assignment_llm_1/assignment_image/install_env_and_run.md

Install Environment (Conda) and Run Scripts

This guide sets up a Conda environment and runs both scripts:

• assignment_image/code/c1.py (train + save checkpoint)
• assignment_image/code/c1_test.py (evaluate + error analysis)

First need to enter this folder:

cd assignment_llm_1/assignment_image

1) Create and activate Conda environment

conda create -n transformer_hw python=3.10 -y
conda activate transformer_hw
python -m pip install --upgrade pip

2) Install dependencies

If there is a requirements.txt file in this folder, run:

pip install -r requirements.txt

3) Run training script (c1.py)

Move to the code directory and run:

python code/c1.py

Expected outputs include:

• saved_model/vit_cifar10_best.pt
• saved_model/vit_cifar10_last.pt

4) Run evaluation script (c1_test.py)

After training completes:

python code/c1_test.py \
  --checkpoint-path ./saved_model/vit_cifar10_best.pt \
  --results-dir ./results

This baseline evaluation run saves:

• results/baseline_analysis.txt
• results/misclassified_examples_test.png

5) Run optional pre-trained ViT comparison

To run transfer learning and compare baseline vs pre-trained ViT:

python code/c1_test.py \
  --checkpoint-path ./saved_model/vit_cifar10_best.pt \
  --results-dir ./results \
  --run-pretrained-experiment

Additional files saved in this mode:

• results/pretrained_vit_analysis.txt
• results/misclassified_examples_pretrained_vit.png
• results/comparison_report.txt

6) Where data and outputs are saved

• Dataset download/cache: ./data
  (both c1.py and c1_test.py load CIFAR-10 from this folder by default)
• Model checkpoints from training: ./saved_model
• Evaluation artifacts/reports: ./results (or the path passed with --results-dir)
• Default checkpoint used by evaluation: ./saved_model/vit_cifar10_best.pt

Quick path summary

• Training command: python code/c1.py
• Baseline evaluation: python code/c1_test.py --checkpoint-path ./saved_model/vit_cifar10_best.pt --results-dir ./results
• Baseline + transfer comparison: python code/c1_test.py --checkpoint-path ./saved_model/vit_cifar10_best.pt --results-dir ./results --run-pretrained-experiment