I have a function that returns a Keras model. However, as an exercise, I’m trying to convert it to a keras.Model class, but I’m having difficulty.

This is the function:

def siamese_loss_network():
    inputs = keras.layers.Input((128, 128, 3))
    x = keras.applications.efficientnet.preprocess_input(inputs)
    base = keras.applications.EfficientNetB0(include_top=False, input_tensor=inputs, pooling = 'max')
    head = base.output
    x = keras.layers.Dense(256, activation="relu")(head)
    x = keras.layers.Dense(32)(x)
    embedding_network = keras.Model(inputs, x)
    input_1 = keras.layers.Input((128, 128, 3),name="input_layer_base_r")
    input_2 = keras.layers.Input((128, 128, 3),name="input_layer_base_l")
    tower_1 = embedding_network(input_1)
    tower_2 = embedding_network(input_2)
    merge_layer = keras.layers.Lambda(euclidean_distance, output_shape=(1,))(
        [tower_1, tower_2]
    )
    output_layer = keras.layers.Dense(1, activation="sigmoid")(merge_layer)
    siamese = keras.Model(inputs=[input_1, input_2], outputs=output_layer)
    return siamese

def euclidean_distance(vects):
    x, y = vects
    sum_square = ops.sum(ops.square(x - y), axis=1, keepdims=True)
    return ops.sqrt(ops.maximum(sum_square, keras.backend.epsilon()))

When running:

model = siamese_loss_network()
model.compile(optimizer=Adam(), loss=loss())
model.summary()

I get the following output:

Model: "functional_1"
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Layer (type)                  ┃ Output Shape              ┃         Param # ┃ Connected to               ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│ input_layer_base_r            │ (None, 128, 128, 3)       │               0 │ -                          │
│ (InputLayer)                  │                           │                 │                            │
├───────────────────────────────┼───────────────────────────┼─────────────────┼────────────────────────────┤
│ input_layer_base_l            │ (None, 128, 128, 3)       │               0 │ -                          │
│ (InputLayer)                  │                           │                 │                            │
├───────────────────────────────┼───────────────────────────┼─────────────────┼────────────────────────────┤
│ functional (Functional)       │ (None, 32)                │       4,385,731 │ input_layer_base_r[0][0],  │
│                               │                           │                 │ input_layer_base_l[0][0]   │
├───────────────────────────────┼───────────────────────────┼─────────────────┼────────────────────────────┤
│ lambda (Lambda)               │ (None, 1)                 │               0 │ functional[0][0],          │
│                               │                           │                 │ functional[1][0]           │
├───────────────────────────────┼───────────────────────────┼─────────────────┼────────────────────────────┤
│ dense_2 (Dense)               │ (None, 1)                 │               2 │ lambda[0][0]               │
└───────────────────────────────┴───────────────────────────┴─────────────────┴────────────────────────────┘
 Total params: 4,385,733 (16.73 MB)
 Trainable params: 4,343,710 (16.57 MB)
 Non-trainable params: 42,023 (164.16 KB)

So here is my adaptation for a class that inherits keras.Model:

class SiameseModel(keras.Model):
    def __init__(self): 
        super().__init__()
        self.inputs = keras.layers.Input((128, 128, 3))
        self.input_1 = keras.layers.Input((128, 128, 3),name="input_layer_base_r")
        self.input_2 = keras.layers.Input((128, 128, 3),name="input_layer_base_l")
        self.base = keras.applications.EfficientNetB0(include_top=False, input_tensor=self.inputs, pooling = 'max')
        self.dense_1 = keras.layers.Dense(256, activation="relu")
        self.dense_2 = keras.layers.Dense(32)
        self.merge_layer = keras.layers.Lambda(euclidean_distance, output_shape=(1,))
        self.output_layer = keras.layers.Dense(1, activation="sigmoid")

    def call(self, inputs):
        head = self.base.output
        x = self.dense_1(head)
        x = self.dense_2(x)
        embedding_network = keras.Model(inputs, x)
        tower_1 = embedding_network(self.input_1)
        tower_2 = embedding_network(self.input_2)
        merge = self.merge_layer([tower_1, tower_2])
        output = self.output_layer(merge)
        return keras.Model(inputs=[self.input_1, self.input_2], outputs=output)

When running:

model = SiameseModel()
model.compile(optimizer=Adam(), loss=loss())
model.summary()

i got the error:

Traceback (most recent call last):
  File "D:[project path]main.py", line 20, in <module>
    model.summary()
  File "C:[user path].condaenvs[env path]libsite-packageskerassrcutilstraceback_utils.py", line 122, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "C:[user path].condaenvs[env path]libsite-packagesoptreeops.py", line 594, in tree_map
    return treespec.unflatten(map(func, *flat_args))
ValueError: Undefined shapes are not supported.

If I try to run mode.fit(), the model trains without any problems, but it’s strange not being able to run model.summary() first. I feel like something is wrong.

I read this issue on Keras repo, but i honestly didn’t understand the reason for the error, nor how to resolve it. Could anyone enlighten me about this?

Python version: Python 3.10.13
pip version: 24.0
Tensorflow version: 2.16.1
Keras version: Version: 3.4.1

Grateful for the attention!