"ValueError: The layer sequential has never been called and thus has no defined output." when the model's been build and called

[Senantq]

I am currently using tensorflow 2.17 with keras 3.4.1 under Ubuntu 24.04 LTS. I have also reproduced the issue with tf-nightly 2.18.0.dev20240731 (keras nightly 3.4.1.dev2024073103).

I encountered the issue when i downloaded a model I have ran on a cluster under tf 2.17/keras 3.4.1. I then tried to obtain some saliency maps on my computer after re-building the model without redifining all its layers from scratch.

See the following google drive for a reprex with the code, model and a data sample: https://drive.google.com/drive/folders/15J_ghWXWbs8EmSVXedH6sJRvJcPUSTIW?usp=sharing

But it raises the following traceback:

`ValueError                                Traceback (most recent call last)
Cell In[1], line 45
     42         class_activation_map = tf.expand_dims(class_activation_map, axis=-1)
     43     return class_activation_map
---> 45 layer_cam_test = layer_cam(img = test_sample, model=model, label_index = 0)

Cell In[1], line 24, in layer_cam(img, label_index, model)
     22 print(layer_names)
     23 for layer_name in layer_names[-1:]:
---> 24     grad_model = tf.keras.models.Model([model.inputs], [model.get_layer(layer_name).output, model.output]) #bug's here
     25     with tf.GradientTape() as tape:
     26         tape.watch(img)

File ~/miniconda3/envs/envtfnightly/lib/python3.11/site-packages/keras/src/ops/operation.py:266, in Operation.output(self)
    256 @property
    257 def output(self):
    258     """Retrieves the output tensor(s) of a layer.
    259 
    260     Only returns the tensor(s) corresponding to the *first time*
   (...)
    264         Output tensor or list of output tensors.
    265     """
--> 266     return self._get_node_attribute_at_index(0, "output_tensors", "output")

File ~/miniconda3/envs/envtfnightly/lib/python3.11/site-packages/keras/src/ops/operation.py:285, in Operation._get_node_attribute_at_index(self, node_index, attr, attr_name)
    269 """Private utility to retrieves an attribute (e.g. inputs) from a node.
    270 
    271 This is used to implement the properties:
   (...)
    282     The operation's attribute `attr` at the node of index `node_index`.
    283 """
    284 if not self._inbound_nodes:
--> 285     raise ValueError(
    286         f"The layer {self.name} has never been called "
    287         f"and thus has no defined {attr_name}."
    288     )
    289 if not len(self._inbound_nodes) > node_index:
    290     raise ValueError(
    291         f"Asked to get {attr_name} at node "
    292         f"{node_index}, but the operation has only "
    293         f"{len(self._inbound_nodes)} inbound nodes."
    294     )

ValueError: The layer sequential has never been called and thus has no defined output.
Click to add a cell.`

There is two workarounds where the value error is not raised: 1°) When using grad_model = keras.models.Model( [model.inputs], [model.get_layer(last_conv_layer_name).output, model.get_layer(Name_of_last_deep_layer).output]) but it results in none gradients in the rest of my code 2°) When redifining completely the model from scratch and loading only the weights, i.e., when using:

model = tf.keras.models.Sequential([
    tf.keras.Input(shape=(27, 75, 93, 81, 1)),  # time_steps, depth, height, width, channels
    tf.keras.layers.TimeDistributed(tf.keras.layers.Conv3D(6, kernel_size=7, activation='relu', kernel_initializer='he_normal')),
    tf.keras.layers.TimeDistributed(tf.keras.layers.MaxPooling3D(pool_size=(2, 2, 2))),
    tf.keras.layers.TimeDistributed(tf.keras.layers.BatchNormalization()),

    tf.keras.layers.TimeDistributed(tf.keras.layers.Conv3D(32, kernel_size=3, activation='relu', kernel_initializer='he_normal')),
    tf.keras.layers.TimeDistributed(tf.keras.layers.MaxPooling3D(pool_size=(2, 2, 2))),
    tf.keras.layers.TimeDistributed(tf.keras.layers.BatchNormalization()),

    tf.keras.layers.TimeDistributed(tf.keras.layers.Conv3D(128, kernel_size=2, activation='relu', kernel_initializer='he_normal')),
    tf.keras.layers.TimeDistributed(tf.keras.layers.MaxPooling3D(pool_size=(2, 2, 2))),
    tf.keras.layers.TimeDistributed(tf.keras.layers.BatchNormalization()),
    tf.keras.layers.TimeDistributed(tf.keras.layers.Conv3D(256, kernel_size=2, activation='relu', kernel_initializer='he_normal')),

    tf.keras.layers.TimeDistributed(tf.keras.layers.Flatten()),
    tf.keras.layers.TimeDistributed(tf.keras.layers.BatchNormalization()),

    tf.keras.layers.Conv1D(256, kernel_size=5, activation='relu', kernel_initializer='he_normal'),
    tf.keras.layers.MaxPooling1D(pool_size=2),
    tf.keras.layers.BatchNormalization(),

    tf.keras.layers.Conv1D(512, kernel_size=3, activation='relu', kernel_initializer='he_normal'),
    tf.keras.layers.MaxPooling1D(pool_size=2),
    tf.keras.layers.BatchNormalization(),

    tf.keras.layers.Conv1D(1024, kernel_size=2, activation='relu', kernel_initializer='he_normal'),
    tf.keras.layers.BatchNormalization(),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(2, activation='softmax')]) 

model.load_weights(...) -> this one doesn't raises any error

Thanks a lot!

[nkovela1]

Hi @Senantq, could you try calling model directly via model(...) rather than model.call? Additionally I see that model.build was called without an input shape. Could you also supply the original source code of the model rather than just the .keras file?

[Senantq]

Hi @nkovela1 and thank you for your help. I have updated the reprex, added the model code into it, and made it a bit more explicit. Do not hesitate to ask if anything's still lacking. Have a nice day,

[ghsanti]

@Senantq

Isn't this the same issue as fixed here?

https://github.com/keras-team/keras/issues/20155#issue-2482970454

You may try using keras-nightly

[Senantq]

I have just installed keras-nightly 3.5.0.dev2024090403 in a new conda environment, and it does'nt solve the problem unfortunately

[ghsanti]

model = keras.Sequential([
    Input(shape=(10, 10, 3)),
   Conv2D(filters=32, kernel_size=3),
])
model_2 = Model([model.inputs], [model.output]) # <--change

errors in a similar way.

output is an output tensor, not a layer. One can use outputs which returns a KerasTensor with a history (layer information.)

Specifically, model.outputs[-1] or the get layer equivalent.

[Senantq]

Hello ghsanti and thank you for answer. I must admit that it is still not totally clear to me. Particularly, I don't see why

from tensorflow import keras
model = keras.Sequential([
    keras.Input(shape=(10, 10, 3)),
   keras.layers.Conv2D(filters=32, kernel_size=3)])
model_2 = keras.Model([model.inputs], [model.output])

worked with keras 2 (installed via tensorflow==2.10) and not with the newer versions. Have I missed something during the change to keras 3? I don't see it listed as one of the new major releases features at https://github.com/keras-team/keras/issues/18467 Plus, it feels like it breaks a of not that old codes such as saliency methods or transfer learning codes.

[ghsanti]

@Senantq

• This works

  model = keras.Sequential([ Input(shape=(4, 4, 3)), Flatten(), Dense(units=5)])
  model(keras.Input((4, 4, 3)))
  model_2 = keras.Model([model.inputs], [model.output])

• This too:

i= Input(shape=(4, 4, 3))
x=  Flatten()(i)
x=  Dense(units=5)(x)
model1 = keras.Model(i, x)
model2 = keras.Model([model1.inputs], [model1.output])

Sequential needs the extra call to build the layers.

Other related threads:

* One needs to call the model with a Symbolic Tensor or an Input layer (which kind of defeats the purpose of Input in the first place? But it's how it's done since it needs an extra call.) Someone comments about [the same contradiction here](https://github.com/tensorflow/tensorflow/issues/44857#issuecomment-727095923) > (...) when I am initializing the densenet model, I am providing the input shape. > Then why do I need to provide it two times? * If you call the model it with a standard list/tensor etc it will not change anything. (this should be clear in the error message.) * I'd expect that the `call` is automatic if Input is present (a fix to the source code.) Here is a [comment suggesting something similar.](https://github.com/tensorflow/tensorflow/issues/44857#issuecomment-727073875)

[Senantq]

I agree that the error message should be at least a bit more explicit if possible. In any case, thank for the time spent.