MNIST/Pylearn2 Simple 2-layer (Sigmod/Softmax) MLP trainer

gistfile1.txt

"""
This trains a simple 2-layered MLP for classifying handwritten digits (MNIST dataset)
For this script to work, you need to have a ./mnist folder that contains the MNIST dataset.

The dataset can be downloaded using the script in the following pylearn2 folder:
<workspace>/pylearn2/pylearn2/scripts/datasets/download_mnist.py

After training the model (neural network containing layers), it will save trained model to disk.
"""

import struct
from digits.main_old import save_model
import numpy as np
from pylearn2.datasets.dense_design_matrix import DenseDesignMatrix
from pylearn2.models import mlp
from pylearn2.termination_criteria import EpochCounter
from pylearn2.training_algorithms import sgd
from pylearn2.utils import serial


TRAINING_IMAGES_PATH = "mnist/train-images-idx3-ubyte"
TRAINING_LABEL_PATH = "mnist/train-labels-idx1-ubyte"
MODEL_SAVE_PATH = "./model.saved"


def read_mnist_images():
    dtype = 'float32'
    with open(TRAINING_IMAGES_PATH, 'rb') as f:
        _, number, rows, cols = struct.unpack('>iiii', f.read(
            16))  # gotta do this line first to move the file pointer to AFTER the first 16 bytes
        array = np.fromfile(f, dtype='uint8').reshape((number, rows, cols))
    dtype = np.dtype(dtype)
    array = array.astype(dtype)
    return array


def get_training_inputs():
    topo_view = read_mnist_images()
    m, r, c = topo_view.shape
    topo_view = topo_view.reshape(m, r, c, 1)

    # rearrange axis to pylearn2's default (in this case, not needed, but added for documentation)
    axes = ['b', 0, 1, 'c']
    default = ('b', 0, 1, 'c')
    topo_view = topo_view.transpose(*[default.index(axis) for axis in axes])

    return topo_view


def get_training_labels():
    with open(TRAINING_LABEL_PATH, 'rb') as f:
        _, number = struct.unpack('>ii',
                                  f.read(
                                      8))  # gotta do this line first to move the file pointer to AFTER the first 8
        # bytes
        label_array = np.fromfile(f, dtype='uint8')
        label_array = np.atleast_2d(label_array).transpose()
    return label_array


def save_model(model):
    model_path = "./model.saved"
    serial.save(model_path, model)


if __name__ == "__main__":
    # build dataset
    training_dataset = DenseDesignMatrix(topo_view=get_training_inputs(), y=get_training_labels(),
                                         axes=['b', 0, 1, 'c'],
                                         y_labels=10)

    # build layer
    hidden_layer = mlp.Sigmoid(layer_name="hidden", dim=500, sparse_init=15) #sigmo
    output_layer = mlp.Softmax(10, "output", irange=.1) #softmax layer, 10 outputs
    layers = [hidden_layer, output_layer]

    # init trainer
    trainer = sgd.SGD(learning_rate=.05, batch_size=10, termination_criterion=EpochCounter(400))

    # init neural network
    neural_network_model = mlp.MLP(layers, nvis=784)  # 784 inputs
    trainer.setup(neural_network_model, training_dataset)

    # begin training the network with dataset
    print "Begin training network.."
    while True:
        trainer.train(dataset=training_dataset)
        neural_network_model.monitor.report_epoch()
        neural_network_model.monitor()
        if not trainer.continue_learning(neural_network_model):
            print "Training ends."
            break

    print "Saving trained network.."
    save_model(neural_network_model)
    print "Network saved"