kcostya / thousand-facial-landmarks

Thousand Facial Landmarks Competition Within MADE Computer Vision course. Predict 971 facial points for VGG dataset.

Thousand Facial Landmarks Competition Solution within MADE Computer Vision Course

Predict 971 facial points for VGG dataset.

https://www.kaggle.com/c/made-thousand-facial-landmarks

General Approach

• ResNet50 backbone with fully connected layer changed
• 128 px crops and ImageNet normalization
• Warm-up with OneCycleLR policy and SGD
• Further training with CosineAnnealingWarmRestarts and AdamW with adaptive learning rates by layer groups
• Wing Loss from pytorch-toolbelt
• No weight decay for bias and batchnorm ("Bag of Tricks for Image Classification with Convolutional Neural Networks")
• Final solution is simple blend with another ResNet50 model

Config

• CROP_SIZE = 128
• backbone = resnet50
• batch size = 32
• epochs = 5
• SGD_momentum = 0.9
• SGD_weight_decay = 1e-04
• OneCycle_max_lr = 0.1
• WingLoss_width = 10
• WingLoss_curvature = 2
• AdamW_weight_decay = 1e-06
• AdamW_amsgrad = True
• Adaptive learning rates:

[
    {"params": model.conv1.parameters(), "lr": 1e-6},
    {"params": model.bn1.parameters(), "lr": 1e-6},
    {"params": model.relu.parameters(), "lr": 1e-5},
    {"params": model.maxpool.parameters(), "lr": 1e-5},
    {"params": model.layer1.parameters(), "lr": 1e-4},
    {"params": model.layer2.parameters(), "lr": 1e-4},
    {"params": model.layer3.parameters(), "lr": 1e-3},
    {"params": model.layer4.parameters(), "lr": 1e-3},
    {"params": model.avgpool.parameters(), "lr": 1e-2},
    {"params": model.fc.parameters(), "lr": 1e-2},
]

Run

pip install -r requirements.txt

python train.py --name "baseline" --data "PATH_TO_DATA" [--gpu]

Basic blending example is available within blender.ipynb notebook

Results