Implementing a tensorflow graph into a Keras model

Question

I am trying to implement roughly the following architecture in Keras (preferably) or Tensorflow.

          ___________      _________      _________     ________    ______
          | Conv    |     | Max    |     | Dense  |    |       |   |     |
Input0--> | Layer 1 | --> | Pool 1 | --> | Layer  | -->|       |   |     |
          |_________|     |________|     |________|    | Sum   |   | Out |
                                                       | Layer |-->|_____|
Input1    ----------- Converted to trainable weights-->|       |              
                                                       |_______|                                                                               |_______|

In short, it is pretty much a model with two inputs, merged into one output using an Add([input0, input1]) layer. The trick is that one of the inputs must be seen as a variable = trainable weight.

Keras layer Add() does not allow this, and it takes input0 and input1 as non-trainable variables:

input0    = Input((28,28,1))
x         = Conv2D(32, kernel_size=(3, 3), activation='relu',input_shape=input_shape)(mod1)
x         = Conv2D(64, (3, 3), activation='relu')(input0)
x         = MaxPooling2D(pool_size=(2, 2))(x)
x         = Flatten()(x)
x         = Dense(128, activation='relu')(x)

input1    = Input((128,))

x         = Add()([x, input1])
x         = Dense(num_classes, activation='softmax')(x)
model     = Model(inputs = [mod1,TPM], outputs = x)
model.summary()
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

I can implement a graph in tensorflow that adds a placeholder X with a weight b, and learns the value for b in respect to a target Y.

train_X = numpy.asarray([1.0, 2.0])
train_Y = numpy.asarray([0.0, 2.5])
n_samples = train_X.shape[0]

# tf Graph Input
X = tf.placeholder("float")
Y = tf.placeholder("float")

# Set model weights
b = tf.Variable([0.0, 0.0], name="bias")

# Construct a linear model
pred = tf.add(X, b)

loss = tf.reduce_mean(tf.square(pred - train_Y))
optimizer = tf.train.GradientDescentOptimizer(learning_rate)
grads_and_vars = optimizer.compute_gradients(loss)

train = optimizer.apply_gradients(grads_and_vars)
#init = tf.initialize_all_variables()
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)

for step in range(epochs):
    sess.run(train, feed_dict={X: train_X, Y: train_Y})

Ths works exaclty how I want. Simple optimizable addition of an input and weights. But I can't include this into a Keras model.I am missing the step how to merge both ideas.

How can I include a layer that only sums one trainable tensor to a non-trainable tensor?

Answer 1

I'm not sure if I fully understand your needs. Based on your tensorflow code, I don't think you will have to feed in the initial value. In that case, I hope the following is at least close to what you want:

import numpy as np
import keras
from keras import backend as K
from keras.engine.topology import Layer
from keras.models import Model
from keras.layers import Input, Conv2D, MaxPooling2D, Flatten, Dense, Add

class MyLayer(Layer):

    def __init__(self, bias_init, **kwargs):
        self.bias_init = bias_init
        super(MyLayer, self).__init__(**kwargs)

    def build(self, input_shape):
        self.bias = self.add_weight(name='bias',
                                    shape=input_shape[1:],
                                    initializer=keras.initializers.Constant(self.bias_init),
                                    trainable=True)
        super(MyLayer, self).build(input_shape)  # Be sure to call this somewhere!

    def call(self, x):
        return x + self.bias

input0    = Input((28,28,1))
x         = Conv2D(32, kernel_size=(3, 3), activation='relu',input_shape=(28,28,1))(input0)
x         = Conv2D(64, (3, 3), activation='relu')(input0)
x         = MaxPooling2D(pool_size=(2, 2))(x)
x         = Flatten()(x)
x         = Dense(128, activation='relu')(x)

input1    = np.random.rand(128)

x         = MyLayer(input1)(x)
x         = Dense(10, activation='softmax')(x)
model     = Model(inputs=input0, outputs=x)
model.summary()
model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])