test.py

tf.reset_default_graph()
input_image = tf.placeholder(tf.float32, shape=[X_train_prep.shape[0], X_train_prep.shape[1]])
U = tf.Variable(tf.truncated_normal([X_train_prep.shape[0], hidden_dim], stddev=0.01), name="U")
V = tf.Variable(tf.truncated_normal([hidden_dim, initial_dim], stddev=0.01), name="V")
approximation_loss = tf.sqrt(tf.reduce_sum(tf.square(tf.sub(input_image, tf.matmul(U, V)))))

# Define optimizer and other constants

init = tf.initialize_all_variables()
sess = tf.InteractiveSession()
sess.run(init)
for num in xrange(num_switches):
    if num % 2 == 0:
        print "Optimization by U", num
        grads_and_vars = optimizer.compute_gradients(approximation_loss, var_list=[V])
    else:
        print "Optimization by V", num
        grads_and_vars = optimizer.compute_gradients(approximation_loss, var_list=[U])
    operation = optimizer.apply_gradients(grads_and_vars)
    for iteration in xrange(num_epoch):
        sess.run(operation, feed_dict={input_image: X_train_prep})