TL;DR

I learned that PyTorch was easier to use than I thought.

Today's Research Goals

• Start reading about Meta's MDETR model
• Learn how to use PyTorch

Completed Tasks

• I installed Nvidia's CUDA toolkit. Whew.
• Followed a PyTorch guide and made a script that trains an image classification model
• Learned how to load data in PyTorch
• Learn how autograd works in PyTorch

Look at the thingies!

![A dog, a truck, a plane, and a dog]("./assets/Dog Truck Plane Dog.png")

Unresolved Things

• Still need to actually read the whole MDETR paper and successfully train the model.
• I have forgotten how to graph an image directly from a tensor using matplotlib.

I am pretty sure this is not an image:

![A dog, a truck, a plane, and a dog]("./assets/Image Graph.png")

Questions From Today

• (Completely off-topic, but) How can one do neurosymbolic learning with PyTorch?
• Is PyTorch Lightning any good?

Other Notes

I feel like a dunce trying to do basic image classification with CIFAR in PyTorch because I feel like I should be past this. However, I suppose it's better to

I used to find autograd completely unapproachable, but the PyTorch docs were relatively approachable.

I also see why researchers have gravitated to PyTorch instead of TensorFlow.