Keras: ValueError: Input tensors to a Model must come from keras.layers.Input. Received: keras.engine.input_layer.InputLayer object

[HansLeonardVanBrueggemann]

Hi all,

quick question: in the examples of the applications webpage (https://keras.io/applications/) there is a section for extracting features in a trained model. In this case you pick the input and the layer of interest in the architecture and build the model as follows:

base_model = VGG19(weights='imagenet')
model = Model(inputs=base_model.input, outputs=base_model.get_layer('block4_pool').output)

So far so good. The problem is when one would like to do the exact same thing on a model that has been trained as submodel of another network. Here is an example:

# Get a trained model
model = VGG16(weights='imagenet', include_top=False)

# embedding into another model
inflow = layers.Input((256, 256, 3))
x = ... some layers ... (inflow)
latent = model(x)
outflow = layers.Dense(128)(latent)
new_model = models.Model(inflow, outflow)

# training
[ ... ]

# extract model
old_model = new_model.get_layer('vgg16')
old_model= models.Model(inputs=old_model.input, output=old_model.get_layer('block5_pool').output)

And error message pops up on the screen

ValueError  Traceback (most recent call last)
[...]
C:\Anaconda2\envs\tensorflow\lib\site-packages\keras\engine\network.py in _init_graph_network(self, inputs, outputs, name)
    163                                  'must come from `keras.layers.Input`. '
    164                                  'Received: ' + str(x) +
--> 165                                  ' (missing previous layer metadata).')
    166             # Check that x is an input tensor.
    167             layer, node_index, tensor_index = x._keras_history

ValueError: Input tensors to a Model must come from `keras.layers.Input`. Received: <keras.engine.input_layer.InputLayer object at 0x0000000006712EB8> (missing previous layer metadata).

I really cannot figure out what is the problem. Any help?

[gabrieldemarmiesse]

This is an edge case, but I believe it's a bug.

[HansLeonardVanBrueggemann]

@gabrieldemarmiesse do you know any way around it?

[gabrieldemarmiesse]

Try to reference the original model. Don't use model.layers.

[HansLeonardVanBrueggemann]

@gabrieldemarmiesse unfortunately that I can't in the project I'm working in. Anyhow thanks for the help so far ;-)

[IS5882]

@HansLeonardVanBrueggemann any luck? I'm facing the same issue

[Kailegh]

Same here, this seems like something that would need to be fixed

[keerthi165]

I am facing the same issue. How do I fix it?

[akashsara]

Heya, facing the same issue here. Using a version of this - a sequence-to-sequence model on a word level with embeddings. The training part works perfectly and if I run the code right after training the model, it works. But if load the model after training, I run into this issue.

    encoder_inputs = model.get_layer('input_1')
    decoder_inputs = model.get_layer('input_2')
    decoder_embedding = model.get_layer('embedding_2')
    encoder_lstm = model.get_layer('lstm_1')
    decoder_lstm = model.get_layer('lstm_2')
    decoder_dense = model.get_layer('dense_1')

    _, enc_state_h, enc_state_c = encoder_lstm.output
    encoder_states = [enc_state_h, enc_state_c]

    # Error occurs at this line.
    encoder_model = Model(encoder_inputs, encoder_states)

ValueError: Input tensors to a Model must come from keras.layers.Input. Received: (missing previous layer metadata).

Edit: Seems like this was unrelated. I solved the issue. I was passing the layer itself instead of the input into the function. For anyone who finds this via Google or something: It should actually be encoder_inputs.output or alternatively model.input[0]

Refer: https://github.com/keras-team/keras/blob/master/examples/lstm_seq2seq_restore.py

On a side note, am I right in assuming that model.input[0] = input_layer.input = input_layer.output?

[bozcani]

I encountered the same problem while implementing YOLO. I don't know the exact reason for this error. Then, I used Sequential API instead of Functional API, and it worked for me. Maybe you might convert your code to Sequential counterpart. It should appear like this:

```# Get a trained model
model = VGG16(weights='imagenet', include_top=False)

embedding into another model

It is better to use Sequential API instead of Functional API

in order to extract submodel.

new_model = Seqeuntial()

DEFINE YOUR CUSTOM LAYERS

input = (Input(shape=(256, 256, 3), name='input'))
x = ...
... Add other layers
outflow = Dense(128)

ADD LAYERS TO YOUR MODEL.

Add initial custom layers.

new_model.add(input)
new_model.add(x)
... add other layers ...

Add VGG layers also

for layer in model.layers:
new_model.add(layer)

Add output layer.

new_model.add(outflow)

training

[ ... ]

extract model

new_model.layers is a list of layers including VGG layers that you added.

You can pick up layers from this list and add your new model.

start_index = ... the index of the first layer of VGG in new_model.layers list...
end_index = ... the index of the 'block5_pool' in new_model.layers...

old_model = Sequential()
for layer in new_model.layers[start_index:end_index]:
old_model.add(layer)
```

[adiv5]

I was getting the same error on Functional API. But i solved the error by making sure the variable that stores the Input layer(in my case model_input) and variable of the last layer(in my case preds) are properly mentioned in

keras.Model(model_input, preds)

Just this small check solved it for me.I think Keras API is not the problem. This seems like edge case.

[ASRodrigo1]

I was trying to put 2 nets together with the Model command:
model = Model(inputs=base_mode, output=top)
And getting the same error.

So i solved using Sequential:
model = Sequential()
model.add(base_model)
model.add(top)