CODEWITHSUNDEEP
tensorflow
Simon
Simon

Reputation: 5412

Variable substitution without duplicating the tensor (or having the graph accepting two different input)

I think it is easier to clarify what I need with a MWE (question is in the comment).

import tensorflow as tf
import numpy as np


class MLP:
    def __init__(self, sizes, activations):
        self.input = last_out = tf.placeholder(dtype=tf.float32, shape=[None, sizes[0]])
        self.layers = []
        for l, size in enumerate(sizes[1:]):
            self.layers.append(last_out)
            last_out = tf.layers.dense(last_out, size, activation=activations[l], kernel_initializer=tf.glorot_uniform_initializer())
        self.layers.append(last_out)



def main():
    session = tf.Session()

    dim = 3
    nn_sizes = [dim, 15, 1]
    nn_activations = [tf.nn.tanh, tf.nn.tanh, tf.identity]
    mynet = MLP(nn_sizes, nn_activations)

    w = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='mynet')
    x1 = tf.placeholder(dtype=tf.float32, shape=[None, dim], name='x1')
    x2 = tf.placeholder(dtype=tf.float32, shape=[None, dim], name='x2')
    x3 = tf.placeholder(dtype=tf.float32, shape=[None, 1], name='x3')

    myfun = tf.reduce_sum(tf.multiply(x3, new_tensor)
    # new_tensor has to be the difference myfun(x2)-myfun(x3).
    # However, the network is the same and its input variable has a different name.
    # I would like to have something like:
    #   substitute(myfun,input,x1)
    #   substitute(myfun,input,x2)
    # without duplicating the network.

    optimizer = tf.contrib.opt.ScipyOptimizerInterface(myfun,var_list=w)

    n = 1000
    x1_samples = np.asmatrix(np.random.rand(n,dim))
    x2_samples = np.asmatrix(np.random.rand(n,dim))
    x3_samples = np.asmatrix(np.random.rand(n,1))

    print(session.run(myfun, {x1: x1_samples, x2: x2_samples, x3: x3_samples}))
    optimizer.minimize(session, {x1: x1_samples, x2: x2_samples, x3: x3_samples})
    print(session.run(myfun, {x1: x1_samples, x2: x2_samples, x3: x3_samples}))


if __name__ == '__main__':
    main()

Upvotes: 0

Views: 50

Answers (1)

javidcf
javidcf

Reputation: 59701

Here's one approach (I assume there is a typo and what you want is x3 * (mynet(x2) - mynet(x1))?):

import tensorflow as tf
import numpy as np

class MLP:
    def __init__(self, x1, x2, sizes, activations):
        x_sizes = [tf.shape(x1)[0], tf.shape(x2)[0]]
        last_out = tf.concat([x1, x2], axis=0)
        self.layers = []
        for l, size in enumerate(sizes[1:]):
            self.layers.append(last_out)
            last_out = tf.layers.dense(last_out, size, activation=activations[l], kernel_initializer=tf.glorot_uniform_initializer())
        self.layers.append(last_out)
        self.x1_eval, self.x2_eval = tf.split(last_out, x_sizes, axis=0)



def main():
    session = tf.Session()

    dim = 3
    nn_sizes = [dim, 15, 1]
    nn_activations = [tf.nn.tanh, tf.nn.tanh, tf.identity]

    w = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='mynet')
    x1 = tf.placeholder(dtype=tf.float32, shape=[None, dim], name='x1')
    x2 = tf.placeholder(dtype=tf.float32, shape=[None, dim], name='x2')
    x3 = tf.placeholder(dtype=tf.float32, shape=[None, 1], name='x3')

    mynet = MLP(x1, x2, nn_sizes, nn_activations)

    myfun = tf.reduce_sum(tf.multiply(x3, (mynet.x2_eval - mynet.x1_eval)))

    optimizer = tf.contrib.opt.ScipyOptimizerInterface(myfun,var_list=w)

    n = 1000
    x1_samples = np.asmatrix(np.random.rand(n,dim))
    x2_samples = np.asmatrix(np.random.rand(n,dim))
    x3_samples = np.asmatrix(np.random.rand(n,1))

    session.run(tf.global_variables_initializer())
    print(session.run(myfun, {x1: x1_samples, x2: x2_samples, x3: x3_samples}))
    optimizer.minimize(session, {x1: x1_samples, x2: x2_samples, x3: x3_samples})
    print(session.run(myfun, {x1: x1_samples, x2: x2_samples, x3: x3_samples}))


if __name__ == '__main__':
    main()

Upvotes: 1

Related Questions