文件列表:

ver 0.72\@cnn\adapt_dw.m (3853, 2009-04-24)
ver 0.72\@cnn\calchx.m (5111, 2009-05-29)
ver 0.72\@cnn\calcje.m (5897, 2009-05-06)
ver 0.72\@cnn\check_finit_dif.m (2250, 2009-05-06)
ver 0.72\@cnn\cnn.m (9176, 2009-05-29)
ver 0.72\@cnn\cnn_size.m (973, 2009-03-28)
ver 0.72\@cnn\init.m (3873, 2009-05-29)
ver 0.72\@cnn\sim.m (2967, 2009-05-29)
ver 0.72\@cnn\subsasgn.m (8760, 2009-05-06)
ver 0.72\@cnn\subsref.m (1708, 2009-03-28)
ver 0.72\@cnn\train.m (2647, 2009-05-29)
ver 0.72\@cnn (0, 2009-09-20)
ver 0.72\back_conv2.m (716, 2009-05-29)
ver 0.72\back_subsample.m (1052, 2009-05-27)
ver 0.72\cnet.mat (524533, 2009-05-29)
ver 0.72\cnet_tool.m (18527, 2009-05-31)
ver 0.72\cudacnn.mexw32 (122880, 2010-08-20)
ver 0.72\eraser.gif (925, 2009-05-29)
ver 0.72\fastFilter2.m (128, 2009-05-27)
ver 0.72\license.txt (1337, 2009-06-05)
ver 0.72\preproc_data.m (1018, 2009-03-28)
ver 0.72\preproc_image.m (413, 2010-08-20)
ver 0.72\rand_std.m (972, 2009-05-27)
ver 0.72\readMNIST.m (2494, 2010-08-20)
ver 0.72\readMNIST_image.m (1016, 2010-08-20)
ver 0.72\rot180.m (57, 2009-03-28)
ver 0.72\subsample.m (1187, 2009-03-19)
ver 0.72\tansig_mod.m (3119, 2009-05-27)
ver 0.72\test_dgt.m (444, 2009-05-29)
ver 0.72\train_cnn.m (4955, 2009-05-29)
ver 0.72 (0, 2010-08-21)

CNN Class, ver 0.72. Change log: Ver 0.72: Sample GUI added, demonstrating use of convolutional network for handwriten digits recognition. Training runs 20% faster. Ver 0.71: Bug fix: training was stoped after 1 epoch. Ver 0.70: First release. This project provides matlab class for implementation of convolutional neural networks. This networks was created by Yann LeCun and have sucessfully used in many practical applications, such as handwriten digits recognition, face detection, robot navigation and others (see references for more info). Because of some architectural features of convolutional networks, such as weight sharing it is imposible to implement it using Matlab Neural Network Toolbox without it's source modifications. That's why this class wokrs almost independently from NN toolbox (coming soon full independence). This release includes sample of handwriten digits recognition using CNN. If you just want to try it run cnet_tool. You'll see a simple GUI. It loads pretrained convolutional neural net from cnet.mat and recognizes image of gigit either pained in painitg area or downloaded from MNIST database. I should note that currently recognition have about 4% error, wich is higher than in Yann LeCun's classifier comparison table (http://yann.lecun.com/exdb/mnist/index.html). This is because my implementation have another activation and error functions (tansig and MSE). Soon will be release with radbas and MLE functions, as stated in [2]. If you interested in training you should open train_cnn.m, set all parameters following to comments and start learning by runing it. The action sequence for creation of arbitrary convolutional neural network is following: 1. Create cnn object. 2. Set archtecture (number of layers, weights, training parameters, etc). 3. Call init method. 4. Define connection matrices for layers if necessary. 5. Load training data. 6. Preprocess training data. 7. Start training. 8. Test neural net. There're several conditions in current version: 1. Network can have only one input for image (e.g. no stereo images simulateously). 2. You have to set connection matrix after the initialization. 3. MNIST database of handwriten digits not included in this distribution, you can download it from http://yann.lecun.com/exdb/mnist/index.html The code is far from optimality both in speed and size, the main reason of it is clarity in the process of initial development. So the next version promise to be much faster. Now the training a sample network consisting of 8 layers with about 100 000 connections on 5000 training samples takes 20 minutes approximately on AMD Opteron 2.6 GHz. So to train the network on all MNIST database it is better to keep it work a night. Next version expectations: 1. Ability to work with several images simulateously. 2. Training speed improvement. 3. Image Aquision toolbox integration. 4. Face detection sample. References: 1. Y. LeCun, L. Bottou, G. Orr and K. Muller: Efficient BackProp, in Orr, G. and Muller K. (Eds), Neural Networks: Tricks of the trade, Springer, 19*** URL:http://yann.lecun.com/exdb/publis/index.html 2. Y. LeCun, L. Bottou, Y. Bengio and P. Haffner: Gradient-Based Learning Applied to Document Recognition, Proceedings of the IEEE, 86(11):2278-2324, November 19*** URL:http://yann.lecun.com/exdb/publis/index.html 3. Patrice Y. Simard, Dave Steinkraus, John C. Platt: Best Practices for Convolutional Neural Networks Applied to Visual Document Analysis URL:http://research.microsoft.com/apps/pubs/?id=68920 4. Thanks to Mike O'Neill for his great article, wich is summarize and generalize all the information in 1-3 for better understandig for programming: URL: http://www.codeproject.com/KB/library/NeuralNetRecognition.aspx 5. Also thanks to Jake Bouvrie for his "Notes on Convolutional Neural Networks", particulary for the idea to debug the neural network using finite differences URL: http://web.mit.edu/jvb/www/cv.html ========== (c) Mikhail Sitotenko. This software distributing under BSD licence.

近期下载者：

相关文件：

评论：[我要评论] [举报此文件]

收藏者：