Ana Pekec - Summer 2023 Research Journal

Tuesday 05/30/23

I began reading papers and met with the group to plan out logistics for the summer.

Wednesday 05/31/23

First day at Nevis Labs. I completed the RCR training, setup my computing account at Nevis Labs and began the ROOT tutorials before the power outage ending the day early.

Thursday 06/01/23

I completed the ROOT tutorials on the Python path using pyroot in Juptyer notebooks. I also had our first meeting regular meeting with Ari Brill on the machine learning project and learned more background knowledge on neural networks.

Friday 06/02/23

Attended Ruo's lecture on Cherenkov radiation and IACT data analysis. For the rest of the day, I read and watched through multiple resources on transformers and attention. I started the PyTorch tutorials but on Google Colabs cloud server since the shevek machine is not available yet.

Monday 06/05/23

I setup a new conda environment for python notebooks. I started and finished the official PyTorch tutorials and also began watching more example tutorials of simple PyTorch networks online.

Tuesday 06/06/23

I completed an additional PyTorch tutorial (this excellent youtube video) where I coded a network for the MNIST and CIFAR10 datasets. Began looking through Ari Brill's code using fake data.

Wednesday 06/07/23

I worked a bit more on the CNN from yesterday's PyTorch tutorial to make the accuracy better at predicting the CIFAR10 images. Spoke with Ari on the basic structure of the CNN in his code. Met in weekly group meeting and discussed progress in each project.

Thursday 06/08/23

Before delving into Ari's code, I watched this lecture to better understand the purpose of attention in transformers. I also watched this youtube video coding a transformer with PyTorch and following along implementing the code into my own juptyer notebook. After, I read through Ari's code from attention until the StereoCNNClassifier class, commenting along the way.

Friday 06/09/23

Finished reviewing and understanding Ari's code on a call with him. Attended an REU lectures and lab tour of the XENON group. We also got the shevek machine set up and were able to open up jupyter notebook server from it.

Monday 06/12/23

Updated the 'cuda' drivers in shevek to make use of the gpu. Finally finished setting up shevek and got code running on shevek. Took only 2 minutes to train Ari's network.

Tuesday 06/13/23

Started working with the CTA data. Manually went through the HDF5 file format and tried to unpack the data into python using hyp5 package. Attended an REU lecture and lab tour of the RARAF group.

Wednesday 06/14/23

Had a meeting with Ari to discuss ways to read the HDF5 CTA files in python. Decided on using pytables and the dl1 reader packages to unpack the data. We also mapped out our model and discussed the attention algorithm more conceptually. Attended VERITAS group meeting and presented model diagram.

Thursday 06/15/23

Catalogued old machines in the shop for most of the day. Used dl1_data_handler to plot the LST data.

Friday 06/16/23

Sick day.

Tuesday 06/20/23

Coded a CTADataset class to load simulated CTA data using dl1 reader. Link for runs: https://docs.google.com/document/d/1laeUW_R9TvvSSxgs6ZXxm37WJkBIqetK841qxCOUJe4/edit?usp=sharing

Wednesday 06/21/23

Attended VERITAS group meeting and chatted with Ari. Started running efficiency analysis by adding num_workers and found that 3 workers seemed to be the most efficient. Also determined a seed for randomness so that each run would be reproducibile.

Thursday 06/22/23

Catalogued machines in the shop (almost at 100 so far) most of today. Worked on fixing a bug in the dl1 reader class so that tel_ids do not have to be sequential. Still having errors from this bug.

Friday 06/23/23