Issue when using Keras NetVLAD with functional api

[fede-vaccaro]

Hi, First of all, thanks for this repository. I'm trying to stack your NetVLAD implementation in keras on the top of a ResNet50, but i'm having some issue doing it.

That's my code:

# instantiating the resnet50
resnet = ResNet50(weights='imagenet', include_top=False, pooling=False, 
input_shape=input_shape)
# deleting the default Input layer
resnet.layers.pop(0)
input_q = Input(shape=(224, 224, 3))
# stacking the resnet on a new input (seems useless but i need this to having multiple inputs)
resnet_q = resnet(input_q)
# permuting the tensor from resnet last layer shape which is (None, 7,7,2048) to (None, 2048, 7,7)
transpose = Permute((3,1,2), input_shape=(7,7,2048))(resnet_p)
# reshaping from (2048, 7,7) to (2048, 7*7), in the form (max_samples, feature_size) like requested from NetVLAD
reshape = Reshape((2048, 7*7))(transpose)
# stacking netvlad
netvlad = NetVLAD(feature_size=7*7, max_samples=2048, cluster_size=64, output_dim=1024)(reshape)

But i'm getting this error:

 ---------------------------------------------------------------------------

 ValueError                                Traceback (most recent call last)

 <ipython-input-20-7cb139e75b78> in <module>
      19 transpose = Permute((3,1,2), input_shape=(7,7,2048))(resnet_p)
      20 reshape = Reshape((2048, 7*7))(transpose)
 ---> 21 netvlad = NetVLAD(feature_size=7*7, max_samples=2048, cluster_size=64, output_dim=1024)(reshape)

 ~/Documenti/netvlad/venv/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py in symbolic_fn_wrapper(*args, **kwargs)
      73         if _SYMBOLIC_SCOPE.value:
      74             with get_graph().as_default():
 ---> 75                 return func(*args, **kwargs)
      76         else:
      77             return func(*args, **kwargs)

 ~/Documenti/netvlad/venv/lib/python3.6/site-packages/keras/engine/base_layer.py in __call__(self, inputs, **kwargs)
     461                                          'You can build it manually via: '
     462                                          '`layer.build(batch_input_shape)`')
 --> 463                 self.build(unpack_singleton(input_shapes))
     464                 self.built = True
     465 

 ~/Documenti/netvlad/loupe_keras.py in build(self, input_shape)
      81         self.cluster_weights = self.add_weight(name='kernel_W1',
      82                                       shape=(self.feature_size, self.cluster_size),
 ---> 83                                       initializer=tf.random_normal_initializer(stddev=1 / math.sqrt(self.feature_size)),
      84                                       trainable=True)
      85         self.cluster_biases = self.add_weight(name='kernel_B1',

 ~/Documenti/netvlad/venv/lib/python3.6/site-packages/keras/engine/base_layer.py in add_weight(self, name, shape, dtype, initializer, regularizer, trainable, constraint)
     280                             dtype=dtype,
     281                             name=name,
 --> 282                             constraint=constraint)
     283         if regularizer is not None:
     284             with K.name_scope('weight_regularizer'):

 ~/Documenti/netvlad/venv/lib/python3.6/site-packages/keras/backend/tensorflow_backend.py in variable(value, dtype, name, constraint)
     618     """
     619     v = tf_keras_backend.variable(
 --> 620         value, dtype=dtype, name=name, constraint=constraint)
     621     if hasattr(value, 'tocoo'):
     622         v._keras_shape = value.tocoo().shape

 ~/Documenti/netvlad/venv/lib/python3.6/site-packages/tensorflow_core/python/keras/backend.py in variable(value, dtype, name, constraint)
     809       dtype=dtypes_module.as_dtype(dtype),
     810       name=name,
 --> 811       constraint=constraint)
     812   if isinstance(value, np.ndarray):
     813     v._keras_shape = value.shape

 ~/Documenti/netvlad/venv/lib/python3.6/site-packages/tensorflow_core/python/ops/variables.py in __call__(cls, *args, **kwargs)
     258       return cls._variable_v1_call(*args, **kwargs)
     259     elif cls is Variable:
 --> 260       return cls._variable_v2_call(*args, **kwargs)
     261     else:
     262       return super(VariableMetaclass, cls).__call__(*args, **kwargs)

 ~/Documenti/netvlad/venv/lib/python3.6/site-packages/tensorflow_core/python/ops/variables.py in _variable_v2_call(cls, initial_value, trainable, validate_shape, caching_device, name, variable_def, dtype, import_scope, constraint, synchronization, aggregation, shape)
     252         synchronization=synchronization,
     253         aggregation=aggregation,
 --> 254         shape=shape)
     255 
     256   def __call__(cls, *args, **kwargs):

 ~/Documenti/netvlad/venv/lib/python3.6/site-packages/tensorflow_core/python/ops/variables.py in <lambda>(**kws)
     233                         shape=None):
     234     """Call on Variable class. Useful to force the signature."""
 --> 235     previous_getter = lambda **kws: default_variable_creator_v2(None, **kws)
     236     for _, getter in ops.get_default_graph()._variable_creator_stack:  # pylint: disable=protected-access
     237       previous_getter = _make_getter(getter, previous_getter)

 ~/Documenti/netvlad/venv/lib/python3.6/site-packages/tensorflow_core/python/ops/variable_scope.py in default_variable_creator_v2(next_creator, **kwargs)
    2554       synchronization=synchronization,
    2555       aggregation=aggregation,
 -> 2556       shape=shape)
    2557 
    2558 

 ~/Documenti/netvlad/venv/lib/python3.6/site-packages/tensorflow_core/python/ops/variables.py in __call__(cls, *args, **kwargs)
     260       return cls._variable_v2_call(*args, **kwargs)
     261     else:
 --> 262       return super(VariableMetaclass, cls).__call__(*args, **kwargs)
     263 
     264 

 ~/Documenti/netvlad/venv/lib/python3.6/site-ackages/tensorflow_core/python/ops/resource_variable_ops.py in __init__(self, initial_value, trainable, collections, validate_shape, caching_device, name, dtype, variable_def, import_scope, constraint, distribute_strategy, synchronization, aggregation, shape)
    1404           aggregation=aggregation,
    1405           shape=shape,
 -> 1406           distribute_strategy=distribute_strategy)
    1407 
    1408   def _init_from_args(self,

 ~/Documenti/netvlad/venv/lib/python3.6/site-packages/tensorflow_core/python/ops/resource_variable_ops.py in _init_from_args(self, initial_value, trainable, collections, caching_device, name, dtype, constraint, synchronization, aggregation, distribute_strategy, shape)
    1487     if isinstance(initial_value, ops.Tensor) and hasattr(
    1488         initial_value, "graph") and initial_value.graph.building_function:
 -> 1489       raise ValueError("Tensor-typed variable initializers must either be "
    1490                        "wrapped in an init_scope or callable "
    1491                        "(e.g., `tf.Variable(lambda : "

 ValueError: Tensor-typed variable initializers must either be wrapped in an init_scope or callable (e.g., `tf.Variable(lambda : tf.truncated_normal([10, 40]))`) when building functions. Please file a feature request if this restriction inconveniences you.

Do you have any idea what could be the issue here? I really can't figure out, i looked at your code and i did some variation trying to spot the bug but nothing changed. Thank you!

[fede-vaccaro]

I should have solved it: the tensorflow weight initializer doesn't work with the last version of keras/tensorflow (keras==2.3.1/tensorflow-gpu==2.0.0) I changed your initializer to:

initializer=keras.initializers.random_normal(stddev=1 / math.sqrt(self.feature_size)),

[shamangary]

Hello sorry for the inconvenience but I am in a middle of something right now