CIFAR 10 - Image Classifier

We are given an image classifier, CIFAR-10, here the input images are processed and classified under certain categories. Our objective is to improve the average accuracy of this image classification. I have improved the accuracy to 63% in this model.

First we begin by understanding the basic CIFAR-10 Image Classifier code obtained.

We start the program by loading the torch package which is helpful in loading and normalizing the training and test datasets.

Output:

Now we get some random training images visualized.

Output:

Now we design a Convolution Neural Network (CNN) with 2 convolution layers conv1 and conv2 and 3 fully connected layers fc1, fc2 and fc3.

Then we design a loss function and optimizer

Now we start training our model through iteration loop.

Output:

We save our trained model now.

We now test the trained data set as a trial.

Output:

We load the previous saved model.

Output:

Now the neural network classifies the data.

We select the index of highest energy.

Output:

Since the trial worked perfectly, we now test the whole data set. Here we find the accuracy of the whole network.

Output:

We also measure the performance of each class.

Output:

---

My Contribution:

Epoch refers to the number of times the model is iterated. As per the referenced blogpost I came across, it is possible to increase accuracy up-to 94% by increasing epoch to 50.

For my image classifier, I increased the epoch from 2 to 9. As a result, the accuracy came up to 63%.

Output:

We compute the final accuracy now.

Output:

My Submission

---

Challenges Faced:

The biggest challenge in running high number of epoch was that the execution time increased drastically. The epoch and the execution time was almost directly proportional.

Execution Time taken with respect to Epoch

Another challenge I faced was overfitting. After a certain point the accuracy went down and the model wasn't usable anymore. Even though the referenced blogpost suggested that epoch 50 can give accuracy of 94%, it did not work on my model due to overfitting.

Accuracy with respect to Epoch

---

Experiments:

I tried to increase the convolution layers from 2 to 3, 4, 5, 6 and even 7, with self-pooling, to improve the accuracy.

Unfortunately, changing these hypermeters wasn't very effective as the maximum accuracy I could reach was 56% in all the 6 trials.

Accuracy with respect to Convolution Layers

---

References:

Reference for the code and it's explanation of the CIFAR-10 Image Classifier: https://pytorch.org/tutorials/beginner/blitz/cifar10_tutorial.html#sphx-glr-beginner-blitz-cifar10-tutorial-py

Reference for the epoch value changing in the section of 'My Contribution': https://blog.fpt-software.com/cifar10-94-of-accuracy-by-50-epochs-with-end-to-end-training

Reference for the image of the blog: https://www.cs.toronto.edu/~kriz/cifar.html