rows = [
    {"experiment": "baseline", "f1": 0.71, "note": "logreg"},
    {"experiment": "+cleaning", "f1": 0.75, "note": "remove leakage"},
]
for row in rows:
    print(row)

baseline
baseline + feature engineering
baseline + feature engineering + tuned hyperparameters

mkdir -p ~/Code/ustc-ai/week12-ai-project-review
cd ~/Code/ustc-ai/week12-ai-project-review
uv init --name week12-ai-project-review
uv venv
source .venv/bin/activate.fish
uv add pandas matplotlib
code .

mkdir -p reports figures src

project-root/
├── README.md
├── solution.md
├── reports/
│   ├── experiments.csv
│   ├── experiment_table.md
│   └── project_review.md
├── figures/
│   └── accuracy_by_experiment.png
└── src/
    └── summarize_experiments.py

id,date,model,change,valid_metric,public_score,notes
exp001,2026-07-26,MLP,baseline,0.921,,first clean baseline
exp002,2026-07-26,MLP,hidden_dim 128 to 256,0.934,,larger model improved validation
exp003,2026-07-26,CNN,small cnn baseline,0.958,,cnn uses image structure
exp004,2026-07-26,CNN,remove second conv layer,0.941,,ablation: second conv helps

from __future__ import annotations
from pathlib import Path
import matplotlib.pyplot as plt
import pandas as pd
EXPERIMENTS_CSV = Path(“reports/experiments.csv”)
TABLE_MD = Path(“reports/experiment_table.md”)
FIGURE_PATH = Path(“figures/accuracy_by_experiment.png”)
def main() -> None:
experiments = pd.read_csv(EXPERIMENTS_CSV)
required_columns = {“id”, “model”, “change”, “valid_metric”, “notes”}
missing_columns = required_columns - set(experiments.columns)
if missing_columns:
raise ValueError(f”missing columns: {sorted(missing_columns)}”)
plain TABLE_MD.parent.mkdir(parents=True, exist_ok=True)
FIGURE_PATH.parent.mkdir(parents=True, exist_ok=True)
plain table = experiments[[“id”, “model”, “change”, “valid_metric”, “notes”]]
TABLE_MD.write_text(table.to_markdown(index=False), encoding=“utf-8”)
plain plt.figure(figsize=(8, 4))
plt.plot(experiments[“id”], experiments[“valid_metric”], marker=“o”)
plt.xlabel(“experiment”)
plt.ylabel(“validation metric”)
plt.title(“Validation metric by experiment”)
plt.xticks(rotation=30, ha=“right”)
plt.tight_layout()
plt.savefig(FIGURE_PATH, dpi=150)
plt.close()
best = experiments.sort_values("valid_metric", ascending=False).iloc[0]
print(f"best_experiment={best['id']} valid_metric={best['valid_metric']}")
print(f"wrote {TABLE_MD}")
print(f"wrote {FIGURE_PATH}")
if name == “main”:
main()

source .venv/bin/activate.fish
python src/summarize_experiments.py

uv add tabulate
python src/summarize_experiments.py

test -f reports/experiment_table.md; and test -f figures/accuracy_by_experiment.png

# AI Competition / Deep Learning Project Review
Task

Problem:
Input:
Output:
Metric:

Dataset

Source:
Size:
Features / modalities:
Train / validation split:
Important preprocessing:

My Contribution

I implemented:
I cleaned:
I trained:
I analyzed:

Baseline

Model:
Why this baseline:
Baseline result:

Improvements

Improvement 1:
Improvement 2:
Improvement 3:

Experiments
Paste or link reports/experiment_table.md here.
Ablation

Removed component:
Result change:
Interpretation:

Result

Best validation score:
Public leaderboard score if any:
Final selected model:

Failure Cases

Case 1:
Case 2:
What these failures suggest:

Limitations

Dataset limitation:
Evaluation limitation:
Model limitation:
Engineering limitation:

What I Learned

Technical lesson:
Statistical lesson:
Engineering lesson:

# Project Name
Summary
One paragraph explaining the task, method, and best result.
File Layout
configs/      experiment configs
src/          training and inference code
reports/      experiment table and review report
figures/      generated charts
data/raw/     local raw data, not committed
Environment
uv venv
source .venv/bin/activate.fish
uv sync
Train
python src/train.py --model mlp --epochs 20
Evaluate
python src/evaluate.py --checkpoint checkpoints/best_mlp.pt
Results




















Model
Validation Metric
Notes
baseline
fill
fill
best
fill
fill
Limitations

Limitation 1.
Limitation 2.
Limitation 3.

- Validation uses a single random split; cross-validation was not performed.
- The dataset is small and clean, so the result may not transfer to noisy real-world data.
- Hyperparameter search was limited to three manual trials due to CPU-only training.
- Failure cases suggest the model confuses visually similar classes.

- My computer was bad.
- The model could be better.
- Need more time.

I participated in the competition.

I cleaned the tabular features, built a RandomForest baseline, fixed the train / inference feature mismatch, and wrote an ablation comparing categorical encoding strategies. The final local macro-F1 improved from 0.742 to 0.781.

我做了什么 -> 解决了什么问题 -> 指标或证据是什么

source .venv/bin/activate.fish
test -f reports/experiments.csv; and test -f reports/project_review.md; and test -f README.md
python src/summarize_experiments.py
test -f reports/experiment_table.md; and test -f figures/accuracy_by_experiment.png

We achieved 0.958 accuracy.

The CNN baseline achieved 0.958 validation accuracy. Compared with the MLP baseline at 0.934, the CNN improved performance by using local image structure.

source .venv/bin/activate.fish