【keras bug】Tensor is unhashable if Tensor equality is enabled. Instead, use tensor.experimental_ ref(

/media/cfs/wangxin436/TCN/future-guide-model/dissales.py:161 call  *
        loss = tf.add(tf.cast(tf.multiply(tf.cast(weights,l1.dtype), loss_st),'float64'),
    /usr/local/anaconda3/lib/python3.6/site-packages/tensorflow_core/python/util/dispatch.py:180 wrapper
        return target(*args, **kwargs)
    /usr/local/anaconda3/lib/python3.6/site-packages/tensorflow_core/python/ops/math_ops.py:686 cast
        base_type = dtypes.as_dtype(dtype).base_dtype
    /usr/local/anaconda3/lib/python3.6/site-packages/tensorflow_core/python/framework/dtypes.py:716 as_dtype
        return _ANY_TO_TF[type_value]
    /usr/local/anaconda3/lib/python3.6/site-packages/tensorflow_core/python/framework/ops.py:713 __hash__
        raise TypeError("Tensor is unhashable if Tensor equality is enabled. "

    TypeError: Tensor is unhashable if Tensor equality is enabled. Instead, use tensor.experimental_ref() as the key.

When encountering such problems , The first point is the version incompatibility , So what are the solutions to incompatibility ,

First of all , Upgraded version , Be compatible , Usually 2.0 The above will not encounter such problems

keras.__version__
tf.__version__
# Observe code version 

Upgraded version

pip install keras
pip install tensorflow

If it can't be solved

In fact, you can check what is wrong in your code , The problem with my code is tf.cast() Use , stay TF2.0+ In the process of execution , In fact, there is no need for personal designation tensor The type of , The code can automatically match the corresponding type , What we need to do is delete the data type specified by individuals

such as

tf.multiply(tf.cast(weights,l1.dtype), l1) + tf.multiply(tf.cast((1-weights,l2.dtype), l2))


# It is amended as follows 

tf.multiply(weights, l1) + tf.multiply(1-weights, l2)

About the version , You can also have a look at

stackoverflow How to explain , In short, modifying personal code is the right way , Others are made compatible , Of course, you can't get away with other people's code , If you don't want to modify it, you can only upgrade the version