|
| 1 | +""" |
| 2 | +@author:wepon |
| 3 | +@blog:http://blog.csdn.net/u012162613/article/details/43157801 |
| 4 | +
|
| 5 | +""" |
| 6 | + |
| 7 | +# -*- coding: utf-8 -*- |
| 8 | +__docformat__ = 'restructedtext en' |
| 9 | + |
| 10 | +import cPickle |
| 11 | +import gzip |
| 12 | +import os |
| 13 | +import sys |
| 14 | +import time |
| 15 | + |
| 16 | +import numpy |
| 17 | + |
| 18 | +import theano |
| 19 | +import theano.tensor as T |
| 20 | + |
| 21 | +""" |
| 22 | +定义Softmax回归模型 |
| 23 | +在deeplearning tutorial中,直接将LogisticRegression视为Softmax, |
| 24 | +而我们所认识的二类别的逻辑回归就是当n_out=2时的LogisticRegression |
| 25 | +""" |
| 26 | +#参数说明: |
| 27 | +#input,输入的一个batch,假设一个batch有n个样本(n_example),则input大小就是(n_example,n_in) |
| 28 | +#n_in,每一个样本的大小,MNIST每个样本是一张28*28的图片,故n_in=784 |
| 29 | +#n_out,输出的类别数,MNIST有0~9共10个类别,n_out=10 |
| 30 | +class LogisticRegression(object): |
| 31 | + def __init__(self, input, n_in, n_out): |
| 32 | + |
| 33 | +#W大小是n_in行n_out列,b为n_out维向量。即:每个输出对应W的一列以及b的一个元素。WX+b |
| 34 | +#W和b都定义为theano.shared类型,这个是为了程序能在GPU上跑。 |
| 35 | + self.W = theano.shared( |
| 36 | + value=numpy.zeros( |
| 37 | + (n_in, n_out), |
| 38 | + dtype=theano.config.floatX |
| 39 | + ), |
| 40 | + name='W', |
| 41 | + borrow=True |
| 42 | + ) |
| 43 | + |
| 44 | + self.b = theano.shared( |
| 45 | + value=numpy.zeros( |
| 46 | + (n_out,), |
| 47 | + dtype=theano.config.floatX |
| 48 | + ), |
| 49 | + name='b', |
| 50 | + borrow=True |
| 51 | + ) |
| 52 | + |
| 53 | +#input是(n_example,n_in),W是(n_in,n_out),点乘得到(n_example,n_out),加上偏置b, |
| 54 | +#再作为T.nnet.softmax的输入,得到p_y_given_x |
| 55 | +#故p_y_given_x每一行代表每一个样本被估计为各类别的概率 |
| 56 | +#PS:b是n_out维向量,与(n_example,n_out)矩阵相加,内部其实是先复制n_example个b, |
| 57 | +#然后(n_example,n_out)矩阵的每一行都加b |
| 58 | + self.p_y_given_x = T.nnet.softmax(T.dot(input, self.W) + self.b) |
| 59 | + |
| 60 | +#argmax返回最大值下标,因为本例数据集是MNIST,下标刚好就是类别。axis=1表示按行操作。 |
| 61 | + self.y_pred = T.argmax(self.p_y_given_x, axis=1) |
| 62 | + |
| 63 | +#params,模型的参数 |
| 64 | + self.params = [self.W, self.b] |
| 65 | + |
| 66 | +#代价函数NLL |
| 67 | +#因为我们是MSGD,每次训练一个batch,一个batch有n_example个样本,则y大小是(n_example,), |
| 68 | +#y.shape[0]得出行数即样本数,将T.log(self.p_y_given_x)简记为LP, |
| 69 | +#则LP[T.arange(y.shape[0]),y]得到[LP[0,y[0]], LP[1,y[1]], LP[2,y[2]], ...,LP[n-1,y[n-1]]] |
| 70 | +#最后求均值mean,也就是说,minibatch的SGD,是计算出batch里所有样本的NLL的平均值,作为它的cost |
| 71 | + def negative_log_likelihood(self, y): |
| 72 | + return -T.mean(T.log(self.p_y_given_x)[T.arange(y.shape[0]), y]) |
| 73 | + |
| 74 | +#batch的误差率 |
| 75 | + def errors(self, y): |
| 76 | + # 首先检查y与y_pred的维度是否一样,即是否含有相等的样本数 |
| 77 | + if y.ndim != self.y_pred.ndim: |
| 78 | + raise TypeError( |
| 79 | + 'y should have the same shape as self.y_pred', |
| 80 | + ('y', y.type, 'y_pred', self.y_pred.type) |
| 81 | + ) |
| 82 | + # 再检查是不是int类型,是的话计算T.neq(self.y_pred, y)的均值,作为误差率 |
| 83 | + #举个例子,假如self.y_pred=[3,2,3,2,3,2],而实际上y=[3,4,3,4,3,4] |
| 84 | + #则T.neq(self.y_pred, y)=[0,1,0,1,0,1],1表示不等,0表示相等 |
| 85 | + #故T.mean(T.neq(self.y_pred, y))=T.mean([0,1,0,1,0,1])=0.5,即错误率50% |
| 86 | + if y.dtype.startswith('int'): |
| 87 | + return T.mean(T.neq(self.y_pred, y)) |
| 88 | + else: |
| 89 | + raise NotImplementedError() |
| 90 | + |
| 91 | +""" |
| 92 | +加载MNIST数据集 |
| 93 | +""" |
| 94 | +def load_data(dataset): |
| 95 | + # dataset是数据集的路径,程序首先检测该路径下有没有MNIST数据集,没有的话就下载MNIST数据集 |
| 96 | + #这一部分就不解释了,与softmax回归算法无关。 |
| 97 | + data_dir, data_file = os.path.split(dataset) |
| 98 | + if data_dir == "" and not os.path.isfile(dataset): |
| 99 | + # Check if dataset is in the data directory. |
| 100 | + new_path = os.path.join( |
| 101 | + os.path.split(__file__)[0], |
| 102 | + "..", |
| 103 | + "data", |
| 104 | + dataset |
| 105 | + ) |
| 106 | + if os.path.isfile(new_path) or data_file == 'mnist.pkl.gz': |
| 107 | + dataset = new_path |
| 108 | + |
| 109 | + if (not os.path.isfile(dataset)) and data_file == 'mnist.pkl.gz': |
| 110 | + import urllib |
| 111 | + origin = ( |
| 112 | + 'http://www.iro.umontreal.ca/~lisa/deep/data/mnist/mnist.pkl.gz' |
| 113 | + ) |
| 114 | + print 'Downloading data from %s' % origin |
| 115 | + urllib.urlretrieve(origin, dataset) |
| 116 | + |
| 117 | + print '... loading data' |
| 118 | +#以上是检测并下载数据集mnist.pkl.gz,不是本文重点。下面才是load_data的开始 |
| 119 | + |
| 120 | +#从"mnist.pkl.gz"里加载train_set, valid_set, test_set,它们都是包括label的 |
| 121 | +#主要用到python里的gzip.open()函数,以及 cPickle.load()。 |
| 122 | +#‘rb’表示以二进制可读的方式打开文件 |
| 123 | + f = gzip.open(dataset, 'rb') |
| 124 | + train_set, valid_set, test_set = cPickle.load(f) |
| 125 | + f.close() |
| 126 | + |
| 127 | + |
| 128 | +#将数据设置成shared variables,主要时为了GPU加速,只有shared variables才能存到GPU memory中 |
| 129 | +#GPU里数据类型只能是float。而data_y是类别,所以最后又转换为int返回 |
| 130 | + def shared_dataset(data_xy, borrow=True): |
| 131 | + data_x, data_y = data_xy |
| 132 | + shared_x = theano.shared(numpy.asarray(data_x, |
| 133 | + dtype=theano.config.floatX), |
| 134 | + borrow=borrow) |
| 135 | + shared_y = theano.shared(numpy.asarray(data_y, |
| 136 | + dtype=theano.config.floatX), |
| 137 | + borrow=borrow) |
| 138 | + return shared_x, T.cast(shared_y, 'int32') |
| 139 | + |
| 140 | + |
| 141 | + test_set_x, test_set_y = shared_dataset(test_set) |
| 142 | + valid_set_x, valid_set_y = shared_dataset(valid_set) |
| 143 | + train_set_x, train_set_y = shared_dataset(train_set) |
| 144 | + |
| 145 | + rval = [(train_set_x, train_set_y), (valid_set_x, valid_set_y), |
| 146 | + (test_set_x, test_set_y)] |
| 147 | + return rval |
| 148 | + |
| 149 | +""" |
| 150 | +将该模型应用于MNIST |
| 151 | +""" |
| 152 | +def sgd_optimization_mnist(learning_rate=0.13, n_epochs=1000, |
| 153 | + dataset='mnist.pkl.gz', |
| 154 | + batch_size=600): |
| 155 | +#加载数据 |
| 156 | + datasets = load_data(dataset) |
| 157 | + train_set_x, train_set_y = datasets[0] |
| 158 | + valid_set_x, valid_set_y = datasets[1] |
| 159 | + test_set_x, test_set_y = datasets[2] |
| 160 | +#计算有多少个minibatch,因为我们的优化算法是MSGD,是一个batch一个batch来计算cost的 |
| 161 | + n_train_batches = train_set_x.get_value(borrow=True).shape[0] / batch_size |
| 162 | + n_valid_batches = valid_set_x.get_value(borrow=True).shape[0] / batch_size |
| 163 | + n_test_batches = test_set_x.get_value(borrow=True).shape[0] / batch_size |
| 164 | + |
| 165 | + ###################### |
| 166 | + # 开始建模 # |
| 167 | + ###################### |
| 168 | + print '... building the model' |
| 169 | + |
| 170 | + |
| 171 | +#设置变量,index表示minibatch的下标,x表示训练样本,y是对应的label |
| 172 | + index = T.lscalar() |
| 173 | + x = T.matrix('x') |
| 174 | + y = T.ivector('y') |
| 175 | + |
| 176 | + |
| 177 | +#定义分类器,用x作为input初始化。 |
| 178 | + classifier = LogisticRegression(input=x, n_in=28 * 28, n_out=10) |
| 179 | + |
| 180 | + |
| 181 | +#定义代价函数,用y来初始化,而其实还有一个隐含的参数x在classifier中。 |
| 182 | +#这样理解才是合理的,因为cost必须由x和y得来,单单y是得不到cost的。 |
| 183 | + cost = classifier.negative_log_likelihood(y) |
| 184 | + |
| 185 | + |
| 186 | +#这里必须说明一下theano的function函数,givens是字典,其中的x、y是key,冒号后面是它们的value。 |
| 187 | +#在function被调用时,x、y将被具体地替换为它们的value,而value里的参数index就是inputs=[index]这里给出。 |
| 188 | +#下面举个例子: |
| 189 | +#比如test_model(1),首先根据index=1具体化x为test_set_x[1 * batch_size: (1 + 1) * batch_size], |
| 190 | +#具体化y为test_set_y[1 * batch_size: (1 + 1) * batch_size]。然后函数计算outputs=classifier.errors(y), |
| 191 | +#这里面有参数y和隐含的x,所以就将givens里面具体化的x、y传递进去。 |
| 192 | + test_model = theano.function( |
| 193 | + inputs=[index], |
| 194 | + outputs=classifier.errors(y), |
| 195 | + givens={ |
| 196 | + x: test_set_x[index * batch_size: (index + 1) * batch_size], |
| 197 | + y: test_set_y[index * batch_size: (index + 1) * batch_size] |
| 198 | + } |
| 199 | + ) |
| 200 | + |
| 201 | + |
| 202 | + validate_model = theano.function( |
| 203 | + inputs=[index], |
| 204 | + outputs=classifier.errors(y), |
| 205 | + givens={ |
| 206 | + x: valid_set_x[index * batch_size: (index + 1) * batch_size], |
| 207 | + y: valid_set_y[index * batch_size: (index + 1) * batch_size] |
| 208 | + } |
| 209 | + |
| 210 | +# 计算各个参数的梯度 |
| 211 | + g_W = T.grad(cost=cost, wrt=classifier.W) |
| 212 | + g_b = T.grad(cost=cost, wrt=classifier.b) |
| 213 | + |
| 214 | +#更新的规则,根据梯度下降法的更新公式 |
| 215 | + updates = [(classifier.W, classifier.W - learning_rate * g_W), |
| 216 | + (classifier.b, classifier.b - learning_rate * g_b)] |
| 217 | + |
| 218 | +#train_model跟上面分析的test_model类似,只是这里面多了updatas,更新规则用上面定义的updates 列表。 |
| 219 | + train_model = theano.function( |
| 220 | + inputs=[index], |
| 221 | + outputs=cost, |
| 222 | + updates=updates, |
| 223 | + givens={ |
| 224 | + x: train_set_x[index * batch_size: (index + 1) * batch_size], |
| 225 | + y: train_set_y[index * batch_size: (index + 1) * batch_size] |
| 226 | + } |
| 227 | + ) |
| 228 | + |
| 229 | + ############### |
| 230 | + # 开始训练 # |
| 231 | + ############### |
| 232 | + print '... training the model' |
| 233 | + |
| 234 | + patience = 5000 |
| 235 | + patience_increase = 2 |
| 236 | +#提高的阈值,在验证误差减小到之前的0.995倍时,会更新best_validation_loss |
| 237 | + improvement_threshold = 0.995 |
| 238 | +#这样设置validation_frequency可以保证每一次epoch都会在验证集上测试。 |
| 239 | + validation_frequency = min(n_train_batches, patience / 2) |
| 240 | + |
| 241 | + |
| 242 | + best_validation_loss = numpy.inf #最好的验证集上的loss,最好即最小。初始化为无穷大 |
| 243 | + test_score = 0. |
| 244 | + start_time = time.clock() |
| 245 | + |
| 246 | + done_looping = False |
| 247 | + epoch = 0 |
| 248 | + |
| 249 | +#下面就是训练过程了,while循环控制的时步数epoch,一个epoch会遍历所有的batch,即所有的图片。 |
| 250 | +#for循环是遍历一个个batch,一次一个batch地训练。for循环体里会用train_model(minibatch_index)去训练模型, |
| 251 | +#train_model里面的updatas会更新各个参数。 |
| 252 | +#for循环里面会累加训练过的batch数iter,当iter是validation_frequency倍数时则会在验证集上测试, |
| 253 | +#如果验证集的损失this_validation_loss小于之前最佳的损失best_validation_loss, |
| 254 | +#则更新best_validation_loss和best_iter,同时在testset上测试。 |
| 255 | +#如果验证集的损失this_validation_loss小于best_validation_loss*improvement_threshold时则更新patience。 |
| 256 | +#当达到最大步数n_epoch时,或者patience<iter时,结束训练 |
| 257 | + while (epoch < n_epochs) and (not done_looping): |
| 258 | + epoch = epoch + 1 |
| 259 | + for minibatch_index in xrange(n_train_batches): |
| 260 | + |
| 261 | + minibatch_avg_cost = train_model(minibatch_index) |
| 262 | + # iteration number |
| 263 | + iter = (epoch - 1) * n_train_batches + minibatch_index |
| 264 | + |
| 265 | + if (iter + 1) % validation_frequency == 0: |
| 266 | + # compute zero-one loss on validation set |
| 267 | + validation_losses = [validate_model(i) |
| 268 | + for i in xrange(n_valid_batches)] |
| 269 | + this_validation_loss = numpy.mean(validation_losses) |
| 270 | + |
| 271 | + print( |
| 272 | + 'epoch %i, minibatch %i/%i, validation error %f %%' % |
| 273 | + ( |
| 274 | + epoch, |
| 275 | + minibatch_index + 1, |
| 276 | + n_train_batches, |
| 277 | + this_validation_loss * 100. |
| 278 | + ) |
| 279 | + ) |
| 280 | + |
| 281 | + # if we got the best validation score until now |
| 282 | + if this_validation_loss < best_validation_loss: |
| 283 | + #improve patience if loss improvement is good enough |
| 284 | + if this_validation_loss < best_validation_loss * \ |
| 285 | + improvement_threshold: |
| 286 | + patience = max(patience, iter * patience_increase) |
| 287 | + |
| 288 | + best_validation_loss = this_validation_loss |
| 289 | + # test it on the test set |
| 290 | + |
| 291 | + test_losses = [test_model(i) |
| 292 | + for i in xrange(n_test_batches)] |
| 293 | + test_score = numpy.mean(test_losses) |
| 294 | + |
| 295 | + print( |
| 296 | + ( |
| 297 | + ' epoch %i, minibatch %i/%i, test error of' |
| 298 | + ' best model %f %%' |
| 299 | + ) % |
| 300 | + ( |
| 301 | + epoch, |
| 302 | + minibatch_index + 1, |
| 303 | + n_train_batches, |
| 304 | + test_score * 100. |
| 305 | + ) |
| 306 | + ) |
| 307 | + |
| 308 | + if patience <= iter: |
| 309 | + done_looping = True |
| 310 | + break |
| 311 | + |
| 312 | +#while循环结束 |
| 313 | + end_time = time.clock() |
| 314 | + print( |
| 315 | + ( |
| 316 | + 'Optimization complete with best validation score of %f %%,' |
| 317 | + 'with test performance %f %%' |
| 318 | + ) |
| 319 | + % (best_validation_loss * 100., test_score * 100.) |
| 320 | + ) |
| 321 | + print 'The code run for %d epochs, with %f epochs/sec' % ( |
| 322 | + epoch, 1. * epoch / (end_time - start_time)) |
| 323 | + print >> sys.stderr, ('The code for file ' + |
| 324 | + os.path.split(__file__)[1] + |
| 325 | + ' ran for %.1fs' % ((end_time - start_time))) |
| 326 | + |
| 327 | +if __name__ == '__main__': |
| 328 | + sgd_optimization_mnist() |
0 commit comments