Thank you for the response.
The script is actually reading all the input data (940MB) into memory and then trains the neural network. I am using Keras to build the neural network. The network is a simple feedforward network with 8 layers (512, 256, 128, 128, 64, 32, 16, 8).
Earlier I had calculated the complete project execution time which included reading the training and test files, and training and testing the model. Since you mentioned that disk read/writes could be the bottleneck, I timed only the training and testing of the model. But I could not still see much improvement in the execution time when compared to the duration taken on my laptop.
I have also made the project public. Please let me know if I am doing something wrong:
Link to the latest execution: