Bug in functional model

[Surya2k1]

I think there seems a bug in functional model.

Case-1: When inputs=outputs for the model construction but training different output shape: Training success

import keras
from keras import layers
import numpy as np

input_1 = layers.Input(shape=(3,))
input_2 = layers.Input(shape=(5,))

model_1 = keras.models.Model([input_1, input_2], [input_1, input_2])
print(model_1.summary())
model_1.compile(optimizer='adam',metrics=['accuracy','accuracy'],loss=['mse'])

#Notice I am passing different output size for training but still training happens
model_1.fit([np.random.normal(size=(10,3)),np.random.normal(size=(10,5))], 
                    [np.random.normal(size=(10,1)),np.random.normal(size=(10,2))])

print('Training completed')

Case 2: Same as Case-1 but different behavior with different mismatched output shapes(than case-1) for training: Error during loss calculation. But I expect Error during graph execution itself.

#With diffrent output shapes than model constructed its raising error while calculating the loss.
#Instead it should have raised shape mismatch error during graph execution.
model_1.fit([np.random.normal(size=(10,3)),np.random.normal(size=(10,5))], 
                    [np.random.normal(size=(10,2)),np.random.normal(size=(10,4))])

Case 3: With Unconnected inputs and outputs

input_1 = layers.Input(shape=(3,))
input_2 = layers.Input(shape=(5,))

input_3 = layers.Input(shape=(1,))
input_4 = layers.Input(shape=(2,))

model_2 = keras.models.Model([input_1, input_2], [input_3, input_4])
model_2.compile(optimizer='adam',metrics=['accuracy','accuracy'],loss=['mse'])

#Passing correct input and ouputs fails because these are not connected.
model_2.fit([np.random.normal(size=(10,3)),np.random.normal(size=(10,5))], [np.random.normal(size=(10,1)),np.random.normal(size=(10,2))])

Got error below which is correct but it is not useful for end users. Instead it should have raised error during graph construction.

177         output_tensors = []
    178         for x in self.outputs:
--> 179             output_tensors.append(tensor_dict[id(x)])
    180 
    181         return tree.pack_sequence_as(self._outputs_struct, output_tensors)

KeyError: "Exception encountered when calling Functional.call().\n\n\x1b[1m139941182292272\x1b[0m\n\nArguments received by Functional.call():\n  • inputs=('tf.Tensor(shape=(None, 3), dtype=float32)', 'tf.Tensor(shape=(None, 5), dtype=float32)')\n  • training=True\n  • mask=('None', 'None')"

I tried to fix an issue similar to case-3 by raising Error during graph build itself in PR #20705 where I noticed this issue related to case1(From failed Test case). Please refer the gist.

[harshaljanjani]

To my knowledge, Keras builds the computation graph in a JIT/lazy fashion, and I'm not too sure if addressing Case 2 is in the spirit of the paradigm. Shape validation typically happens when the model actually starts training, which is probably why the Case 2 error comes up during loss calculation and not graph construction. Having said that, I’d love to hear if there’s anything else to consider when trying to bypass it, especially since I’ve only recently started contributing to the library.