NYCU Deep Learning Labs

This repository collects my coursework implementations for the NYCU Deep Learning course. The projects are written mainly in PyTorch and cover both computer vision and reinforcement learning workflows, from model implementation to experiment tracking, evaluation, and final reporting.

Labs

Lab	Topic	Main Work	Results / Artifacts
Lab 2	Binary Semantic Segmentation	Implemented U-Net and ResNet34-UNet from scratch for Oxford-IIIT Pet foreground segmentation. Built dataset loading, preprocessing, BCE + Dice loss training, validation, inference, and Kaggle submission scripts.	U-Net reached best validation Dice 0.9187. Both U-Net and ResNet34-UNet checkpoints and Kaggle submission pipelines are included.
Lab 5	Value-based Reinforcement Learning	Implemented DQN for CartPole and Atari Pong, then extended it with Double DQN, Prioritized Experience Replay, multi-step returns, and Dueling DQN. Used W&B for training curves and evaluation tracking.	CartPole achieved 500.00 average reward over seeds 0-19. Vanilla Pong DQN reached 19.05 average reward. Enhanced DQN reached 19.10 at 600k steps and 20.15 at 1.5M steps.
Lab 6	Conditional Image Generation	Implemented a conditional DDPM for i-CLEVR using a PyTorch U-Net, cosine noise schedule, classifier-free guidance, EMA weights, DDIM sampling, and evaluator-based checkpoint selection.	Reached evaluator accuracy 0.875000 on test.json and 0.857143 on new_test.json. Generated ordered PNG outputs, testing grids, and a denoising process grid.

Repository Layout

.
|-- lab2/
|   |-- src/
|   |   |-- models/
|   |   |   |-- unet.py
|   |   |   `-- resnet34_unet.py
|   |   |-- train.py
|   |   |-- evaluate.py
|   |   |-- inference.py
|   |   |-- oxford_pet.py
|   |   `-- utils.py
|   |-- saved_models/
|   |-- requirements.txt
|   `-- Lab2_Binary_Semantic_Segmentation_2026_Spring.pdf
|
|-- lab5/
    |-- LAB5_B11107027_Code/
    |   |-- dqn.py
    |   |-- test_model.py
    |   |-- eval_task1.sh
    |   |-- eval_task2.sh
    |   |-- eval_task3.sh
    |   `-- requirements.txt
    |-- figure/
    |-- report.tex
    |-- LAB5_B11107027.pdf
    `-- LAB5_B11107027.zip

`-- lab6/
    |-- src/
    |   |-- dataset.py
    |   |-- diffusion.py
    |   |-- ema.py
    |   |-- evaluate.py
    |   |-- models.py
    |   |-- sample.py
    |   |-- train.py
    |   |-- utils.py
    |   `-- __init__.py
    |-- images/
    |-- requirements.txt
    |-- README.md
    `-- DL_LAB6_B11107027_陳映宬_report.pdf

Lab 2: Binary Semantic Segmentation

Lab 2 focuses on foreground segmentation for the Oxford-IIIT Pet dataset. The goal is to predict a binary pet mask from an RGB image.

Implemented components:

• U-Net with valid convolutions and skip connections.
• ResNet34-UNet with a ResNet34-style encoder and U-Net decoder.
• Oxford-IIIT Pet dataset loader with mask preprocessing.
• BCE + Dice loss to reduce foreground/background collapse.
• Validation with global Dice and IoU.
• Inference scripts for Kaggle CSV submission generation.

Key result:

• U-Net best validation Dice: 0.9187.
• Generated 739-image Kaggle prediction files for both U-Net and ResNet34-UNet.

Lab 5: Value-based Reinforcement Learning

Lab 5 studies DQN and enhanced DQN methods on CartPole and Atari Pong.

Implemented components:

• Vanilla DQN with replay buffer, target network, Bellman target computation, and epsilon-greedy exploration.
• Atari preprocessing with grayscale conversion, 84 x 84 resizing, and 4-frame stacking.
• CNN Q-network for Pong.
• Double DQN for decoupled action selection and target evaluation.
• Prioritized Experience Replay with importance-sampling weights.
• Multi-step return support.
• Dueling DQN architecture.
• W&B logging for reward curves, evaluation scores, and ablation comparisons.

Key results:

Task	Environment	Result
Task 1	CartPole-v1	Average reward 500.00, min 500.0, max 500.0 over seeds 0-19.
Task 2	ALE/Pong-v5	Vanilla DQN average reward 19.05 over seeds 26-45.
Task 3	ALE/Pong-v5	Enhanced DQN average reward 19.10 at 600k steps and 20.15 at 1.5M steps.

The Lab 5 report also includes required ablations for no PER, no Double DQN, and no multi-step return, plus additional bonus analysis for dueling networks, learning-rate decay, n-step choices, soft target updates, and target update frequency.

Lab 6: Conditional DDPM for i-CLEVR

Lab 6 implements a conditional DDPM to generate i-CLEVR images from multi-label object conditions.

Implemented components:

• i-CLEVR dataset loader and multi-hot condition encoding.
• Conditional U-Net noise predictor with residual blocks and self-attention.
• Sinusoidal timestep embeddings and MLP-based condition embeddings.
• DDPM training with a cosine beta schedule and Huber noise-prediction loss.
• Classifier-free guidance by randomly dropping conditions during training.
• EMA parameter tracking for more stable sampling.
• DDIM sampling for faster generation.
• Evaluation script using the provided frozen ResNet18 evaluator.

Key results:

Split	Accuracy
test.json	0.875000
new_test.json	0.857143

The generated image folder contains ordered PNG files for both testing splits, two synthetic image grids, and the required denoising process grid.

Reproducing Experiments

Each lab has its own dependency file.

Lab 2:

cd lab2
pip install -r requirements.txt
python src/train.py --model unet
python src/evaluate.py --model unet --weight saved_models/unet_best.pth

Lab 5:

cd lab5/LAB5_B11107027_Code
pip install -r requirements.txt
bash eval_task1.sh
bash eval_task2.sh
bash eval_task3.sh

Lab 6:

cd lab6
pip install -r requirements.txt
python -m src.sample --meta-dir file/file --ckpt /path/to/best_ema.pt --out-dir /content/images --split both
python -m src.evaluate --meta-dir file/file --image-dir /content/images --split both

Notes

• The repository is organized by lab folder so each assignment can be inspected independently.
• Large datasets, checkpoints, and generated submission artifacts may be kept locally or in submission packages depending on course requirements.
• The code is coursework-oriented and prioritizes reproducibility of the lab results and reports.