Mnist handwriting recognition case

MNIST is a large handwritten digit recognition database, which is widely used for training and testing in the field of machine learning. Including 60,000 training sets and 10,000 test sets, each image has been normalized in size and data is centered. The fixed size is 28×28 pixels. 

data set

Use the get_mnist.sh script to download under data/mnist in the Caffe source code directory.

$ cd data/mnist/
$ ./get_mnist.sh

mnist data format description:

There are four more files in the data/mnist directory

file name	Description
train-images-idx3-ubyte	Training set, picture
train-labels-idx1-ubyte	Training set, label
t10k-images-idx3-ubyte	Test set, picture
t10k-labels-idx1-ubyte	Test set, label

The format of test images and training images is the same;

Conversion format

The pixels in the picture file are organized in rows, white (pixel value 0) is the background, and black (pixel value 255) is the foreground;

The downloaded data set is a binary file and needs to be converted to LMDB or LEVELDB to be recognized by Caffe.

$ ./examples/mnist/create_mnist.sh
Creating lmdb...
Done.

Two directories, mnist_train_lmdb/ and mnist_test_lmdb/, are generated in the examples/mnist directory, and there are two files in each directory: data.mdb and lock.mdb. Mnist_train_lmdb is a training set in LMDB format, and mnist_test_lmdb is a test set in LMDB format.

tips：

Caffe uses LMDB and LEVELDB for reason 1: to provide a unified data type format; reason 2: to improve disk IO utilization;

Network model

Review: Lenet's network structure (proto configuration)examples/mnist/lenet_solver.prototxt

Training network

tips: Be sure to read the printed log;

GLOG output format: date, time, process number, source file: code line number] output information;

It is convenient to track the remote operation process and analyze the operation efficiency;

CPU mode: the last line of examples/mnist/lenet_solver.prototxt is modified as follows

# solver mode: CPU or GPU
solver_mode: CPU

Run the examples/mnist/train_lenet.sh script

$ ./examples/mnist/train_lenet.sh 

The final result of training is saved in examples/mnist/lenet_iter_10000.caffemodel.

test

Use the MNIST test set to test the trained model

$ ./build/tools/caffe.bin test \
-model examples/mnist/lenet_train_test.prototxt \
-weights examples/mnist/lenet_iter_10000.caffemodel \
-iterations 100

Command line parameters:

• ./build/tools/caffe.bin test means that only prediction (forward propagation) is made, and no parameter update is made;
• -model examples/mnist/lenet_train_test.prototxt specifies the model description file
• -weights examples/mnist/lenet_iter_10000.caffemodel specifies the trained weight file
• iterations 100 Specify the number of test iterations, the number of participating tests=iterations*batch_size, the data volume of each iteration is set batch_size: 100 in the model description file, and 100 iterations just cover 10000 samples of the test set.

supplement:

./build/tools/caffe.bin