TensorFlow运行MNIST测试时出现数据类型转换错误求助

Hey there, let's get this error sorted out right away! The core issue here is integer division behavior in Python 2.7—since you're using Python 2, the expression 1/3 evaluates to 0 (an int32 integer), but your input tensor conv2d(x_image, W_conv1) + b_conv1 is of type float32. TensorFlow throws this mismatch error because it can't safely convert your floating-point tensor to integer type for the Leaky ReLU operation.

Quick Fixes #

You just need to force float division instead of integer division for the alpha parameter. Pick one of these options:

• Use a float literal in the division: alpha=1.0/3
• Or swap the float literal: alpha=1/3.0
• Or explicitly cast to float: alpha=float(1)/3

Your corrected line will look like this:

h_conv1 = tf.nn.leaky_relu(conv2d(x_image, W_conv1) + b_conv1, alpha=1.0/3)

Why This Happened #

In Python 2, dividing two integers with / does floor division (truncates the result to the nearest integer), so 1/3 becomes 0 (an int32 value). TensorFlow's tf.nn.leaky_relu expects the alpha value to match the data type of the input tensor. When you pass an int32 alpha with a float32 tensor, it tries to convert the tensor to int32—which fails because your tensor contains floating-point values that can't be represented as integers.

If you want to avoid this kind of issue long-term, add from __future__ import division at the very top of your script. This makes Python 2 use Python 3-style float division for all / operations, so 1/3 will evaluate to 0.333... instead of 0.

内容的提问来源于stack exchange，提问作者Leo