/UCBerkeley-Internship

Papers-Notes:

This repository contains my tasks summary at the end of each day, paper (Pa) and posts (Po) reading notes on deep learning and machine learning, as well as tutorials covered, mainly focused on my thesis "Deep Learning based Multi-Object Tracking and Motion Prediction for Intelligent Vehicles applications" during my internship at UCBerkeley (April - July 2022).

Courses:

(NYU) Deep Learning Course: https://atcold.github.io/pytorch-Deep-Learning/
(UPenn) Graph Neural Networks Course: https://gnn.seas.upenn.edu/

Notes:

April

1st:

  • First contact with the MSC lab.

  • Interview with the advisor (Wei Zhan) and PostDoc student (Chen Tang) to define my tasks.

  • Start to sort my UCBerkeley documents.

4th:

  • Finish to sort my UCBerkeley documents.

  • Start to study the NYU Deep Learning course -> Week 1 (50:33 YT video).

  • Start to study the UPenn GNNs course (Introduction covered)

  • Join to MSC_Auto Slack and get papers to read about the new project: Explainable AI to study behavior prediction models

5th:

  • Almost all paperwork is done.

  • Read Project "Explainable AI for behaviour-prediction models".

  • Finish to watch the first lecture (NYU-DL): "History, motivation, and evolution of Deep Learning". What's left: Highlight the theory of 1.1. and 1.2.

  • See Lecture 1.1. "Graph Neural Networks" (UPenn-GNNs).

  • Install LaneGCN and start reading its paper.

6th:

  • See Practicum 1 (NYU-DL). What's left: Highlighy Lecture 1 and Practicum 1 theory, and run practicum 1 code.

  • See Lecture 1.2. "Machine Learning on Graphs: The Why" (UPenn-GNNs).

  • See Video 1 (Essence of Linear Algebra) and SVD (Singular Value Decomposition).

  • Start reading "You Mostly Walk Alone. Analizing Feature Attribution in Trajectory Prediction" paper (arXiv preprint).

7th:

  • Finish reading "You Mostly Walk Alone. Analizing Feature Attribution in Trajectory Prediction" paper (arXiv preprint).

8th:

  • Read post "Kullback-Leibler Divergence for Machine Learning".

  • Read post "A Quick Intro to Leave-One-Out Cross-Validation (LOOCV)".

  • Read post "Review. Multimodal Trajectory Predictions for Autonomous Driving Using Deep Convolutional Networks".

  • Almost finished with "Exploring Social Posterior Collapse in Variational Autoencoder for Interaction Modeling" paper.

  • Start reading "Interventional Behavior Prediction Avoiding Overly Confident Anticipation in Interactive Prediction" paper.

  • Start the second lecture (NYU-DL): "Stochastic gradient descent and backpropagation" (6:53 YT video)

13th:

  • Finish reading "Exploring Social Posterior Collapse in Variational Autoencoder for Interaction Modeling" paper (NeurIPS 2021)

  • Request Cal 1 Card

  • Finish reading "Interventional Behavior Prediction: Avoiding Overly Confident Anticipation in Interactive Prediction" (in submission, IROS 2022)

  • Start reading about Shapley values https://christophm.github.io/interpretable-ml-book/shapley.html

18th:

  • Read post "How to Manage and Restore Tmux Sessions in Linux".

  • Start reading "Learning Lane Graph Representations for Motion Forecasting" paper (ECCV 2020).

  • Enhanced README of "Exploring Attention GAN for Vehicle Motion Prediction" paper (in submission, ITSC 2022).

  • Finish 00-logic_neuron_programming.ipynb assignment (DL course).

  • Continue reading Shapley values.

  • Continue studying GNNs Lecture 1.

19th:

  • Finish reading "Learning Lane Graph Representations for Motion Forecasting" paper (ECCV 2020).

  • Start reading "VectorNet: Encoding HD Maps and Agent Dynamics from Vectorized Representation"

  • Finish 01-tensor_tutorial.ipynb assignment (DL course).

  • Finish reading Shapley values.

20th:

  • Organize Workshop: "Accelerating_CUDA_C++_Applications_with_Multiple_GPUs" (NVIDIA DLI).

  • Read Part 1 Workshop: "Building Transformer-Based Natural Language Processing Applications".

  • Finish Part 1 (Hand-on) Workshop: "Building Transformer-Based Natural Language Processing Applications".

  • Continue reading "VectorNet: Encoding HD Maps and Agent Dynamics from Vectorized Representation" paper (CVPR 2020).

21th:

  • Set final objectives for ICRA 2nd round.

  • Finish reading "VectorNet: Encoding HD Maps and Agent Dynamics from Vectorized Representation" paper (CVPR 2020).

25th:

  • Submit ICRA 2nd round. Let's pray!!!!.

  • Finish reading post "Supervised, Semi-Supervised, Unsupervised, and Self-Supervised Learning".

  • Fix SSH issue with DENSO workstation .

26th:

  • Read post "Weight Decay == L2 Regularization?".

  • Start training LaneGCN in the DENSO workstation. Not working in my UAH pc (almost sure due to CUDA and cuDNN versions).

  • Download Argoverse 2.0.

  • Read post "Interpretability vs Explainability: The Black Box of Machine Learning".

27th:

  • Finish reading "A Unified Approach to Interpreting Model Predictions" paper (NeurIPS 2021).

  • Start working with SHAP values.

Total papers: 6
Total posts: 8 \

May

2nd:

  • Read post "Partial Dependence and Individual Conditional Expectation plots".

  • Read post "An introduction to explainable AI with Shapley values"

  • Run "Census income classification with Keras" notebook (SHAP values for DL models using Kernel Explainer).

  • Start studying SHAP values for Motion Prediction (Implementation for Trajectron++).

3rd - 6th:

  • Keep studying SHAP values.

  • Project XAI is discarded due to its complexity and lack of knowledge in GNNs + XAI.

9th - 13th:

  • Studied Week 1 - Week 5 (DL course NYU)

16th:

  • Reading post "Regularization techniques for training deep neural networks"

  • Fix CARLA Leaderboard CUDA issue -> --gpus all flag in the launch_docker_image_v2.sh script

  • Studying Week 6 (RNNs) (DL course NYU)

17th - 20th:

  • Start refactoring Motion Prediction repository.

  • Prepare camera-ready for ICRA - Fresh Perspectives on the Future of Autonomous Driving workshop

23rd:

  • Refactorize dataset.

  • Start studying Set transformer (solution from Lyft dataset)

24th:

  • Read post "Winning solution for Kaggle challenge: Lyft Motion Prediction for Autonomous Vehicles"

  • Continue working with the Set transformer + Goals

25th - 27th:

31st:

  • Add option to select CUDA device (mapfe4mp)

  • Work on IV 2022 qualitative results for oral presentation

Total papers:
Total posts: 3 \

June

1st - 10th:

  • Attend IV (Intelligent Vehicles) 2022 conference

13th:

  • Work on mapfe4mp code

  • Check "check_accuracy" function in trainer. It seems to work fine

  • Add map_location option to torch.load (if you want to restore checkpoint) -> torch.load(restore_path, map_location=current_cuda)

  • Study the model to observe what I can improve

  • Add visualizator. Check that goal point generator and visualizator show the same scenes. Organize model pptx. Extract conclusions from train and validation split

Total papers:
Total posts: \

July

Total papers:
Total posts: \