Image classifier using cifar 100, train accuracy not increasing

Question

I was trying to train an image classifier model using cifar100 dataset in tensorflow, but accuracy is not increasing over 1.2%. I googled the issue and found several solutions but still my model is not doing well.

I implemented a few steps such as:

1. increasing CNN layer and pooling along with drop outs and normalization
2. changing no. of dense layers
3. changing batch size and epochs
4. changing optimizers

A common thing I noticed is that with epoch=10 and batch size=256 & epoch=500 and batch size=512 training loss and accuracy is varying in the same way.

To prevent over-fitting I also tried dropout regularisation, this shows some change (train acc. varies in between 0.5 and 1.2%), with the same parameters when I increased epochs nothing changed(train and model acc.)..

I wanted to know whether this is a problem with the dataset or with the model definition.

classifier model:

def classifierModel(inp):
    layer1=tf.nn.relu(tf.nn.conv2d(inp, filter=tf.Variable(tf.truncated_normal([5,5,3,16])), 
                                   strides=[1,2,2,1], padding='SAME'))
    layer1=tf.nn.bias_add(layer1, tf.Variable(tf.truncated_normal([16])))
    layer1=tf.nn.relu(tf.nn.max_pool(layer1, ksize=[1,1,1,1], strides=[1,2,2,1], padding='SAME'))

    layer2=tf.nn.relu(tf.nn.conv2d(layer1, filter=tf.Variable(tf.truncated_normal([5,5,16,32])), 
                                   strides=[1,2,2,1], padding='SAME'))
    layer2=tf.nn.bias_add(layer2, tf.Variable(tf.truncated_normal([32])))
    layer2=tf.nn.relu(tf.nn.max_pool(layer2, ksize=[1,1,1,1], strides=[1,2,2,1], padding='SAME'))

    layer3=tf.nn.relu(tf.nn.conv2d(layer2, filter=tf.Variable(tf.truncated_normal([5,5,32, 64])), 
                                   strides=[1,2,2,1], padding='SAME'))
    layer3=tf.nn.bias_add(layer3, tf.Variable(tf.truncated_normal([64])))

    layer3=tf.nn.relu(tf.nn.max_pool(layer3, ksize=[1,1,1,1], strides=[1,2,2,1], padding='SAME'))
    layer3=tf.nn.dropout(layer3, keep_prob=0.7)
    print(layer3.shape)


    fclayer1=tf.reshape(layer3, [-1, weights['fc1'].get_shape().as_list()[0]])
    fclayer1=tf.add(tf.matmul(fclayer1, weights['fc1']), biases['fc1'])
    fclayer1= tf.nn.dropout(fclayer1, keep_prob=0.5)
    fclayer2=tf.add(tf.matmul(fclayer1, weights['fc2']), biases['fc2'])
    fclayer2=tf.nn.dropout(fclayer2, keep_prob=0.5)
    fclayer3=tf.add(tf.matmul(fclayer2, weights['fc3']), biases['fc3'])
    fclayer3=tf.nn.dropout(fclayer3, keep_prob=0.7)
    outLayer=tf.nn.softmax(tf.add(tf.matmul(fclayer3, weights['out']), biases['out']))
    return outLayer

Optimizers, cost, accuracy:

cost=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=model, labels=y))
optimizer=tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)
correct_pred=tf.equal(tf.argmax(model, 1), tf.argmax(y, 1))
accuracy=tf.reduce_mean(tf.cast(correct_pred, tf.float32))

Training:

with tf.Session() as sess:
sess.run(init)
for i in range(epochs):
    #shuffle(idx)
    #train_features=train_features[idx, :, :, :]
    #train_labels=train_labels[idx, ]
    for batch_features, batch_labels in get_batches(batch_size, train_features, train_labels):
        sess.run(optimizer, feed_dict={x:batch_features, y:batch_labels})
    if (i%display_step==0):

        epoch_stats(sess, i, batch_features, batch_labels)

model_acc=sess.run(accuracy, feed_dict={x:test_features, y:test_labels})
saver.save(sess, save_file)

writer.add_graph(sess.graph)

Results:

1. epoch : 0 - cost : 4.62 - acc: 0.01
2. epoch : 1 - cost : 4.62 - acc: 0.01
3. epoch : 2 - cost : 4.62 - acc: 0.008
4. epoch : 3 - cost : 4.61 - acc: 0.012
5. epoch : 4 - cost : 4.61 - acc: 0.005
6. epoch : 5 - cost : 4.62 - acc: 0.006
7. epoch : 6 - cost : 4.62 - acc: 0.016
8. epoch : 7 - cost : 4.62 - acc: 0.012
9. epoch : 8 - cost : 4.61 - acc: 0.014
10. epoch : 9 - cost : 4.62 - acc: 0.009
11. Model accuracy - 0.010499999858438969

Answer 1

The first argument you are passing to softmax_cross_entropy_with_logits_v2 is incorrect. You must pass the "previous" values to apply the softmax. That's because softmax_cross_entropy_with_logits_v2 is really cross_entropy (softmax (x)). The justification is that the derivative can be simplified.

In model you should do the following:

def classifierModel(inp):
    layer1=tf.nn.relu(tf.nn.conv2d(inp, filter=tf.Variable(tf.truncated_normal([5,5,3,16])), 
                                   strides=[1,2,2,1], padding='SAME'))
    layer1=tf.nn.bias_add(layer1, tf.Variable(tf.truncated_normal([16])))
    layer1=tf.nn.relu(tf.nn.max_pool(layer1, ksize=[1,1,1,1], strides=[1,2,2,1], padding='SAME'))

    layer2=tf.nn.relu(tf.nn.conv2d(layer1, filter=tf.Variable(tf.truncated_normal([5,5,16,32])), 
                                   strides=[1,2,2,1], padding='SAME'))
    layer2=tf.nn.bias_add(layer2, tf.Variable(tf.truncated_normal([32])))
    layer2=tf.nn.relu(tf.nn.max_pool(layer2, ksize=[1,1,1,1], strides=[1,2,2,1], padding='SAME'))

    layer3=tf.nn.relu(tf.nn.conv2d(layer2, filter=tf.Variable(tf.truncated_normal([5,5,32, 64])), 
                                   strides=[1,2,2,1], padding='SAME'))
    layer3=tf.nn.bias_add(layer3, tf.Variable(tf.truncated_normal([64])))

    layer3=tf.nn.relu(tf.nn.max_pool(layer3, ksize=[1,1,1,1], strides=[1,2,2,1], padding='SAME'))
    layer3=tf.nn.dropout(layer3, keep_prob=0.7)
    print(layer3.shape)


    fclayer1=tf.reshape(layer3, [-1, weights['fc1'].get_shape().as_list()[0]])
    fclayer1=tf.add(tf.matmul(fclayer1, weights['fc1']), biases['fc1'])
    fclayer1= tf.nn.dropout(fclayer1, keep_prob=0.5)
    fclayer2=tf.add(tf.matmul(fclayer1, weights['fc2']), biases['fc2'])
    fclayer2=tf.nn.dropout(fclayer2, keep_prob=0.5)
    fclayer3=tf.add(tf.matmul(fclayer2, weights['fc3']), biases['fc3'])
    fclayer3=tf.nn.dropout(fclayer3, keep_prob=0.7)
    logits = tf.add(tf.matmul(fclayer3, weights['out']), biases['out'])
    outLayer=tf.nn.softmax(logits)
    return outLayer, logits

In loss function:

model, logits = classifierModel(inp)
cost=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=y))
optimizer=tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)
correct_pred=tf.equal(tf.argmax(model, 1), tf.argmax(y, 1))
accuracy=tf.reduce_mean(tf.cast(correct_pred, tf.float32))