用Lasagne来实现MLP，测试mnist

虽然Lasagne自带的examples中带有MNIST的例子，但是这个例子的风格更接近用Theano来写。使用nolearn.lasagne中的NeuralNet类来写MLP似乎更符合Lasagne的设计初衷，也更像Caffe的风格。

code

没有比代码更清楚地了：

# Using lasagne to fit mnist.
# show how to use NeuralNet
# Tarrega, 150611.

import lasagne
from lasagne import layers
from lasagne.updates import nesterov_momentum
from nolearn.lasagne import NeuralNet
import os
from urllib import urlretrieve
import gzip
import pickle
import numpy

DATA_URL = 'http://deeplearning.net/data/mnist/mnist.pkl.gz'
DATA_FILENAME = 'mnist.pkl.gz'

def pickle_load(f, encoding):
    return pickle.load(f)

def _load_data(url=DATA_URL, filename=DATA_FILENAME):
    """Load data from `url` and store the result in `filename`."""
    print 'filename for the minist datatset:',filename
    if not os.path.exists(filename):
        print("Downloading MNIST dataset")
        urlretrieve(url, filename)

    with gzip.open(filename, 'rb') as f:
        return pickle_load(f, encoding='latin-1')


def load():
    """Get data with labels, split into training, validation and test set."""
    data = _load_data()
    X_train, y_train = data[0]
    y_train = numpy.asarray(y_train,dtype='int32')
    #X_valid, y_valid = data[1]
    #X_test, y_test = data[2]
    print 'size:',X_train.shape, y_train.shape, X_train.dtype, y_train.dtype
    return X_train,y_train


net1 = NeuralNet(
    layers=[  # three layers: one hidden layer
        ('input', layers.InputLayer),
        ('hidden', layers.DenseLayer),
        ('output', layers.DenseLayer),
        ],
    # layer parameters:
    input_shape=(None, 28*28),  # 28x28 input pixels per batch
    hidden_num_units=200,  # number of units in hidden layer
    output_nonlinearity=lasagne.nonlinearities.softmax,  # output layer
    output_num_units=10,  # 10 target values

    # optimization method:
    update=nesterov_momentum,
    update_learning_rate=0.01,
    update_momentum=0.9,

    regression=False,  # flag to indicate we're dealing with regression problem
    max_epochs=400,  # we want to train this many epochs
    verbose=1,
    )

X, y = load()
net1.fit(X, y)

output

运行结果：

  epoch    train loss    valid loss    train/val    valid acc  dur
-------  ------------  ------------  -----------  -----------  -----
      1       0.57989       0.32038      1.80998      0.91009  2.60s
      2       0.30194       0.26267      1.14949      0.92692  2.64s
      3       0.25044       0.22917      1.09281      0.93482  2.63s
      4       0.21578       0.20542      1.05046      0.94105  2.83s
      5       0.18969       0.18748      1.01177      0.94560  2.73s
      6       0.16917       0.17336      0.97585      0.94876  2.73s
      7       0.15259       0.16223      0.94059      0.95232  2.80s
      8       0.13882       0.15313      0.90655      0.95509  2.79s
      9       0.12723       0.14541      0.87498      0.95776  2.71s
     10       0.11725       0.13897      0.84373      0.96043  2.65s
     11       0.10859       0.13341      0.81392      0.96152  2.74s
     12       0.10100       0.12864      0.78511      0.96271  2.65s
     13       0.09424       0.12444      0.75735      0.96389  2.65s
     14       0.08822       0.12075      0.73056      0.96508  2.75s
     15       0.08280       0.11754      0.70446      0.96567  2.66s
     16       0.07790       0.11458      0.67988      0.96647  2.55s
     17       0.07343       0.11192      0.65606      0.96676  2.68s
     18       0.06936       0.10953      0.63324      0.96785  2.69s
     19       0.06561       0.10741      0.61082      0.96874  2.61s
     20       0.06218       0.10538      0.59000      0.96914  2.60s
     21       0.05899       0.10357      0.56959      0.96943  2.64s
     22       0.05604       0.10197      0.54953      0.96973  2.59s
     23       0.05329       0.10048      0.53033      0.97022  2.61s
     24       0.05070       0.09921      0.51105      0.97052  2.71s
     25       0.04829       0.09795      0.49296      0.97072  2.68s

Questions:

1. 为什么运行速度比Lasagne自带的Example快这么多？NeuralNet类自动把Training分为80%Train，20%Test，可能计算量小了。