NTHU Face Segmentation

Environment Setup

1. CUDA 11.8 Installation:

   For optimal performance, install CUDA 11.8 from the official Nvidia website. We recommend using the runfile(local) option for installation.

2. Anaconda Installation:

   Download and install Anaconda from the official website: Anaconda Download.

3. Environment Setup:

   Open a terminal and run the following commands to create an environment using the provided environment.yml file:

   conda env create -f environment.yml
   conda activate face

4. Install dependency

   Open a terminal and run the instruction

   pip install facenet-pytorch

Data Preparation

CelebAMask-HQ

1. Download the CelebAMask-HQ.zip file from the dataset source and unzip it into the ./data directory. Ensure the extracted folder structure matches the one shown below.

2. Download the test_idx.txt and train_idx.txt files from the following Kaggle dataset Link. Place these files in the ./data directory.

3. After extracting the files and preparing the directory structure, run the script dataset/preprocess.py to preprocess the images. This script will automatically process and combine the images, placing them in the ./data/CelebAMask-HQ/CelebAMask-HQ-combined_mask folder.

Unseen Dataset

1. Download the Unseen.zip dataset from the kaggle competition website

2. Extract the files into ./data/Unseen folder. Thats all

   Tree Structure Example

   Face_Competition
   ├── checkpoint
   │ └── checkpoint.pth
   ├── data
   │ ├── CelebAMask-HQ
   │ │ ├── CelebA-HQ-img
   │ │ │ ├── 0.jpg
   │ │ │ ├── 1.jpg
   │ │ │ ├── 2.jpg
   │ │ │ ├── ...
   │ │ │ └── 2999.jpg
   │ │ ├── CelebAMask-HQ-mask-anno
   │ │ │ ├── 0
   │ │ │ ├── 1
   │ │ │ ├── 2
   │ │ │ ├── ...
   │ │ │ └── 14
   │ │ ├── CelebAMask-HQ-combined_mask
   │ │ │ ├── 0.jpg
   │ │ │ ├── 1.jpg
   │ │ │ ├── 2.jpg
   │ │ │ ├── ...
   │ │ │ └── 2999.jpg
   │ │ ├── CelebA-HQ-to-CelebA-mapping.txt
   │ │ ├── CelebAMask-HQ-attribute-anno.txt
   │ │ └── CelebAMask-HQ-pose-anno.txt
   │ ├── Unseen
   │ │ ├── 2569520_1.jpg
   │ │ ├── 3013103_0.jpg
   │ │ ├── ...
   │ │ └── 14039645295_0.jpg
   │ ├── test_idx.txt
   │ └── train_idx.txt
   ├── dataset
   ├── model
   ├── Readme.md
   ├── run.py
   └── visualize.py
   

Training and Testing

1. Run the Notebook:

Open the run.ipynb file and select the face kernel. Click Run All to execute the entire notebook.

2. Parameter Explanation:

The notebook defines several key parameters:

Paths:

• CHECKPOINT_PATH: Path to the model weights for resuming training or performing inference.
• TRAIN_INDEX_PATH: Path to the list of training data indexes.
• TEST_INDEX_PATH: Path to the list of testing data indexes.
• UNSEEN_INDEX_PATH: Path to the list of unseen data indexes.

Global Variables:

• MODE: Training mode (train) or testing mode (test).
• RECORD: Whether to use Weights & Biases (wandb) to record training losses.
• SAVE_MODEL_NAME: Name of the saved model in the checkpoint/ folder.
• SAVE_IMAGES: If the mode is in csv mode. Set weather to save the output image to file.
• EXPORT_TO_CSV_AFTER_TRAIN: Export the prediction mask.csv after training
• JUPYTER_NOTEBOOK: Use for identify the python script or jupyter notebook
• UNSEEN: (Boolean) Set the testing or csv mode for unseen data or CelebAMask-HQ data
• SAVE_CSV: (Debug) Set if save the csv file if in csv mode
• SHOW_IMAGES_PER_INDEX: Show pictures every [SHOW_IMAGES_PER_INDEX] frames

Hyperparameters:

• EPOCHS: Maximum number of training epochs.
• BATCH_SIZE: Batch size for both the training and testing data loaders.
• LEARNING_RATE: Optimizer learning rate.

Checkpoints

• Google Drive