CODEWITHSUNDEEP
tensorflowkeras
josh
josh

Reputation: 126

How to sum up the prediction vectors of a keras model into a single vector

I have 2 keras models. The first gets as input a string and gives a prediction for example, five classes.

In the second model I want to use this output. However, the output of the first model should be summed up into a single output for multiple inputs.

I want single prediction for the sum of all entered strings and not a prediction for each entered string.

model1 = tf.keras.Sequential()
model1.add(Input(shape=(len(inputs[0]),), dtype=tf.float32))
model1.add(Dense(256, activation='relu'))
model1.add(Dense(len(helper_classes), activation='softmax'))

model2 = tf.keras.Sequential()
model2.add(model1)
model2.add(Dense(16))
model2.add(Dense(len(classes), activation=tf.nn.softmax))
model2.layers[0].trainable = False
model2.compile(loss='categorical_crossentropy',optimizer='adam', metrics=['accuracy'])

model2.summary()

For explanation: the strings are preprocessed to a float vector.

Actual output of model1:
Input: "Hello","World", ...
Output: [0.2, 0, 0, 0.8, 0],[0, 0, 0.4, 0, 0.6], ...

What i need:
Input: "Hello","World", ...
Output: [0.2 + 0.0 + ... , 0 + 0.0 + ... , 0 + 0.4 + ... , 0.8 + 0.0 + ... , 0 + 0.6 + ...]

Image of model1
Image of model1 after adding Reduction Layer


Solution
Okay I solved it now. My first mistake was that I summed up on axis 1. What I could fix with the help of vlad. The second mistake was that I did not keep the dimensions with keep_dims = true.

The solution was to insert a lambda layer in the second model which basically does what Vlad and Thibault proposed:

model2 = tf.keras.Sequential()
model2.add(model1)

model2.add(Lambda(lambda x: K.sum(x, axis=0,keepdims=True)))

model2.add(Dense(16))
model2.add(Dense(len(classes), activation=tf.nn.softmax))
model2.layers[0].trainable = False
model2.compile(loss='categorical_crossentropy',optimizer='adam', metrics=['accuracy'])

Upvotes: 0

Views: 3009

Answers (2)

Vlad
Vlad

Reputation: 8585

Use tf.reduce_sum():

import tensorflow as tf
output = tf.Variable([[0.2, 0.0, 0.0, 0.8, 0],[0.0, 0.0, 0.4, 0, 0.6],])
reduced = tf.reduce_sum(output, axis=0)

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    print(reduced.eval())
    # [0.2 0.  0.4 0.8 0.6]

To use it within Keras define a custom layer like this:

from tensorflow.keras import layers

class ReductionLayer(layers.Layer):
    def __init__(self):
        super(ReductionLayer, self).__init__()

    def call(self, inputs):
        return tf.reduce_sum(inputs, axis=0)

and add it to your Sequential() model:

model.add(ReductionLayer())

Upvotes: 2

Thibault Bacqueyrisses
Thibault Bacqueyrisses

Reputation: 2331

If i understand well your problem, all you need is to sum your last Dense layer of model1. You can achieve that by unsing Keras backend Sum : 

keras.backend.sum(x, axis=None, keepdims=False)

You can found doc here : https://keras.io/backend/#sum

Upvotes: 1

Related Questions