CRF Layer fails tensor dtype cast on Keras 2.2.5

[dataframing]

Hi all,

I noticed this morning that the recent release of Keras 2.2.5 breaks this library's implementation of the CRF layer.

I ran the official CRF example here using keras==2.2.4 and was able to successfully train both the CRF and bi-LSTM CRF. However, upgrading to keras==2.2.5 led to the following issue:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/tensorflow/python/framework/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)
    470                 preferred_dtype=default_dtype,
--> 471                 as_ref=input_arg.is_ref)
    472             if input_arg.number_attr and len(

~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/tensorflow/python/framework/ops.py in internal_convert_n_to_tensor(values, dtype, name, as_ref, preferred_dtype, ctx)
   1292             preferred_dtype=preferred_dtype,
-> 1293             ctx=ctx))
   1294   return ret

~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/tensorflow/python/framework/ops.py in internal_convert_to_tensor(value, dtype, name, as_ref, preferred_dtype, ctx, accept_symbolic_tensors, accept_composite_tensors)
   1223     if ret is None:
-> 1224       ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
   1225 

~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/tensorflow/python/framework/ops.py in _TensorTensorConversionFunction(t, dtype, name, as_ref)
   1017         "Tensor conversion requested dtype %s for Tensor with dtype %s: %r" %
-> 1018         (dtype.name, t.dtype.name, str(t)))
   1019   return t

ValueError: Tensor conversion requested dtype bool for Tensor with dtype float32: 'Tensor("crf_1/zeros_like_4:0", shape=(?, ?), dtype=float32)'

During handling of the above exception, another exception occurred:

TypeError                                 Traceback (most recent call last)
<ipython-input-1-6055c06dd262> in <module>
     80 model.add(Embedding(len(vocab), EMBED_DIM, mask_zero=True))  # Random embedding
     81 crf = CRF(len(class_labels), sparse_target=True)
---> 82 model.add(crf)
     83 model.summary()
     84 

~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/keras/engine/sequential.py in add(self, layer)
    180                 self.inputs = network.get_source_inputs(self.outputs[0])
    181         elif self.outputs:
--> 182             output_tensor = layer(self.outputs[0])
    183             if isinstance(output_tensor, list):
    184                 raise TypeError('All layers in a Sequential model '

~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/keras/engine/base_layer.py in __call__(self, inputs, **kwargs)
    449             # Actually call the layer,
    450             # collecting output(s), mask(s), and shape(s).
--> 451             output = self.call(inputs, **kwargs)
    452             output_mask = self.compute_mask(inputs, previous_mask)
    453 

~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/keras_contrib/layers/crf.py in call(self, X, mask)
    290 
    291         if self.test_mode == 'viterbi':
--> 292             test_output = self.viterbi_decoding(X, mask)
    293         else:
    294             test_output = self.get_marginal_prob(X, mask)

~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/keras_contrib/layers/crf.py in viterbi_decoding(self, X, mask)
    562                 input_energy, mask, self.left_boundary, self.right_boundary)
    563 
--> 564         argmin_tables = self.recursion(input_energy, mask, return_logZ=False)
    565         argmin_tables = K.cast(argmin_tables, 'int32')
    566 

~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/keras_contrib/layers/crf.py in recursion(self, input_energy, mask, go_backwards, return_sequences, return_logZ, input_length)
    514 
    515         if mask is not None:
--> 516             mask2 = K.cast(K.concatenate([mask, K.zeros_like(mask[:, :1])], axis=1),
    517                            K.floatx())
    518             constants.append(mask2)

~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/keras/backend/tensorflow_backend.py in concatenate(tensors, axis)
   2174         return tf.sparse_concat(axis, tensors)
   2175     else:
-> 2176         return tf.concat([to_dense(x) for x in tensors], axis)
   2177 
   2178 

~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/tensorflow/python/util/dispatch.py in wrapper(*args, **kwargs)
    178     """Call target, and fall back on dispatchers if there is a TypeError."""
    179     try:
--> 180       return target(*args, **kwargs)
    181     except (TypeError, ValueError):
    182       # Note: convert_to_eager_tensor currently raises a ValueError, not a

~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/tensorflow/python/ops/array_ops.py in concat(values, axis, name)
   1297               tensor_shape.scalar())
   1298       return identity(values[0], name=scope)
-> 1299   return gen_array_ops.concat_v2(values=values, axis=axis, name=name)
   1300 
   1301 

~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/tensorflow/python/ops/gen_array_ops.py in concat_v2(values, axis, name)
   1254   _attr_N = len(values)
   1255   _, _, _op = _op_def_lib._apply_op_helper(
-> 1256         "ConcatV2", values=values, axis=axis, name=name)
   1257   _result = _op.outputs[:]
   1258   _inputs_flat = _op.inputs

~/.local/share/virtualenvs/information-collection-model-IthLzmFg/lib/python3.7/site-packages/tensorflow/python/framework/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)
    497                                 (prefix, dtype.name))
    498               else:
--> 499                 raise TypeError("%s that don't all match." % prefix)
    500             else:
    501               raise TypeError(

TypeError: Tensors in list passed to 'values' of 'ConcatV2' Op have types [bool, float32] that don't all match.

From the 2.2.5 release notes, I can only suspect that the point regarding the new dtype argument on the base keras.Layer could be introducing some sort of breaking change. This issue may also be a duplicate of #498 .

cc: @lzfelix

[WEN-MIN]

Hi, I got the same question when keras update to 2.25. if there any solution for this problem except back keras version to 2.24?

[lan2720]

the same question...

[RayX-X]

so is this can be solved? or just waiting for tensorflow addons?

[FanYaning]

for Keras 2.2.5 --> 516 mask2 = K.cast(K.concatenate([mask, K.zeros_like(mask[:, :1])], axis=1), 517 K.floatx())

try change to mask2 = K.concatenate([K.cast(mask, K.floatx()), K.zeros_like(mask[:, :1])], axis=1)

[kdja90]

@FanYaning I use keras 2.6.0 i not fined where this line
--> 516 mask2 = K.cast(K.concatenate([mask, K.zeros_like(mask[:, :1])], axis=1), 517 K.floatx()) which file is it in?

[FanYaning]

@kdja90

https://github.com/keras-team/keras-contrib/blob/master/keras_contrib/layers/crf.py in this file.