This repo contains solutions to some of the exercises in Jacob Hilton's Deep Learning Curriculum.
1-transformers: a pytorch implementation of a decoder-only transformer
for sequence tasks in transformer.py. It successfully completes a sequence reversal
task (on the second half of the sequence) in transformer_reverser.ipynb. In
transformer-shakespeare.ipynb I train it on the complete works of Shakespeare.
2-scaling: replicating scaling law results using convolutional nets of
different sizes on the MNIST dataset. The figure at the bottom of MNIST_scaling.ipynb
is a replication is Figure 2 in Kaplan et al.
3-parallelization: an implementation of data parallelization using MPI in data_parallel.py.
A comparison of this parallel training and sequential training of CNNs on MNIST in compare_parallel_single.ipynb.
6-RL: trained a Cartpole agent using naive policy gradient in policy_gradient_cartpole.ipynb.
Wrote a pytorch implementation of PPO in ppo.py and used it to trian a Lunar Lander agent in
ppo_lunar_lander.ipynb.