按照这theano的tutorial开始跟着写了,因为去年年底之前学习过一段时间,但是当时时间少,并且很多地方也没搞懂,大多都是看着书来模仿,结果出错不好找地方之外,自己如果根据自己的想法随便写下结果就出错了。这几天好好的学习了下。我来总结下
在softmax或者逻辑回归的代码中:
1:在教程中写法是,先写一个类,在类的init方法中初始化w和b,以及计算概率。
2:然后分别再另外的函数中计算cost和error。
3:在类外来进行训练。
以上方法由于在init方法中,要进行概率计算,所以在初始化类的时候至少要传递进去x,在计算cost和error的时候要传递进去y。所以如果此时训练方法写在类内的话就不行。因为数据也是在训练方法内生成的,在类别没办法传递进去(此时,你也许会说在类外面单独加载数据怎么样?因为训练方法内其实传递进去的是数据的索引,所以这个在外面单独加载数据的话也不太现实。所以加载数据还是和训练方法写在一起是最方便的)
为了解决以上问题,很容易,就是把init方法中,把在类外面不能传递的参数,全部去掉,然后在别的计算的地方用到的话,再传递进去。比如把概率预测拿出来,写成一个函数,把x传递进去。然后其余的保持不变。这样在训练方法内,因为有对x和y的声明,以及对应数据的传递进去,所以不会出问题
我这里写好了一个,发上来作为备忘吧,为下一步的更高层次的封装做准备:
- import numpy, theano, theano.tensor as T, gzip, cPickle
-
- class NN():
-
- def __init__(self, n_in, n_out):
- self.w = theano.shared(numpy.asarray(numpy.zeros([n_in, n_out]), theano.config.floatX))
- self.b = theano.shared(numpy.asarray(numpy.zeros(n_out), theano.config.floatX))
- def get_probalblity(self, x):
- return T.nnet.softmax(T.dot(x, self.w) + self.b)
- def get_prediction(self, x, y):
- return T.argmax(self.get_probalblity(x), 1)
- def cost(self, x, y):
- p_y_given_x = self.get_probalblity(x)
- return -T.mean(T.log(p_y_given_x[T.arange(y.shape[0]), y]))
- def error(self, x, y):
- prediction = self.get_prediction(x, y)
- return T.mean(T.neq(prediction, y))
- def load_data(self):
- f = gzip.open('mnist.pkl.gz')
- trainxy, validatexy, testxy = cPickle.load(f)
- def share_data(xy):
- x,y = xy
- x = theano.shared(numpy.asarray(x, theano.config.floatX))
- y = theano.shared(numpy.asarray(y, theano.config.floatX))
- return [x, T.cast(y, 'int32')]
- trainx, trainy = share_data(trainxy)
- validatex,validatey = share_data(validatexy)
- testx, testy = share_data(testxy)
- return [(trainx,trainy),(validatex,validatey),(testx,testy)]
- def train(self):
- x = T.matrix('x', theano.config.floatX)
- y = T.ivector('y')
- [(trainx,trainy),(validatex,validatey),(testx,testy)] = self.load_data()
- gw,gb = T.grad(self.cost(x,y), [self.w, self.b])
- index = T.lscalar()
-
- batch_size = 600
- trainModel = theano.function([index], self.cost(x,y), updates=[(self.w, self.w-0.13*gw), (self.b, self.b-0.13*gb)], givens={x:trainx[index*batch_size:(index+1)*batch_size], y:trainy[index*batch_size:(index+1)*batch_size]})
- validateModel = theano.function([index], self.error(x,y), givens={x:validatex[index*batch_size:(index+1)*batch_size], y:validatey[index*batch_size:(index+1)*batch_size]})
- testModel = theano.function([index], self.error(x,y), givens={x:testx[index*batch_size:(index+1)*batch_size], y:testy[index*batch_size:(index+1)*batch_size]})
-
- best_validate_error = numpy.Inf
- best_test_error = 0
- patience = 5000
- increasement = 2
- train_batchs = trainx.get_value().shape[0]/batch_size
- validate_batchs = validatex.get_value().shape[0]/batch_size
- test_batchs = testx.get_value().shape[0]/batch_size
- validate_frequency = min(patience/2, train_batchs)
- epochs = 1000
- epoch = 1
- ite = 0
- stopping = False
- while (epoch < epochs) and (not stopping):
- for i in xrange(train_batchs):
- ite += 1
- this_cost = trainModel(i)
- if ite%validate_frequency == 0:
- this_validate_error = numpy.mean([validateModel(j) for j in xrange(validate_batchs)])
- print ('ite:%d/%d, cost:%f, validate:%f'%(ite, epoch, this_cost, this_validate_error))
- if this_validate_error < best_validate_error:
- if this_validate_error < 0.995*best_validate_error:
- patience = max(patience, ite*increasement)
- this_test_error = numpy.mean([testModel(j) for j in xrange(test_batchs)])
- best_validate_error = this_validate_error
- best_test_error = this_test_error
- print ('ite:%d/%d, cost:%f, validate:%f, test:%f'%(ite, epoch, this_cost, this_validate_error, this_test_error))
- if patience <= ite:
- stopping = True
- break
- epoch +=1
- print ('best validate error:%f, best test error:%f'%(best_validate_error, best_test_error))
-
-
- if __name__ == '__main__':
- nn = NN(784, 10)
- nn.train()
以上训练结果,和tutorial给出的基本一致。
。。。。。
ite:5810/70, cost:0.329380, validate:0.075104
ite:5810/70, cost:0.329380, validate:0.075104, test:0.075000
ite:5893/71, cost:0.329054, validate:0.075208
ite:5976/72, cost:0.328735, validate:0.075104
ite:5976/72, cost:0.328735, validate:0.075104, test:0.075104
ite:6059/73, cost:0.328422, validate:0.075000
ite:6059/73, cost:0.328422, validate:0.075000, test:0.074896
ite:6142/74, cost:0.328116, validate:0.074792
ite:6142/74, cost:0.328116, validate:0.074792, test:0.074896
best validate error:0.074792, best test error:0.074896