theano函数在前馈神经网络的梯度优化过程中不更新参数

Question

通过从三层前馈神经网络的一个非常简单的实现开始并在 mnist 数据集上测试它，试图让我的手弄湿 theano 和深度网络。

我开始时使用的是随机梯度下降的基本实现，但网络训练不正确。网络参数未更新。

想知道是否有人可以指出我做错了什么。

下面的代码是我的lstm模块。我之所以这样称呼它，是因为我计划在未来实施 lstm 网络。

import theano, theano.tensor as T
import numpy as np
from collections import OrderedDict

np_rng = np.random.RandomState(1234)

class FeedForwardLayer(object):

  def __init__(self, input_size, hidden_size, activation):
    self.input_size = input_size
    self.hidden_size = hidden_size
    self.activation = activation
    self.create_layer()

  def create_layer(self):
    self.W = create_shared(self.hidden_size, self.input_size, "weight") 
    self.b = create_shared(self.hidden_size, name="bias")

  def activate(self, x):
    if x.ndim > 1:
      return self.activation(T.dot(self.W, x.T) + self.b[:, None]).T
    else:
      return self.activation(T.dot(self.W, x) + self.b)

  @property
  def params(self):
    return [self.W, self.b]

  @params.setter
  def params(self, param_list):
    self.W.set_value(param_list[0])
    self.b.set_value(param_list[1])

class Network(object):

  def __init__(self, input_size, celltype=FeedForwardLayer, layer_sizes=None):
    self.input_size = input_size
    self.celltype = celltype
    self.layer_sizes = layer_sizes
    self.create_layers()

  def create_layers(self):
    self.layers = []
    input_size = self.input_size

  for layer_size in self.layer_sizes:
    self.layers.append(self.celltype(input_size, layer_size, activation=T.nnet.sigmoid))
    input_size = layer_size

  def forward(self, x):
    out = []
    layer_input = x

    for layer in self.layers:
      layer_out = layer.activate(layer_input)
      out.append(layer_out)
      layer_input = layer_out

    return out

  @property
  def params(self):
    return [param for layer in self.layers for param in layer.params]

  @params.setter
  def params(self, param_list):
    start = 0
    for layer in self.layers:
      end = start + len(layer.params)
      layer.params = param_list[start:end]
      start = end

 def create_shared(m, n=None, name=None):
   if n is None:
     return theano.shared(np_rng.standard_normal((m, )), name=name)
   else:
     return theano.shared(np_rng.standard_normal((m, n)), name=name)

def optimization_updates(cost, params, lr=.01):
  """
  implements stochastic gradient descent 
  Inputs
  ---------------
  cost -- theano variable to minimize 
  params -- network weights to take gradient with respect to
  lr -- learning rate
  """
  lr = theano.shared(np.float64(lr).astype(theano.config.floatX))

  gparams = T.grad(cost, params)

  updates = OrderedDict()

  for gparam, param in zip(gparams, params):
    updates[param] = param - lr * gparam

  return updates

以下代码是我在 mnist 数据集上创建、训练和测试一个简单的三层前馈网络的地方。

from lstm import Network
import theano, theano.tensor as T
import numpy as np
import lstm as L
from sklearn.datasets import load_digits
from sklearn.cross_validation import train_test_split
from sklearn.metrics import confusion_matrix, classification_report
from sklearn.preprocessing import LabelBinarizer

# load and normalize dataset
digits = load_digits()
X = digits.data
y = digits.target
X -= X.min()
X /= X.max()

# create network
model = Network(64, layer_sizes=[100, 10])  

# prepare training and test data
X_train, X_test, y_train, y_test = train_test_split(X, y)
labels_train = LabelBinarizer().fit_transform(y_train)
labels_test = LabelBinarizer().fit_transform(y_test)

data = T.vector()
result = model.forward(data)[-1]
label = T.vector()

cost = (result - label).norm(L=2)
updates = L.optimization_updates(cost, model.params)

update = theano.function([data, label], cost, updates=updates, allow_input_downcast=True)
predict = theano.function([data], result, allow_input_downcast=True)

for X, y in zip(X_train, labels_train):
    c = update(X, y)

predictions = []

for X in X_test:
    prediction = predict(X)
    predictions.append(np.argmax(prediction))

print(confusion_matrix(y_test, predictions))
print(classification_report(y_test, predictions))

我面临的问题是参数没有正确更新。我不确定那是因为我没有正确计算梯度，还是因为我没有正确使用 theano 函数。

Answer 1

使用随机梯度下降时，您必须对数据集进行多次传递。分类错误和混淆矩阵在第一个时期变化不大的情况并不罕见，尤其是在数据集很小的情况下。

我对您的代码进行了以下更改以训练 100 个时期

for i in xrange(100):
    for X, y in zip(X_train, labels_train):
        c = update(X, y)

混淆矩阵似乎开始改善了：

[[ 0  0 18  0 13  4  5  0  5  0]
 [ 0 42  0  2  0  0  0  0  2  0]
 [ 0  0 51  0  0  0  0  1  0  0]
 [ 0  0  0 45  0  1  0  1  2  0]
 [ 0  0  0  0 33  0  0  0  0  0]
 [ 0  0  0  0  0 47  0  0  0  0]
 [ 0  0  0  0  0  0 45  0  0  0]
 [ 0  0  0  0  1  0  0 48  0  0]
 [ 0  2  1  0  0  0  0  0 34  0]
 [ 0  1  0 25  0  3  0  2 16  0]]

theano函数在前馈神经网络的梯度优化过程中不更新参数

theano function not updating parameters during gradient optimization in feed forward neural net

neural-network

python-3.x

theano