使用TensorFlow 踩过的坑

***转载：https://blog.youkuaiyun.com/wang2008start/article/details/71516198***

问题一：TypeError: Expected int32, got list containing Tensors of type ‘_Message’ instead.
tensorflow 函数tf.cocat([fw,bw],2)出错:

Expected int32, got list containing Tensors of type ‘_Message’ inst
查看原因是11版本的函数形式为:tf.concat(2,[fw,bw]),即应把串联的维度与串联值位置调换即可.

问题二：Input ‘split_dim’ of ‘Split’ Op has type float32 that does not match expected type of int32
This is because in Tensorflow versions < 0.12.0 the split function takes the arguments as:

x = tf.split(0, n_steps, x) # tf.split(axis, num_or_size_splits, value)
The tutorial you are working from was written for versions > 0.12.0, which has been changed to be consistent with Numpy’s split syntax:

x = tf.split(x, n_steps, 0) # tf.split(value, num_or_size_splits, axis)

问题三：TypeError: concat() got an unexpected keyword argument ‘axis’
tf.concat(concat_dim=axis, values=inputs, name=name)
修改为： tf.concat(inputs,1,name=name)

问题四：ValueError: ‘size’ must be a 1-D Tensor of 2 elements
img = tf.image.resize_images(img, new_shape[0], new_shape[1])
改为
img = tf.image.resize_images(img, new_shape)

问题五： ‘module’ object has no attribute ‘pack’
因为TF后面的版本修改了这个函数的名称，把 tf.pack 改为 tf.stack。

问题六：The value of a feed cannot be a tf.Tensor object. Acceptable feed values include Python scalars, strings, lists, or numpy ndarrays
数据集是feed输入的，feed的数据格式是有要求的
解决：img,label = sess.run[img,label],用返回值

问题七：AttributeError: ‘module’ object has no attribute ‘per_image_whitening’
For anyone else who has this problem, per_image_whitening was replaced by per_image_standardization in v0.12.

问题八：AttributeError: ‘module’ object has no attribute ‘image_summary’
tf.image_summary should be renamed to tf.summary.image;

问题九：AttributeError: ‘module’ object has no attribute ‘mul’
tf.mul(a,b) 这里的矩阵a和矩阵b的shape必须相等 tf.mul()是矩阵的element-wise相乘（即Hadamard乘积）
tf.matmul(a,b) 这里的矩阵a和矩阵b的shape应是a的行数对等与b的列数，tf.matmul()是矩阵的一般相乘。
解决：[tf.mul，tf.sub ] 和 [tf.neg] 不再使用，改为 [tf.multiply]，[tf.subtract] 和 [tf.negative]。

问题十：AttributeError: ‘module’ object has no attribute ‘scalar_summary’
修改为:tf.summary.scalar(‘batch_loss’, loss)原因:新版本做了调整 …