Search code examples
pythonmachine-learningkeraskeras-layersoftmax

In Keras, how to apply softmax function on each row of the weight matrix?

from keras.models import Model
from keras.models import Input
from keras.layers import Dense

a = Input(shape=(3,))
b = Dense(2, use_bias=False)(a)
model = Model(inputs=a, outputs=b)

Suppose that the weights of the Dense layer in the above code is [[2, 3], [3, 1], [-1, 1]]. If we give [[2, 1, 3]] as an input to the model, then the output will be:

no softmax

But I want to apply the softmax function to each row of the Dense layer, so that the output will be:

with softmax

How can I do this?

Solution

  • One way to achieve what you are looking for is to define a custom layer by subclassing the Dense layer and overriding its call method:

    from keras import backend as K

class CustomDense(Dense):
    def __init__(self, units, **kwargs):
        super(CustomDense, self).__init__(units, **kwargs)

    def call(self, inputs):
        output = K.dot(inputs, K.softmax(self.kernel, axis=-1))
        if self.use_bias:
            output = K.bias_add(output, self.bias, data_format='channels_last')
        if self.activation is not None:
            output = self.activation(output)
        return output

    Test to make sure it works:

    model = Sequential()
model.add(CustomDense(2, use_bias=False, input_shape=(3,)))

model.compile(loss='mse', optimizer='adam')

import numpy as np

w = np.array([[2,3], [3,1], [1,-1]])
inp = np.array([[2,1,3]])

model.layers[0].set_weights([w])
print(model.predict(inp))

# output
[[4.0610714 1.9389288]]

    Verify it using numpy:

    soft_w = np.exp(w) / np.sum(np.exp(w), axis=-1, keepdims=True)
print(np.dot(inp, soft_w))

[[4.06107115 1.93892885]]