Working with dynamic input shape

[Florient CHOUTEAU]

Hi,

When trying to apply the GroupNormalization Layer to fully convolutionnal networks with a dynamic input shape (as there are not mathematical constraints for a fixed input shape and keras usually allows most of its layers to have a dynamic input shape) i have found the Layer unable to work. This may be due to the K.reshape call that only allows one dimension (the batch size) to be None,

Would you have a solution in mind for GroupNormalization to work w/ dynamic input shapes ? (Is it even possible to implement it this way, like BN ?)

Regards,

---

Here is the code I'm trying to launch:

i = Input(shape=(None,None,3))
c = Conv2D(64, (3,3), padding="same")(i)
o = GroupNormalization(axis=3)(c)
m = Model(inputs=i,outputs=o)

And the error:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
/opt/conda/lib/python3.6/site-packages/tensorflow/python/framework/tensor_util.py in make_tensor_proto(values, dtype, shape, verify_shape)
    516     try:
--> 517       str_values = [compat.as_bytes(x) for x in proto_values]
    518     except TypeError:

/opt/conda/lib/python3.6/site-packages/tensorflow/python/framework/tensor_util.py in <listcomp>(.0)
    516     try:
--> 517       str_values = [compat.as_bytes(x) for x in proto_values]
    518     except TypeError:

/opt/conda/lib/python3.6/site-packages/tensorflow/python/util/compat.py in as_bytes(bytes_or_text, encoding)
     66     raise TypeError('Expected binary or unicode string, got %r' %
---> 67                     (bytes_or_text,))
     68 

TypeError: Expected binary or unicode string, got -1

During handling of the above exception, another exception occurred:

TypeError                                 Traceback (most recent call last)
<ipython-input-5-91a4d6819a5e> in <module>()
      1 i = Input(shape=(None,None,3))
      2 c = Conv2D(64, (3,3), padding="same")(i)
----> 3 o = GroupNormalization(axis=3)(c)
      4 m = Model(inputs=i,outputs=o)

/opt/conda/lib/python3.6/site-packages/keras/engine/topology.py in __call__(self, inputs, **kwargs)
    617 
    618             # Actually call the layer, collecting output(s), mask(s), and shape(s).
--> 619             output = self.call(inputs, **kwargs)
    620             output_mask = self.compute_mask(inputs, previous_mask)
    621 

/home/code/keras_shipdetection/layers/group_normalization.py in call(self, inputs, **kwargs)
    135         needs_broadcasting = (sorted(reduction_axes) != list(range(ndim))[:-1])
    136 
--> 137         inputs = K.reshape(inputs, group_shape)
    138 
    139         mean = K.mean(inputs, axis=group_reduction_axes[2:], keepdims=True)

/opt/conda/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py in reshape(x, shape)
   1896         A tensor.
   1897     """
-> 1898     return tf.reshape(x, shape)
   1899 
   1900 

/opt/conda/lib/python3.6/site-packages/tensorflow/python/ops/gen_array_ops.py in reshape(tensor, shape, name)
   6111   if _ctx is None or not _ctx._eager_context.is_eager:
   6112     _, _, _op = _op_def_lib._apply_op_helper(
-> 6113         "Reshape", tensor=tensor, shape=shape, name=name)
   6114     _result = _op.outputs[:]
   6115     _inputs_flat = _op.inputs

/opt/conda/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)
    511           except TypeError as err:
    512             if dtype is None:
--> 513               raise err
    514             else:
    515               raise TypeError(

/opt/conda/lib/python3.6/site-packages/tensorflow/python/framework/op_def_library.py in _apply_op_helper(self, op_type_name, name, **keywords)
    508                 dtype=dtype,
    509                 as_ref=input_arg.is_ref,
--> 510                 preferred_dtype=default_dtype)
    511           except TypeError as err:
    512             if dtype is None:

/opt/conda/lib/python3.6/site-packages/tensorflow/python/framework/ops.py in internal_convert_to_tensor(value, dtype, name, as_ref, preferred_dtype, ctx)
   1102 
   1103     if ret is None:
-> 1104       ret = conversion_func(value, dtype=dtype, name=name, as_ref=as_ref)
   1105 
   1106     if ret is NotImplemented:

/opt/conda/lib/python3.6/site-packages/tensorflow/python/framework/constant_op.py in _constant_tensor_conversion_function(v, dtype, name, as_ref)
    233                                          as_ref=False):
    234   _ = as_ref
--> 235   return constant(v, dtype=dtype, name=name)
    236 
    237 

/opt/conda/lib/python3.6/site-packages/tensorflow/python/framework/constant_op.py in constant(value, dtype, shape, name, verify_shape)
    212   tensor_value.tensor.CopyFrom(
    213       tensor_util.make_tensor_proto(
--> 214           value, dtype=dtype, shape=shape, verify_shape=verify_shape))
    215   dtype_value = attr_value_pb2.AttrValue(type=tensor_value.tensor.dtype)
    216   const_tensor = g.create_op(

/opt/conda/lib/python3.6/site-packages/tensorflow/python/framework/tensor_util.py in make_tensor_proto(values, dtype, shape, verify_shape)
    519       raise TypeError("Failed to convert object of type %s to Tensor. "
    520                       "Contents: %s. Consider casting elements to a "
--> 521                       "supported type." % (type(values), values))
    522     tensor_proto.string_val.extend(str_values)
    523     return tensor_proto

TypeError: Failed to convert object of type <class 'list'> to Tensor. Contents: [-1, 32, None, None, 2]. Consider casting elements to a supported type.

I am running tensorflow 1.8.0 and keras 2.1.6

[Somshubra Majumdar]

As you have seen, reshape allows only 1 dimension to be unknown, therefore it is not possible to have dynamic shape here.