All layer names should be unique.

[Tahedi1]

I tested the package methods with VGG16 and it works. But, I have tested it with trained InceptionV3 model constructed in separate train script as following:

base_model = InceptionV3(include_top=False, weights='imagenet',input_shape = (299,299,3))
x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
x = Dropout(0.2)(x)
predictions = Dense(nbr_classes, activation='softmax')(x)
model = Model(input=base_model.input, output=predictions)

And loaded in visualisation script as following (I based on imagenet_compare_methods.ipynb) :

model_path = 'path_of_model'
model_weights_path = 'path_of_model_weights'
model = load_model(model_path)
model.load_weights(model_weights_path)
model = keras.models.Model(model.inputs, model.outputs)
model.compile(optimizer="adam", loss="categorical_crossentropy")
preprocess_f=keras.applications.inception_v3.preprocess_input
color_coding="RGB"
channels_first = keras.backend.image_data_format() == "channels_first"
color_conversion = "BGRtoRGB" if color_coding == "BGR" else None
input_range = (-1,1)
noise_scale = (input_range[1]-input_range[0]) * 0.1

I got the following error with iutils.keras.graph.model_wo_softmax function:

Traceback (most recent call last):
  File "imagenet_compare_methods_incepV3.py", line 91, in <module>
    model_wo_softmax = iutils.keras.graph.model_wo_softmax(model)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/innvestigate/utils/keras/graph.py", line 312, in model_wo_softmax
    name=model.name)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/keras/engine/network.py", line 93, in __init__
    self._init_graph_network(*args, **kwargs)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/keras/engine/network.py", line 231, in _init_graph_network
    self.inputs, self.outputs)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/keras/engine/network.py", line 1455, in _map_graph_network
    ' times in the model. '
ValueError: The name "dense_1" is used 2 times in the model. All layer names should be unique.

To handle this error I have modified the names of my dense layers and I added this code:

model.layers[-3].name='dense1'
model.layers[-1].name='dense_final'

These methodes [Input, Gradient] work proprly and SmoothGrad works but with out of memory warning. But with Guided Backprop I got an another error :

Traceback (most recent call last):
  File "imagenet_compare_methods_incepV3.py", line 133, in <module>
    a = analyzer.analyze(x_pp)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/innvestigate/analyzer/base.py", line 479, in analyze
    self.create_analyzer_model()
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/innvestigate/analyzer/base.py", line 443, in create_analyzer_model
    outputs=analysis_outputs+debug_outputs)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/keras/engine/network.py", line 93, in __init__
    self._init_graph_network(*args, **kwargs)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/keras/engine/network.py", line 231, in _init_graph_network
    self.inputs, self.outputs)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/keras/engine/network.py", line 1455, in _map_graph_network
    ' times in the model. '
ValueError: The name "activation_1" is used 2 times in the model. All layer names should be unique.

Please help me to solve this issue.

[albermax]

How about:

base_model = InceptionV3(include_top=False, weights='imagenet',input_shape = (299,299,3))
x = base_model.output
x = GlobalAveragePooling2D(name='special_name_01')(x)
x = Dense(1024, activation='relu', name='special_name_02')(x)
x = Dropout(0.2, name='special_name_03')(x)
predictions = Dense(nbr_classes, activation='softmax', name='special_name_04')(x)
model = Model(input=base_model.input, output=predictions)

[Tahedi1]

I will retrain the model as you suggested and I will test the script.
The problem is I have already many trained model and I would like to use them directly. I have examined all layers names and I found them with differents names.
I dont understand why this error is appeared.

[albermax]

Retraining will not help you, I did also not suggest a that.
If you look I suggested to change the name of the layers before you apply them.

The error appears because Keras does accidentally reuse layer names.

[Tahedi1]

Ok, albermax. I will test your proposition.

[Tahedi1]

In ordre to modify layers name as you sugested :

model.layers[-1].name='special_name_04'
model.layers[-2].name='special_name_03'
model.layers[-3].name='special_name_02'
model.layers[-4].name='special_name_01'

But I have the same error:

Traceback (most recent call last):
  File "visio_incepV3.py", line 134, in <module>
    a = analyzer.analyze(x_pp)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/innvestigate/analyzer/base.py", line 479, in analyze
    self.create_analyzer_model()
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/innvestigate/analyzer/base.py", line 443, in create_analyzer_model
    outputs=analysis_outputs+debug_outputs)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/keras/legacy/interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/keras/engine/network.py", line 93, in __init__
    self._init_graph_network(*args, **kwargs)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/keras/engine/network.py", line 231, in _init_graph_network
    self.inputs, self.outputs)
  File "/home_nfs/brahimim/.conda/envs/mohammed.tf.1.9.0/lib/python3.6/site-packages/keras/engine/network.py", line 1455, in _map_graph_network
    ' times in the model. '
ValueError: The name "activation_1" is used 2 times in the model. All layer names should be unique.

[albermax]

I am afraid but this is not what I suggested.
This worked for me:

base_model = InceptionV3(include_top=False, weights='imagenet',input_shape = (299,299,3))
x = base_model.output
x = GlobalAveragePooling2D(name='special_name_01')(x)
x = Dense(1024, activation='relu', name='special_name_02')(x)
x = Dropout(0.2, name='special_name_03')(x)
predictions = Dense(nbr_classes, activation='softmax', name='special_name_04')(x)
model = Model(input=base_model.input, output=predictions)
# workaround to get unique names
for l in model.layers:
  l.name = "%s_workaround" % l.name
# Create model with new names
model = Model(input=model.input, output=model.output)

Hope this helps!

[Tahedi1]

I'm not working with imagnet dataset. For this reason, I'm not using your code. I have already trained a model and I have two files (model.h5, model_weights.h5).

[Tahedi1]

I have solved the issue but the problem is analyser is very fast with VGG16 but with inceptionV3 is very slow.

[albermax]

Again you are not very specific. If you need help please provide some example code and more evidence.
We don't experience those problems. Once compiled the code is typically as fast as a gradient computation.

I close the issue your problems is fixed.

[dalia1992]

I have solved the issue but the problem is analyser is very fast with VGG16 but with inceptionV3 is very slow.

Could you tell me how you solved it? I'm having the same issue