TensorFlow卷积层报错：输入为float tensor要求为double tensor如何解决？

问题根源 #

• 你混淆了TensorFlow和PyTorch的API：当前代码使用的是TensorFlow框架，张量类型为EagerTensor，而你写的.type()、torch.from_numpy、DoubleTensor均为PyTorch语法，完全不兼容TensorFlow，因此会报属性不存在的错误
• conv.double()写法无效：你定义的conv是lambda函数，本身没有double方法，这行代码无法起到转换张量类型的作用
• 代码存在逻辑错误：函数参数传入的是x，但你内部调用了全局变量X_train_tf，后续复用该层会出现异常

正确解决方法 #

提供两种可行方案，按需选择即可：

方案1：转换所有输入和参数为double类型（匹配conv要求的double输入） #

修改后的卷积层代码如下：

import tensorflow as tf

def convLayer(x, kHeight, kWidth, strideX, strideY,
              featureNum, name, padding = "SAME", groups = 1):
    """convolution"""
    # 用函数参数x，不要硬编码全局的X_train_tf
    channel = int(x.get_shape()[-1])
    # 卷积计算时统一将输入转为double类型，TensorFlow中double对应tf.float64
    def conv(a, b):
        a_double = tf.cast(a, dtype=tf.float64)
        b_double = tf.cast(b, dtype=tf.float64)
        return tf.nn.conv2d(a_double, b_double, strides = [1, strideY, strideX, 1], padding = padding)

    with tf.compat.v1.variable_scope(name) as scope:
        # 权重参数直接定义为float64类型，避免后续转换损耗
        w =  tf.compat.v1.get_variable("w", 
                                      shape = [kHeight, kWidth, int(channel/groups), featureNum],
                                      dtype=tf.float64)
        b =  tf.compat.v1.get_variable("b", 
                                      shape = [featureNum],
                                      dtype=tf.float64)
        
        xNew = tf.split(value = x, num_or_size_splits = groups, axis = 3)
        wNew = tf.split(value = w, num_or_size_splits = groups, axis = 3)
    
        featureMap = [conv(t1, t2) for t1, t2 in zip(xNew, wNew)]
        mergeFeatureMap = tf.concat(axis = 3, values = featureMap)
        print(mergeFeatureMap.shape)
        out = tf.nn.bias_add(mergeFeatureMap, b)
        return tf.nn.relu(tf.reshape(out, mergeFeatureMap.get_shape().as_list()), name = scope.name)

方案2：调整卷积要求为float类型（更推荐，训练速度更快、显存占用更低） #

如果没有强制的double精度要求，直接将所有参数和输入统一为float32即可，无需转double：

import tensorflow as tf

def convLayer(x, kHeight, kWidth, strideX, strideY,
              featureNum, name, padding = "SAME", groups = 1):
    """convolution"""
    channel = int(x.get_shape()[-1])
    conv = lambda a, b: tf.nn.conv2d(a, b, strides = [1, strideY, strideX, 1], padding = padding)

    with tf.compat.v1.variable_scope(name) as scope:
        # 显式指定权重为float32，和TensorFlow默认输入类型匹配
        w =  tf.compat.v1.get_variable("w", 
                                      shape = [kHeight, kWidth, int(channel/groups), featureNum],
                                      dtype=tf.float32)
        b =  tf.compat.v1.get_variable("b", 
                                      shape = [featureNum],
                                      dtype=tf.float32)
        
        xNew = tf.split(value = x, num_or_size_splits = groups, axis = 3)
        wNew = tf.split(value = w, num_or_size_splits = groups, axis = 3)
    
        featureMap = [conv(t1, t2) for t1, t2 in zip(xNew, wNew)]
        mergeFeatureMap = tf.concat(axis = 3, values = featureMap)
        print(mergeFeatureMap.shape)
        out = tf.nn.bias_add(mergeFeatureMap, b)
        return tf.nn.relu(tf.reshape(out, mergeFeatureMap.get_shape().as_list()), name = scope.name)

注意事项 #

• TensorFlow中张量类型转换统一使用tf.cast(张量, dtype=目标类型)，float32对应单精度float，float64对应双精度double
• 不要跨框架混用API，使用TensorFlow时所有张量操作都调用tf开头的方法，不要调用PyTorch相关方法
• 自定义层尽量使用传入的参数，不要硬编码全局变量，避免后续调用出现异常

内容的提问来源于stack exchange，提问作者Francisco Nieto