A convolutional neural networks (CNN)-based deep learning algorithm using the Keras Sequential API for image classification.
Trained with the CIFAR-100 dataset.
Expected: motorcycle
Expected: tractor
Expected: lion
Well... it (kinda) works!
While CIFAR-10 is easy to work with, training on CIFAR-100 certainly poses some limitations as it is not easy to train a really deep learning model and is not feasible to get results as high as 90% such as in MNIST. However, this has really been an interesting challenge for me to see how far this dataset could go. To save training time, I have used a batch size of 250 with 20 epochs and 0.2 validation split on my neural network, using ReLu as activation function. Due to this nature, after a series of trail, this model has only been able to give a testing accuracy of 40.28%. If we go for a more robust model with more computing resources, it is possible to achieve a better result like what we see on https://paperswithcode.com/sota/image-classification-on-cifar-100.
An alternate apporach would be the model used by Djork-Arné Clevert, Thomas Unterthiner, Sepp Hochreiter which uses ELU as activation function as re-implemented in this article https://medium.com/@birdortyedi_23820/deep-learning-lab-episode-5-cifar-100-a557e19219ba. While this method could give a top-1 accuracy of 75.72%, it would reuqire harsh requirements such as 165000 epochs. For the specs of my pc (2070) it is still expected to run at least 20 days non-stop. Anyway, I have learnt quite a lot in this project, and that's what important ^^.
Inspired by Youtube Channel Computer Science https://www.youtube.com/watch?v=iGWbqhdjf2
CIFAR-10 and 100 https://www.cs.toronto.edu/~kriz/cifar.html