ./ - Top directory. ./README.md - This readme file. ./run_me_first.m - The scipt that you need to run first. ./demo.m - Demonstration script to check and understand this package easily. ./download.m - Script to download datasets. |networks/ - Contains various network classes. |layers/ - Contains various layer classes. |optimizer/ - Contains optimization solvers. |test_samples/ - Contains test samples. |datasets/ - Contains dasetsets (to be downloaded). </pre> <br /> First to do: configure path ---------------------------- Run `run_me_first` for path configurations. ```Matlab %% First run the setup script run_me_first; ``` Second to do: download datasets ---------------------------- Run `download` for downloading datasets. ```Matlab %% Run the downloading script download; ``` - If your computer is behind a proxy server, please configure your Matlab setting. See [this](http://jp.mathworks.com/help/matlab/import_export/proxy.html?lang=en). <br /> Simplest usage example: 5 steps! ---------------------------- Just execute `demo_two_layer_neuralnet` for the simplest demonstration of this package. This is a forward backward neural network. ```Matlab %% load dateaset [x_train, t_train, train_num, x_test, t_test, test_num, class_num, dimension, ~, ~] = ... load_dataset('mnist', './datasets/', inf, inf, false); %% set network network = two_layer_net(x_train, t_train, x_test, t_test, 784, 50, 10, []); %% set trainer trainer = nn_trainer(network); %% train info = trainer.train(); %% plot display_graph('epoch', 'cost', {'Tow layer net'}, {}, {info}); train_info = info; test_info = info; train_info.accuracy = info.train_acc; test_info.accuracy = info.test_acc; display_graph('epoch', 'accuracy', {'Train', 'Test'}, {}, {train_info, test_info}); ``` <br /> <br /> Let's take a closer look at the code above bit by bit. The procedure has only **5 steps**! **Step 1: Load dataset** First, we load a dataset including train set and test set using a data loader function `load_dataset()`. The output include train set and test set, and related other data. ```Matlab [x_train, t_train, train_num, x_test, t_test, test_num, class_num, dimension, ~, ~] = ... load_dataset('mnist', './datasets/', inf, inf, false); ``` **Step 2: Set network** The next step defines the network architecture. This example uses a two layer neural network with the input size 784, the hidden layer size 50, and the output layer size 10. Datasets are also delivered to this class. ```Matlab %% set network network = two_layer_net(x_train, t_train, x_test, t_test, 784, 50, 10, []); ``` **Step 3: Set trainer** You also set the network to be used. Some options for training could be configured using the second argument, which is not used in this example, though. ```Matlab %% set trainer trainer = nn_trainer(network); ``` **Step 4: Perform trainer** Now, you start to train the network. ```Matlab %% train info = trainer.train(); ``` It returns the statistics information that include the histories of epoch numbers, cost values, train and test accuracies, and so on. **Step 5: Show result** Finally, `display_graph()` provides output results of decreasing behavior of the cost values in terms of the number of epoch. The accuracy results for the train and the test are also shown. ```Matlab % plot display_graph('epoch', 'cost', {'Two layer net'}, {}, {info}); train_info = info; test_info = info; train_info.accuracy = info.train_acc; test_info.accuracy = info.test_acc; display_graph('epoch', 'accuracy', {'Train', 'Test'}, {}, {train_info, test_info}); ``` That's it! <br /> More plots ---------------------------- TBA. <br /> License ------- - The SimpleDeepNetToolbox is **free** and **open** source. - The code provided in SimpleDeepNetToolbox should only be used for **academic/research purposes**. - This toolbox was originally ported from [python library](https://github.com/oreilly-japan/deep-learning-from-scratch). <br /> Notes ------- - As always, parameters such as the learning rate should be configured properly in each dataset and network. <br /> Problems or questions --------------------- If you have any problems or questions, please contact the author: [Hiroyuki Kasai](http://kasai.kasailab.com/) (email: kasai **at** is **dot** uec **dot** ac **dot** jp) <br /> Release notes -------------- * Version 1.0.3 (Nov. 14, 2018) - Some class structures are re-configured. * Version 1.0.2 (Nov. 09, 2018) - Some class structures are re-configured. * Version 1.0.1 (Nov. 07, 2018) - Some class structures are re-configured. * Version 1.0.0 (Oct. 08, 2018) - Initial version.