# How to Fine-Tune a Pre-Trained PFLD Model on Custom Facial Landmark Data > Learn how to fine-tune a pre-trained PFLD model with your custom facial landmark data. This guide covers data preparation, weight loading, hyperparameter tuning, and training for accurate results. - Repository: [Guoqiang QI/pfld](https://github.com/guoqiangqi/pfld) - Tags: how-to-guide - Published: 2026-03-03 --- **Fine-tuning a pre-trained PFLD model involves converting your dataset to NumPy format with [`data/SetPreparation.py`](https://github.com/guoqiangqi/pfld/blob/main/data/SetPreparation.py), loading pre-trained weights in [`model2.py`](https://github.com/guoqiangqi/pfld/blob/main/model2.py), adjusting hyperparameters in [`train_model.py`](https://github.com/guoqiangqi/pfld/blob/main/train_model.py) (typically lowering the learning rate to 1e-5 and freezing early backbone layers), and running the training loop before validating with [`test_model.py`](https://github.com/guoqiangqi/pfld/blob/main/test_model.py) and deploying via [`camera.py`](https://github.com/guoqiangqi/pfld/blob/main/camera.py).** The PFLD (Progressive Face Localization and Detection) repository provides a lightweight facial-landmark detector optimized for real-time applications. This guide explains how to fine-tune a pre-trained PFLD model on your own dataset to improve accuracy on domain-specific imagery while retaining the model's efficient MobileNet-style architecture. ## Prepare Your Training Dataset ### Convert Raw Annotations to NumPy Format The training pipeline expects face images and their corresponding 68-point (or 98-point) landmark annotations in NumPy format. Use the helper script [`data/SetPreparation.py`](https://github.com/guoqiangqi/pfld/blob/main/data/SetPreparation.py) to convert raw data into the expected format. ```bash python data/SetPreparation.py --input_dir data/raw_images \ --anno_file data/annotations.txt \ --output_dir data/prepared ``` ### Optional Data Augmentation For dataset compression or augmentation, utilize utilities in the `tools/` directory. The [`tools/frame_cut.py`](https://github.com/guoqiangqi/pfld/blob/main/tools/frame_cut.py) script handles frame extraction, while [`tools/emotion_compress.py`](https://github.com/guoqiangqi/pfld/blob/main/tools/emotion_compress.py) manages compression tasks. ## Configure the Model Architecture ### Load Pre-Trained Weights The core network implementation resides in [`model2.py`](https://github.com/guoqiangqi/pfld/blob/main/model2.py). This file defines a MobileNet-style backbone followed by depth-wise separable convolutions that output landmark coordinates. It contains the logic to load pre-trained weights trained on the original 300-W dataset, serving as the foundation for fine-tuning. ## Configure Fine-Tuning Parameters ### Adjust Hyperparameters for Transfer Learning Hyperparameters including learning rate, batch size, number of epochs, and loss weights are defined in [`train_model.py`](https://github.com/guoqiangqi/pfld/blob/main/train_model.py). For effective fine-tuning: - **Lower the learning rate** from the default `1e-4` to `1e-5` to prevent catastrophic forgetting - **Freeze early backbone layers** to preserve generic feature representations - **Allow later layers to adapt** to new data distributions ## Execute the PFLD Fine-Tuning Process The [`train_model.py`](https://github.com/guoqiangqi/pfld/blob/main/train_model.py) script orchestrates the complete training workflow: - Loads data from the NumPy files generated in the preparation step - Constructs the model architecture - Computes losses using auxiliary pose regularization terms from [`euler_angles_utils.py`](https://github.com/guoqiangqi/pfld/blob/main/euler_angles_utils.py) - Updates optimizer states - Logs progress to [`data/log0.txt`](https://github.com/guoqiangqi/pfld/blob/main/data/log0.txt) - Saves periodic checkpoints to the specified directory ```bash python train_model.py \ --train_data data/prepared/train.npy \ --val_data data/prepared/val.npy \ --pretrained model2.py:pretrained_weights.pth \ --batch_size 64 \ --epochs 100 \ --learning_rate 1e-5 \ --freeze_backbone True \ --log_file data/log0.txt \ --checkpoint_dir checkpoints/ ``` ## Validate and Deploy the Fine-Tuned Model ### Evaluate Model Performance Run [`test_model.py`](https://github.com/guoqiangqi/pfld/blob/main/test_model.py) to assess accuracy on held-out validation data: ```bash python test_model.py \ --model checkpoints/best_finetuned.pth \ --test_data data/prepared/test.npy ``` ### Real-Time Inference Export the fine-tuned weights (saved as `.pth` files) and integrate them into the inference pipeline. Replace the original checkpoint in [`camera.py`](https://github.com/guoqiangqi/pfld/blob/main/camera.py) to observe updated facial-landmark predictions in real time: ```bash python camera.py --model checkpoints/best_finetuned.pth ``` ## Summary - **Data Preparation**: Convert raw images and 68-point or 98-point annotations to NumPy format using [`data/SetPreparation.py`](https://github.com/guoqiangqi/pfld/blob/main/data/SetPreparation.py) - **Model Initialization**: Load pre-trained weights via [`model2.py`](https://github.com/guoqiangqi/pfld/blob/main/model2.py) to leverage 300-W dataset knowledge - **Fine-Tuning Strategy**: Reduce learning rate to `1e-5` and freeze backbone layers in [`train_model.py`](https://github.com/guoqiangqi/pfld/blob/main/train_model.py) to preserve generic features - **Training Execution**: Run [`train_model.py`](https://github.com/guoqiangqi/pfld/blob/main/train_model.py) with pose regularization from [`euler_angles_utils.py`](https://github.com/guoqiangqi/pfld/blob/main/euler_angles_utils.py) and monitor logs in [`data/log0.txt`](https://github.com/guoqiangqi/pfld/blob/main/data/log0.txt) - **Deployment**: Validate with [`test_model.py`](https://github.com/guoqiangqi/pfld/blob/main/test_model.py) and deploy via [`camera.py`](https://github.com/guoqiangqi/pfld/blob/main/camera.py) for real-time facial landmark detection ## Frequently Asked Questions ### What is the expected input format for fine-tuning a PFLD model? The training pipeline expects NumPy arrays containing face images paired with 68-point or 98-point landmark coordinates. Use [`data/SetPreparation.py`](https://github.com/guoqiangqi/pfld/blob/main/data/SetPreparation.py) to convert raw images and annotation files into the required format, specifying input directories and annotation file paths via command-line arguments. ### How do I prevent overfitting when fine-tuning on a small dataset? Freeze the early layers of the MobileNet-style backbone by setting `--freeze_backbone True` in [`train_model.py`](https://github.com/guoqiangqi/pfld/blob/main/train_model.py), which preserves generic low-level feature extractors. Additionally, lower the learning rate to `1e-5` and utilize data augmentation tools like [`tools/frame_cut.py`](https://github.com/guoqiangqi/pfld/blob/main/tools/frame_cut.py) and [`tools/emotion_compress.py`](https://github.com/guoqiangqi/pfld/blob/main/tools/emotion_compress.py) to artificially expand your training set. ### Where does the PFLD model store training checkpoints and logs? During training, [`train_model.py`](https://github.com/guoqiangqi/pfld/blob/main/train_model.py) writes progress logs to [`data/log0.txt`](https://github.com/guoqiangqi/pfld/blob/main/data/log0.txt) and saves model checkpoints to the directory specified by `--checkpoint_dir`. The best model weights are stored as `.pth` files, which can be directly loaded into [`test_model.py`](https://github.com/guoqiangqi/pfld/blob/main/test_model.py) for evaluation or [`camera.py`](https://github.com/guoqiangqi/pfld/blob/main/camera.py) for real-time inference. ### Can I fine-tune the PFLD model for a different number of landmarks? While the pre-trained weights in [`model2.py`](https://github.com/guoqiangqi/pfld/blob/main/model2.py) are optimized for 68-point or 98-point configurations, you can modify the final output layer in the architecture definition to match your specific landmark count. However, this requires adjusting the model structure before loading pre-trained weights and retraining the modified layers from scratch while potentially freezing the backbone.