keras: Keras model cannot be loaded if it contains a Lambda layer calling tf.image.resize_images

Try using custom_objects argument in load_model (or model_from_json) function.

For your case load_model("model_path", custom_objects={"tf": tf}) after importing tensorflow as tf would do the job.

@lauphedo do you defined tf in the Lamba function? I found out if I import tensorflow as tf inside the function used by Lamba layer, and it is fine

got same error when loading a model contains Lambda

load_model(“model_path”, custom_objects={“tf”: tf}) did not work for me with keras 2.0.6 and python3.6

The problem is probably the extra Python lambda that you’re using. Just do this instead:

model.add(Lambda(resize_normalize, input_shape=(80,318,3), output_shape=(66, 200, 3)))

@lauphedo That’s another known issue, see #5088. One workaround is to import inside the lambda function.

(maybe) similar problem: trying to load a model that contains: model_nspect2mask.add(TimeDistributed(Lambda(lambda x: tf.log(tf.abs(x))), input_shape = (None,feat_num)))

get error:

NameError                                 Traceback (most recent call last)
<ipython-input-59-c7a565cc9dba> in <module>()
----> 1 model_nspect2mask = load_model(newexp_folder_path+model_filename)

/usr/local/lib/python2.7/dist-packages/keras/models.pyc in load_model(filepath, custom_objects)
    140         raise ValueError('No model found in config file.')
    141     model_config = json.loads(model_config.decode('utf-8'))
--> 142     model = model_from_config(model_config, custom_objects=custom_objects)
    143 
    144     # set weights

/usr/local/lib/python2.7/dist-packages/keras/models.pyc in model_from_config(config, custom_objects)
    191                         'Maybe you meant to use '
    192                         '`Sequential.from_config(config)`?')
--> 193     return layer_from_config(config, custom_objects=custom_objects)
    194 
    195 

/usr/local/lib/python2.7/dist-packages/keras/utils/layer_utils.pyc in layer_from_config(config, custom_objects)
     41                                        custom_objects=custom_objects)
     42     else:
---> 43         return layer_class.from_config(config['config'])
     44 
     45 

/usr/local/lib/python2.7/dist-packages/keras/models.pyc in from_config(cls, config, layer_cache)
   1084         else:
   1085             layer = get_or_create_layer(first_layer)
-> 1086             model.add(layer)
   1087 
   1088         for conf in config[1:]:

/usr/local/lib/python2.7/dist-packages/keras/models.pyc in add(self, layer)
    297                 else:
    298                     input_dtype = None
--> 299                 layer.create_input_layer(batch_input_shape, input_dtype)
    300 
    301             if len(layer.inbound_nodes) != 1:

/usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in create_input_layer(self, batch_input_shape, input_dtype, name)
    399         # and create the node connecting the current layer
    400         # to the input layer we just created.
--> 401         self(x)
    402 
    403     def add_weight(self, shape, initializer, name=None,

/usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in __call__(self, x, mask)
    570         if inbound_layers:
    571             # This will call layer.build() if necessary.
--> 572             self.add_inbound_node(inbound_layers, node_indices, tensor_indices)
    573             # Outputs were already computed when calling self.add_inbound_node.
    574             outputs = self.inbound_nodes[-1].output_tensors

/usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in add_inbound_node(self, inbound_layers, node_indices, tensor_indices)
    633         # creating the node automatically updates self.inbound_nodes
    634         # as well as outbound_nodes on inbound layers.
--> 635         Node.create_node(self, inbound_layers, node_indices, tensor_indices)
    636 
    637     def get_output_shape_for(self, input_shape):

/usr/local/lib/python2.7/dist-packages/keras/engine/topology.pyc in create_node(cls, outbound_layer, inbound_layers, node_indices, tensor_indices)
    164 
    165         if len(input_tensors) == 1:
--> 166             output_tensors = to_list(outbound_layer.call(input_tensors[0], mask=input_masks[0]))
    167             output_masks = to_list(outbound_layer.compute_mask(input_tensors[0], input_masks[0]))
    168             # TODO: try to auto-infer shape

/usr/local/lib/python2.7/dist-packages/keras/layers/wrappers.pyc in call(self, inputs, mask)
    126             # (nb_samples * timesteps, ...)
    127             inputs = K.reshape(inputs, (-1,) + input_shape[2:])
--> 128             y = self.layer.call(inputs)  # (nb_samples * timesteps, ...)
    129             # (nb_samples, timesteps, ...)
    130             output_shape = self.get_output_shape_for(input_shape)

/usr/local/lib/python2.7/dist-packages/keras/layers/core.pyc in call(self, x, mask)
    638         if 'mask' in arg_spec.args:
    639             arguments['mask'] = mask
--> 640         return self.function(x, **arguments)
    641 
    642     def get_config(self):

/usr/local/lib/python2.7/dist-packages/keras/layers/core.pyc in <lambda>(x)
      8 
      9 # conversion to dB  # f = lambda x: tf.log(tf.abs(x))
---> 10 model_nspect2mask.add(TimeDistributed(Lambda(lambda x: tf.log(tf.abs(x))), input_shape = (None,feat_num)))
     11 
     12 # normalize per feature per batch

NameError: global name 'tf' is not defined ```

Like above, loading works if model does not have that Lambda layer. 

I working with keras-2.2.0 python3.6.4 @adwin5

def Repeat_layer(repeat_times, name):
    # import tensorflow as tf 
    return Lambda(lambda x: tf.tile(tf.expand_dims(x, 2),
                                    tf.stack([1, 1, repeat_times])),
                  name=name)

if I import tensorflow as tf inside the Repeat_layer, it will get error can't pickle module objects when I save model via model.save or model.to_json().

@1524045patrick
I import tensorflow as tf outside Repeat_layer i.e., global scope. And load my model via load_model("model_path", custom_objects={"tf": tf}) , it works for me.

Is this issue fixed? Same issue in TF. 2.x

Yes, as @antonimmo says, using custom_objects works, and it works for any other things that load_model complains about not being found e.g. custom loss functions.

Using tf version 2.4 I still have this problem

@adwin5, @allanzelener, @lauphedo ,

for me that workaround did not work when I am using the lamda layer like this:

model.add(Lambda(lambda x: resize_normalize(x), input_shape=(80,318,3), output_shape=(66, 200, 3)))

and my function looks like this. Itried both tensorflow and keras backend import explicite inside the function:

def resize_normalize(image):
    import tensorflow as tf
    #from keras.backend import tf as ktf
    #resize 
    resized = tf.image.resize_images(image, (66, 200))
    #normalize
    resized = resized/255.0 - 0.5

    return resized

When I load my model I got the following traceback:

@adwin5, @allanzelener Wow, this was fast and really helpful! Thank you very much! Before, I couln’t understand #5088

Interestingly, load_weights() performs well without the mentioned fix…

I have a very similar error with a Lambda layer. I am working with complex numbers and at some point need to consider the real part of my tensors.

Here is a minimal failing example:

from keras.layers import Input, Lambda
from keras.models import Model, load_model
import tensorflow as tf

complex_input = Input((None,), dtype='complex64')
x_real = Lambda(tf.math.real)(complex_input)
model = Model(inputs=complex_input, outputs=x_real)

model.save('tmp.hdf5')

load_model('tmp.hdf5', custom_objects={'tf.math.real': tf.math.real, 'tf': tf, 'tf.math': tf.math})

Which fails with:

TypeError                                 Traceback (most recent call last)
~/workspace/tests/venv/lib/python3.6/site-packages/keras/layers/core.py in wrapper(*args, **kwargs)
    179     try:
--> 180       return target(*args, **kwargs)
    181     except (TypeError, ValueError):

TypeError: 'str' object is not callable

During handling of the above exception, another exception occurred:

NameError                                 Traceback (most recent call last)
<ipython-input-1-883b5508b892> in <module>
      9 model.save('tmp.hdf5')
     10 
---> 11 load_model('tmp.hdf5', custom_objects={'tf.math.real': tf.math.real, 'tf': tf, 'tf.math': tf.math})

~/workspace/tests/venv/lib/python3.6/site-packages/keras/engine/saving.py in load_model(filepath, custom_objects, compile)
    417     f = h5dict(filepath, 'r')
    418     try:
--> 419         model = _deserialize_model(f, custom_objects, compile)
    420     finally:
    421         if opened_new_file:

~/workspace/tests/venv/lib/python3.6/site-packages/keras/engine/saving.py in _deserialize_model(f, custom_objects, compile)
    223         raise ValueError('No model found in config.')
    224     model_config = json.loads(model_config.decode('utf-8'))
--> 225     model = model_from_config(model_config, custom_objects=custom_objects)
    226     model_weights_group = f['model_weights']
    227 

~/workspace/tests/venv/lib/python3.6/site-packages/keras/engine/saving.py in model_from_config(config, custom_objects)
    456                         '`Sequential.from_config(config)`?')
    457     from ..layers import deserialize
--> 458     return deserialize(config, custom_objects=custom_objects)
    459 
    460 

~/workspace/tests/venv/lib/python3.6/site-packages/keras/layers/__init__.py in deserialize(config, custom_objects)
     53                                     module_objects=globs,
     54                                     custom_objects=custom_objects,
---> 55                                     printable_module_name='layer')

~/workspace/tests/venv/lib/python3.6/site-packages/keras/utils/generic_utils.py in deserialize_keras_object(identifier, module_objects, custom_objects, printable_module_name)
    143                     config['config'],
    144                     custom_objects=dict(list(_GLOBAL_CUSTOM_OBJECTS.items()) +
--> 145                                         list(custom_objects.items())))
    146             with CustomObjectScope(custom_objects):
    147                 return cls.from_config(config['config'])

~/workspace/tests/venv/lib/python3.6/site-packages/keras/engine/network.py in from_config(cls, config, custom_objects)
   1030                 if layer in unprocessed_nodes:
   1031                     for node_data in unprocessed_nodes.pop(layer):
-> 1032                         process_node(layer, node_data)
   1033 
   1034         name = config.get('name')

~/workspace/tests/venv/lib/python3.6/site-packages/keras/engine/network.py in process_node(layer, node_data)
    989             # and building the layer if needed.
    990             if input_tensors:
--> 991                 layer(unpack_singleton(input_tensors), **kwargs)
    992 
    993         def process_layer(layer_data):

~/workspace/tests/venv/lib/python3.6/site-packages/keras/engine/base_layer.py in __call__(self, inputs, **kwargs)
    455             # Actually call the layer,
    456             # collecting output(s), mask(s), and shape(s).
--> 457             output = self.call(inputs, **kwargs)
    458             output_mask = self.compute_mask(inputs, previous_mask)
    459 

~/workspace/tests/venv/lib/python3.6/site-packages/keras/layers/core.py in call(self, inputs, mask)
    685         if has_arg(self.function, 'mask'):
    686             arguments['mask'] = mask
--> 687         return self.function(inputs, **arguments)
    688 
    689     def compute_mask(self, inputs, mask=None):

~/workspace/tests/venv/lib/python3.6/site-packages/keras/layers/core.py in wrapper(*args, **kwargs)
    182       # Note: convert_to_eager_tensor currently raises a ValueError, not a
    183       # TypeError, when given unexpected types.  So we need to catch both.
--> 184       result = dispatch(wrapper, *args, **kwargs)
    185       if result is not OpDispatcher.NOT_SUPPORTED:
    186         return result

NameError: name 'dispatch' is not defined

The error is very cryptic because it points only to Keras lines, whereas the code areas referred to are actually in TensorFlow… That’s a first problem, but also as you can see using the custom_objects kwarg, like @antonimmo suggested, doesn’t help fixing the problem.

Any ideas on how to solve this? (as a side note, it’s not actually too much of a problem since I can just recreate the model and load the weights but it would be more convenient if I could just load the model).

Thx alot for commenting this question. So meanwhile I fixed it.

• one resaon and there u are right was the lamda.
• second fix is to define the resize_normalize() inside model architecture like shown below

def architecture():

def resize_normalize(image):
    import cv2
    from keras.backend import tf as ktf   
    
    # resize to width 200 and high 66 liek recommended
    # in the nvidia paper for the used CNN
    # image = cv2.resize(image, (66, 200)) #first try
    resized = ktf.image.resize_images(image, (32, 128))
    #normalize 0-1
    resized = resized/255.0 - 0.5

    return resized

model = Sequential()
model.add(Cropping2D(cropping=((60,20), (1,1)), input_shape=(160,320,3)))
model.add(Lambda(resize_normalize, input_shape=(160, 320, 3), output_shape=(32, 128, 3)))

… …

I’ve run into the same issue, here’s a small gist that reproduces the error.

A different error occurs when deserializing from yaml instead of json. I couldn’t get a Lambda wrapper for K.resize_images to work either.

Edit: The JSON errors seems to be because defaults is serialized as [None] but is expected to be either None or a tuple but not a list. I’ve added PR #5350 to resolve this.

The NameError issue is referenced in #5088 and has a workaround. Simply import what’s needed in a wrapped function. This gist fixes my example above.