Question

仅对一批进行训练时，神经网络就很好，但是对整个数据集进行训练时，则无法学习。

当我发现培训准确性和测试准确性都没有提高时，问题就开始了。我尝试只使用一批（四个样本）来训练网络，以确保代码本身能够正常工作，并且可以正常工作并提高准确性。

另一方面，当尝试使用整个数据集训练网络时，网络不会训练。事实证明，在处理整个数据集时，权重没有得到更新（最终输出层上的偏差项除外），而仅处理一批时，所有权重都得到了正确的更新。

关于可能发生的事情的任何想法吗？

网络架构： 3个转换，1个最大池，1个转换，1个最大池，2个FC和1个输出层。

这是CNN的代码：

anat_x = tf.placeholder("float", [None, anat_img_depth, anat_img_dim1, anat_img_dim2, 1])
y = tf.placeholder("float", [None, n_classes])
kp = tf.placeholder("float", shape=())

def conv3d(x, W, b, s=2):
    # Conv2D wrapper, with bias and relu activation
    x = tf.nn.conv3d(x, W, strides=[1, s, s, s, 1], padding='SAME')
    x = tf.nn.bias_add(x, b)
    return tf.nn.relu(x)

def maxpool3d(x, k=2, s=2):
    return tf.nn.max_pool3d(x, ksize=[1, k, k, k, 1], strides=[1, s, s, s, 1], padding='SAME')

    weights_anat = {
    'wc1': tf.get_variable('W0_A', shape=(3,3,3,1,32), initializer=tf.contrib.layers.xavier_initializer()),
    'wc2': tf.get_variable('W1_A', shape=(3,3,3,32,32), initializer=tf.contrib.layers.xavier_initializer()),
    'wc3': tf.get_variable('W2_A', shape=(3,3,3,32,64), initializer=tf.contrib.layers.xavier_initializer()),
    'wc4': tf.get_variable('W3_A', shape=(3,3,3,64,64), initializer=tf.contrib.layers.xavier_initializer()),
    'wd1': tf.get_variable('W4_A', shape=(3*4*4*64,1024), initializer=tf.contrib.layers.xavier_initializer()),
    'wd2': tf.get_variable('W5_A', shape=(1024,512), initializer=tf.contrib.layers.xavier_initializer()),
    'out': tf.get_variable('W6_A', shape=(512,2), initializer=tf.contrib.layers.xavier_initializer()),
}
biases_anat = {
    'bc1': tf.get_variable('B0_A', shape=(32), initializer=tf.contrib.layers.xavier_initializer()),
    'bc2': tf.get_variable('B1_A', shape=(32), initializer=tf.contrib.layers.xavier_initializer()),
    'bc3': tf.get_variable('B2_A', shape=(64), initializer=tf.contrib.layers.xavier_initializer()),
    'bc4': tf.get_variable('B3_A', shape=(64), initializer=tf.contrib.layers.xavier_initializer()),
    'bd1': tf.get_variable('B4_A', shape=(1024), initializer=tf.contrib.layers.xavier_initializer()),
    'bd2': tf.get_variable('B5_A', shape=(512), initializer=tf.contrib.layers.xavier_initializer()),
    'out': tf.get_variable('B6_A', shape=(2), initializer=tf.contrib.layers.xavier_initializer()),
}

def conv_net(x, weights, biases):

    print("input shape: " + str(x.shape))  

    # Here we call the conv3d function we had defined above and pass the input image x, weights wc1 and bias bc1.
    conv1_anat = conv3d(x, weights['wc1'], biases['bc1'], s=2)
    print("conv1_anat shape: " + str(conv1_anat.shape))  

    # Here we call the conv3d function we had defined above and pass the previous conv1 layer, weights wc2 and bias bc2.    
    conv2_anat = conv3d(conv1_anat, weights['wc2'], biases['bc2'], s=2)
    print("conv2_anat shape: " + str(conv2_anat.shape))  

    # Here we call the conv3d function we had defined above and pass the previous conv2 layer, weights wc3 and bias bc3.    
    conv3_anat = conv3d(conv2_anat, weights['wc3'], biases['bc3'], s=2)
    print("conv3_anat shape: " + str(conv3_anat.shape))  

    # Max Pooling (down-sampling), this chooses the max value from a 2*2 matrix window and outputs a 14*14 matrix.
    pooling1_anat = maxpool3d(conv3_anat, k=2, s=2)
    print("pooling1_anat shape: " + str(pooling1_anat.shape))  

    conv4_anat = conv3d(pooling1_anat, weights['wc4'], biases['bc4'], s=2)
    print("conv4_anat shape: " + str(conv4_anat.shape))  

    pooling2_anat = maxpool3d(conv4_anat, k=2, s=2)
    print("pooling2_anat shape: " + str(pooling2_anat.shape))  

    fc1 = tf.reshape(pooling2_anat, [-1, weights['wd1'].get_shape().as_list()[0]])
    fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])
    fc1 = tf.nn.relu(fc1)
    print("fc1_anat shape: " + str(fc1.shape))  

    fc2 = tf.reshape(fc1, [-1, weights['wd2'].get_shape().as_list()[0]])
    fc2 = tf.add(tf.matmul(fc2, weights['wd2']), biases['bd2'])
    fc2 = tf.nn.relu(fc2)
    print("fc2_anat shape: " + str(fc2.shape))  

    out = tf.add(tf.matmul(fc2, weights['out']), biases['out'])
    print("out_anat shape: " + str(out.shape))  

    return out

pred = conv_net(anat_x, weights_anat, biases_anat)

cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=pred, labels=y))

optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)

在训练整个数据集时权重不会更新

0 个答案: