The output of ops.prod doesn't work in layer's argument. To make it work, it's required to cast, i.e. int(ops.prod). Now, still its an issue to use ops.prod in call method anyway.

import keras
from keras import layers, ops
import numpy as np

class ProdDenseLayer(layers.Layer):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)

    def build(self, input_shape):
        self.input_prod = ops.prod(input_shape[1:])
        self.dense = layers.Dense(self.input_prod, activation='relu')
        self.dense.build(input_shape)

    def call(self, inputs):
        scale_factor = ops.prod(ops.shape(inputs)[1:])
        scaled_inputs = inputs * ops.cast(scale_factor, inputs.dtype)
        return self.dense(scaled_inputs)

# Main model using ops.prod
class ProdModel(keras.Model):
    def __init__(self, **kwargs):
        super().__init__(**kwargs)
        self.prod_layer1 = ProdDenseLayer()

    def build(self, input_shape):
        self.prod_layer1.build(input_shape)

    def call(self, inputs):
        batch_size = ops.shape(inputs)[0]
        total_elements = ops.prod(ops.shape(inputs)[1:])
        normalized_inputs = inputs / ops.cast(total_elements, inputs.dtype)
        x = self.prod_layer1(normalized_inputs)
        return x

# Create dummy data
batch_size = 32
input_shape = (10,)
X_train = np.random.randn(batch_size * 10, *input_shape).astype(np.float32)
y_train = np.random.randint(0, 2, (batch_size * 10, 1)).astype(np.float32)
model = ProdModel()
model.compile(
    optimizer='adam',
    loss='binary_crossentropy',
    metrics=['accuracy']
)
history = model.fit(
    X_train, y_train,
    epochs=3,
    batch_size=batch_size,
    validation_split=0.2,
    verbose=1
)
model.summary()

Epoch 1/3
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
/tmp/ipykernel_36/1432196851.py in <cell line: 0>()
     45     metrics=['accuracy']
     46 )
---> 47 history = model.fit(
     48     X_train, y_train,
     49     epochs=3,

/usr/local/lib/python3.11/dist-packages/keras/src/utils/traceback_utils.py in error_handler(*args, **kwargs)
    120             # To get the full stack trace, call:
    121             # `keras.config.disable_traceback_filtering()`
--> 122             raise e.with_traceback(filtered_tb) from None
    123         finally:
    124             del filtered_tb

/tmp/ipykernel_36/1432196851.py in build(self, input_shape)
     25 
     26     def build(self, input_shape):
---> 27         self.prod_layer1.build(input_shape)
     28 
     29     def call(self, inputs):

/tmp/ipykernel_36/1432196851.py in build(self, input_shape)
     11         self.input_prod = ops.prod(input_shape[1:])
     12         self.dense = layers.Dense(self.input_prod, activation='relu')
---> 13         self.dense.build(input_shape)
     14 
     15     def call(self, inputs):

ValueError: Invalid dtype: <property object at 0x79c67d66d3f0>

Comment From: SamanehSaadat

Hi @innat

I made two changes to your code that made it work: 1. As you mentioned, I cast the output of np.prod to int. 2. I added self.built=True to the custom layer and model and built the model. Here is the working colab.

I think it's reasonable that the output of op.prod is float32. Right?

Comment From: innat

First of all, setting built=1 is not needed and irrelevant here.

About casting ops.prod to int - in that case I can just use np.prod instead of ops.prod. The behaviour of them are not same.

The issue with ops.prod will also hit when you use it in call method with ops.reshape or layers.Reshape methods.

Comment From: SamanehSaadat

Could you explain why you need to use ops.prod here?

Comment From: innat

@SamanehSaadat To reproduce the error.

Comment From: SamanehSaadat

@SamanehSaadat To reproduce the error.

I mean what's your usecase that you need to use ops.prod and you can't use something like math.prod?

I'm trying to understand the problem better.

The output of ops.prod doesn't work in layer's argument.

Isn't this because its type is float? Are there other reasons?

The issue with ops.prod will also hit when you use it in call method with ops.reshape or layers.Reshape methods.

What's the issue you're seeing with the call method and why do you think the root cause of the issue is ops.prod?

Comment From: github-actions[bot]

This issue is stale because it has been open for 14 days with no activity. It will be closed if no further activity occurs. Thank you.

Comment From: georgeneedles60-hub

ok thanks for the update