Keras: How to Multiply()?

Question

TensorFlow 2.0 RC1

import tensorflow as tf
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Multiply

import numpy as np

Expected output:

Multiply()([np.array([1,2,3,4,4,4]).reshape(2,3), np.array([1,0])])

Problem:

input_1 = Input(shape=(None,3))
mask_1 = Input(shape=(None,))

net = Multiply()([input_1, mask_1])
net = Model(inputs=[input_1, mask_1], outputs=net)

net.predict([np.array([1,2,3,4,4,4]).reshape(1,2,3), np.array([1,0]).reshape(1,2)]) # 1 = batch size

How to fix this problem?

score 3 · Answer 1 · answered Sep 30 '19 at 11:30

3

Reshape the second array in the last line of code as np.array([1,0]).reshape(-1)

net.predict([np.array([1,2,3,4,4,4]).reshape(1,2,3), np.array([1,0]).reshape(-1)]) # 1 = batch size

answered Sep 30 '19 at 11:30

stephen_mugisha

889
1
8
18

2

For further information about `.reshape(-1)` you can check the [link](https://stackoverflow.com/questions/18691084/what-does-1-mean-in-numpy-reshape) – Physicing Sep 30 '19 at 11:38
Thanks @Physicing for the add-on – stephen_mugisha Sep 30 '19 at 11:39
@stephen_mugisha Thank you for you answer! But this does not work. – Alexey Golyshev Sep 30 '19 at 11:54
@AlexeyGolyshev what error are you getting when you use this method? – stephen_mugisha Sep 30 '19 at 11:56
`ValueError: Mismatch between expected batch size and model output batch size. Output shape = (2, 2, 3), expected output shape = shape (1, 2, 3)` – Alexey Golyshev Sep 30 '19 at 11:57
@AlexeyGolyshev do you mind showing you input dimensions?? Your output is 2-dimensional and you need to reduce it to 1-dimensional – stephen_mugisha Sep 30 '19 at 12:08

Manoj Mohan · Answer 2 · 2019-09-30T12:08:40.677

3

It depends on how the input shape is specified. In the Multiply() example(element-wise multiplication), the batch size is 2 and the feature size is 3 for Input and 1 for mask. So, when specifying the input shape in Keras, only the feature size needs to be specified.

input_1 = Input(shape=(3,))
mask_1 = Input(shape=(1,))
net = Multiply()([input_1, mask_1])
net = Model(inputs=[input_1, mask_1], outputs=net)
output = net.predict([np.array([1,2,3,4,4,4]).reshape(2,3), np.array([1,0])])
print(output)

[[1. 2. 3.] [0. 0. 0.]]

edited Sep 30 '19 at 12:08

answered Sep 30 '19 at 12:01

Manoj Mohan

5,654
1
17
21

Thank you for your answer! I do character embedding. So in my example batch size is 1. 2 words of 3 characters. – Alexey Golyshev Sep 30 '19 at 12:13

score 2 · Accepted Answer · answered Sep 30 '19 at 12:03

Number of dimensions should match, by modifying the input shape of the 2nd input to (None, 1) and adding an extra-dimension to the [1, 0] array

import numpy as np
from tensorflow.keras.layers import Multiply
from tensorflow.keras import Model, Input

input_1 = Input(shape=(2,3))
mask_1 = Input(shape=(2,1))

net = Multiply()([input_1, mask_1])
net = Model(inputs=[input_1, mask_1], outputs=net)

net.summary()

print(net.predict([np.array([1,2,3,4,4,4]).reshape((1,2,3)), np.array([1,0]).reshape((1,2,1))]))

Keras: How to Multiply()?

3 Answers3