一维转换错误 - 输入和过滤器大小需要相同？

Question

我正在尝试在 eeg 信号上实现 1d CNN，但我收到一个错误，上面写着

ValueError: 两个形状的维度 1 必须相等，但分别是 492 和 1 将形状 0 与其他形状合并。对于具有 > 输入形状的“MaxPool/input”（操作：“Pack”）：[?,492,64], [50,1,64]。

[?, 492, 64] (batchsize, in_width, channels) 我相信这是第一个 Conv1d 层的输出张量

[50, 1, 64] (filter_width, in_channels, out_channels) 是第一个 Conv1d 权重的形状。

为什么 492 和 1 必须相等？我不理解阻止我发现问题的错误。这是我使用张量流的第一周，我们将不胜感激。谢谢。导致以下错误的代码。

# Convolutional Layer 1s
filter_size_1s = 50
num_filters_1s = 64
stride_1s = 6
# Convolutional Layer 2s , 3s , 4s
filter_size_s = 8
num_filters_s = 128
stride_s = 1

#weights and biases
# filter tensor of shape [filter_width, in_channels, out_channels]
W_1s = tf.Variable(tf.truncated_normal([50, 1, 64], stddev=0.1))  
B_1s = tf.Variable(tf.constant(0.1, tf.float32, [64]))
W_2s = tf.Variable(tf.truncated_normal([8, 64, 128], stddev=0.1))  
B_2s = tf.Variable(tf.constant(0.1, tf.float32, [128]))
W_3s = tf.Variable(tf.truncated_normal([8, 128, 128], stddev=0.1)) 
B_3s = tf.Variable(tf.constant(0.1, tf.float32, [128]))
W_4s = tf.Variable(tf.truncated_normal([8, 128, 128], stddev=0.1))  
B_4s = tf.Variable(tf.constant(0.1, tf.float32, [128]))


def CNN_small(input, phase_test, iteration):

    conv1s = new_conv_layer(input, W_1s, B_1s, stride_1s, phase_test, iteration)
    max_pool1s = tf.nn.max_pool(conv1s, 
                           ksize=[1, pool_size_1s, 1, 1],
                           strides=[1, pool_stride_1s, 1, 1], 
                           padding='VALID')

    dropout_s = tf.nn.dropout(max_pool1s, dropout_prob)    
    conv2s = new_conv_layer(dropout_s, W_2s, B_2s, stride_s, phase_test, iteration)    
    conv3s = new_conv_layer(conv2s, W_3s, B_3s, stride_s, phase_test, iteration)    
    conv4s = new_conv_layer(conv3s, W_4s, B_4s, stride_s, phase_test, iteration)
    max_pool2s = tf.nn.max_pool(conv4s, 
                           ksize=[1, pool_size_2s, 1, 1],
                           strides=[1, pool_stride_2s, 1, 1], 
                           padding='VALID')
    return max_pool2s

def new_conv_layer(input, weights, bias, stride, phase_test, iteration):

    conv = tf.nn.conv1d(value=input, filters=weights, stride=stride, padding='VALID') + bias
    #bn = batch_norm(conv, biases, phase_test, iteration)    #biases added into batch_norm
    activation = tf.nn.relu(conv)
    return activation, weights

x = tf.placeholder(tf.float32, shape=[None, stage_length], name='x')
x_stage = tf.reshape(x, [-1, stage_length, num_channels])   #batch, in_width, channels

#And the line which is giving the error
#cnn layer
max_pool2s = CNN_small(x_stage, phase_test, iteration)

Answer 1

您的new_conv_layer 函数正在返回两个张量，而您正试图像只返回一个一样使用它。如果层后不使用权重，只需将return语句改为

return activation

您还可以找到tf.layers API 清理器。