I've been combing through this code for a week now trying to figure out why my cost function is increasing as in the following image. Reducing the learning rate does help but very little. Can anyone spot why the cost function isn't working as expected?

I realise a CNN would be preferable, but I still want to understand why this simple network is failing. Please help:)

Runaway Cost Function

import numpy as np
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data

import matplotlib.pyplot as plt

mnist = input_data.read_data_sets("MNIST_DATA/",one_hot=True)

def createPlaceholders():
    xph = tf.placeholder(tf.float32, (784, None))
    yph = tf.placeholder(tf.float32, (10, None))
    return xph, yph

def init_param(layers_dim):
    weights = {}
    L = len(layers_dim)

    for l in range(1,L):
        weights['W' + str(l)] = tf.get_variable('W' + str(l), shape=(layers_dim[l],layers_dim[l-1]), initializer= tf.contrib.layers.xavier_initializer())
        weights['b' + str(l)] = tf.get_variable('b' + str(l), shape=(layers_dim[l],1), initializer= tf.zeros_initializer())

    return weights

def forward_prop(X,L,weights):
    parameters = {}
    parameters['A0'] = tf.cast(X,tf.float32)

    for l in range(1,L-1):
        parameters['Z' + str(l)] = tf.add(tf.matmul(weights['W' + str(l)], parameters['A' + str(l-1)]), weights['b' + str(l)])
        parameters['A' + str(l)] = tf.nn.relu(parameters['Z' + str(l)])

    parameters['Z' + str(L-1)] = tf.add(tf.matmul(weights['W' + str(L-1)], parameters['A' + str(L-2)]), weights['b' + str(L-1)])
    return parameters['Z' + str(L-1)]

def compute_cost(ZL,Y):
    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels = tf.cast(Y,tf.float32), logits = ZL))
    return cost

def randomMiniBatches(X,Y,minibatch_size):
    m = X.shape[1]
    shuffle = np.random.permutation(m)
    temp_X = X[:,shuffle]
    temp_Y = Y[:,shuffle]

    num_complete_minibatches = int(np.floor(m/minibatch_size))

    mini_batches = []

    for batch in range(num_complete_minibatches):
        mini_batches.append((temp_X[:,batch*minibatch_size: (batch+1)*minibatch_size], temp_Y[:,batch*minibatch_size: (batch+1)*minibatch_size]))

    mini_batches.append((temp_X[:,num_complete_minibatches*minibatch_size:], temp_Y[:,num_complete_minibatches*minibatch_size:]))

    return mini_batches

def model(X, Y, layers_dim, learning_rate = 0.001, num_epochs = 20, minibatch_size = 64):
    costs = []

    xph, yph = createPlaceholders()
    weights = init_param(layers_dim)
    ZL = forward_prop(xph, len(layers_dim), weights)
    cost = compute_cost(ZL,yph)
    optimiser = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)

    init = tf.global_variables_initializer()

    with tf.Session() as sess:

        for epoch in range(num_epochs):
            minibatches = randomMiniBatches(X,Y,minibatch_size)
            epoch_cost = 0

            for b, mini in enumerate(minibatches,1):
                mini_x, mini_y = mini
                _,c = sess.run([optimiser,cost],feed_dict={xph:mini_x,yph:mini_y})
                epoch_cost += c
            print('epoch: ',epoch+1,'/ ',num_epochs)

            epoch_cost /= len(minibatches)


X_train = mnist.train.images.T
n_x = X_train.shape[0]
Y_train = mnist.train.labels.T
n_y = Y_train.shape[0]
layers_dim = [n_x,10,n_y]

model(X_train, Y_train, layers_dim)
  • 1
    $\begingroup$ What's the x-axis? What's the y-axis? $\endgroup$
    – SmallChess
    Commented Apr 23, 2018 at 14:05
  • $\begingroup$ on the graph? y is cost and x is epoch_number. The cost is increasing like crazy! $\endgroup$ Commented Apr 23, 2018 at 14:08
  • $\begingroup$ I think this question is off-topic as it's about debugging a code that doesn't work. $\endgroup$
    – SmallChess
    Commented Apr 23, 2018 at 14:09
  • $\begingroup$ happy to hear suggestions of what to try without you looking at the code. I've duplicated a model that worked for another basic classification with a monotonically decreasing cost, but for some reason with MNIST my cost is increasing. $\endgroup$ Commented Apr 23, 2018 at 14:15

2 Answers 2


Tensorflow's softmax function only works if the number of batches are in the rows and the output in the columns. If these are reversed, then you need to transpose the tensors in the cost function.


Can you try the same after normalizing the data, sometimes most of us in the beginning does not provide enough attention to that specific area. One of the reasons that this could happen is that on the train set you might have normalized but passed the raw data as labels when fitting to the network.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.