CODEWITHSUNDEEP
python-2.7tensorflow
Mehran
Mehran

Reputation: 60

Difficulty Understanding TensorFlow Computations

I'm new to TensorFlow and have difficulty understanding how the computations works. I could not find the answer to my question on the web.

For the following piece of code, the last time I print "d" in the for loop of the "train_neural_net()" function, I'm expecting the values to be identical to when I print "test_distance.eval". But they are way different. Can anyone tell me why this is happening? Isn't TensorFlow supposed to cache the Variable results learned in the for loop and use them when I run "test_distance.eval"?

def neural_network_model1(data):
    nn1_hidden_1_layer = {'weights': tf.Variable(tf.random_normal([5, n_nodes_hl1])), 'biasses': tf.Variable(tf.random_normal([n_nodes_hl1]))}
    nn1_hidden_2_layer = {'weights': tf.Variable(tf.random_normal([n_nodes_hl1, n_nodes_hl2])), 'biasses': tf.Variable(tf.random_normal([n_nodes_hl2]))}
    nn1_output_layer = {'weights': tf.Variable(tf.random_normal([n_nodes_hl2, vector_size])), 'biasses': tf.Variable(tf.random_normal([vector_size]))}

    nn1_l1 = tf.add(tf.matmul(data, nn1_hidden_1_layer["weights"]), nn1_hidden_1_layer["biasses"])
    nn1_l1 = tf.sigmoid(nn1_l1)

    nn1_l2 = tf.add(tf.matmul(nn1_l1, nn1_hidden_2_layer["weights"]), nn1_hidden_2_layer["biasses"])
    nn1_l2 = tf.sigmoid(nn1_l2)

    nn1_output = tf.add(tf.matmul(nn1_l2, nn1_output_layer["weights"]), nn1_output_layer["biasses"])

    return nn1_output

def neural_network_model2(data):
    nn2_hidden_1_layer = {'weights': tf.Variable(tf.random_normal([5, n_nodes_hl1])), 'biasses': tf.Variable(tf.random_normal([n_nodes_hl1]))}
    nn2_hidden_2_layer = {'weights': tf.Variable(tf.random_normal([n_nodes_hl1, n_nodes_hl2])), 'biasses': tf.Variable(tf.random_normal([n_nodes_hl2]))}
    nn2_output_layer = {'weights': tf.Variable(tf.random_normal([n_nodes_hl2, vector_size])), 'biasses': tf.Variable(tf.random_normal([vector_size]))}

    nn2_l1 = tf.add(tf.matmul(data, nn2_hidden_1_layer["weights"]), nn2_hidden_1_layer["biasses"])
    nn2_l1 = tf.sigmoid(nn2_l1)

    nn2_l2 = tf.add(tf.matmul(nn2_l1, nn2_hidden_2_layer["weights"]), nn2_hidden_2_layer["biasses"])
    nn2_l2 = tf.sigmoid(nn2_l2)

    nn2_output = tf.add(tf.matmul(nn2_l2, nn2_output_layer["weights"]), nn2_output_layer["biasses"])

    return nn2_output

def train_neural_net():
    prediction1 = neural_network_model1(x1)
    prediction2 = neural_network_model2(x2)

    distance = tf.sqrt(tf.reduce_sum(tf.square(tf.subtract(prediction1, prediction2)), reduction_indices=1))
    cost = tf.reduce_mean(tf.multiply(y, distance))
    optimizer = tf.train.AdamOptimizer().minimize(cost)

    hm_epochs = 500

    test_result1 = neural_network_model1(x3)
    test_result2 = neural_network_model2(x4)
    test_distance = tf.sqrt(tf.reduce_sum(tf.square(tf.subtract(test_result1, test_result2)), reduction_indices=1))

    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())

        for epoch in range(hm_epochs):
            _, d = sess.run([optimizer, distance], feed_dict = {x1: train_x1, x2: train_x2, y: train_y})
            print("Epoch", epoch, "distance", d)

        print("test distance", test_distance.eval({x3: train_x1, x4: train_x2}))

train_neural_net()

Upvotes: 0

Views: 96

Answers (2)

Alldb
Alldb

Reputation: 130

You don't need to define two function for generating models, you can use tf.name_scope, and pass a model name to the function to use it as a prefix for variable declaration. On the other hand, you defined two variables for distance, first is distance and second is test_distance . But your model will learn from train data to minimize cost which is only related to first distance variable. Therefore, test_distance is never used and the model which is related to it, will never learn anything! Again there is no need for two distance functions. You only need one. When you want to calculate train distance, you should feed it with train data and when you want to calculate test distance you should feed it with test data. Anyway, if you want second distance to work, you should declare another optimizer for it and also you have to learn it as you have done for first one. Also you should consider the fact that models are learning base on their initial values and training data. Even if you feed both models with exactly same training batches, you can't expect to have exactly similar characteristics models since initial values for weights are different and this could cause falling into different local minimum of error surface. At the end notice that whenever you call neural_network_model1 or neural_network_model2 you will generate new weights and biases, because tf.Variable is generating new variables for you.

Upvotes: 1

mrry
mrry

Reputation: 126154

Each time you call the functions neural_network_model1() or neural_network_model2(), you create a new set of variables, so there are four sets of variables in total.

  • The call to sess.run(tf.global_variables_initializer()) initializes all four sets of variables.

  • When you train in the for loop, you only update the first two sets of variables, created with these lines:

    prediction1 = neural_network_model1(x1)
prediction2 = neural_network_model2(x2)

  • When you evaluate with test_distance.eval(), the tensor test_distance depends only on the variables that were created in the last two sets of variables, which were created with these lines:

    test_result1 = neural_network_model1(x3)
test_result2 = neural_network_model2(x4)

    These variables were never updated in the training loop, so the evaluation results will be based on the random initial values.

TensorFlow does include some code for sharing weights between multiple calls to the same function, using with tf.variable_scope(...): blocks. For more information on how to use these, see the tutorial on variables and sharing on the TensorFlow website.

Upvotes: 1

Related Questions