Awesome

Kaggle-TGS

PyTorch implementation for Kaggle TGS Salt Identification Challenge

15th / 3291 - Team with atom1231

I only share code for my best model in this repository.

The scores of my best 10-fold model by training from scratch: 0.887 on private LB / 0.872 on public LB.

Model Design

A small U-shape [1] model (~8MB) with

• Pyramid pooling module, (dilated) residual blocks and auxiliary training [2]
• Spatial and channel squeeze & excitation module [3]
• Guided upsampling module [4]

Training

• Input size: 129x129 with reflect padding
• Batch size: 40
• Augmentation
  • Random horizontally flip
  • Random translate with reflect padding
  • Random crop and resize
• Optimizer: Adam
• Weight decay: 1e-4
• Initial learning rate: 1e-3
• Total loss function
  • 1.00 Cross-entropy loss
  • 1.00 Lovász-softmax loss [5]
  • 0.05 Topology-aware loss [6]
• 10-fold CV (random split)

Final prediction

• Mean of the probability maps from all 10-fold models (TTA with horizontal flip)
• Threshold of salt probability: 0.5

Progress

• U-shape network with pyramid pooling module, (dilated) residual blocks and auxiliary training by cross-entropy and Lovász-softmax loss
  • Step LR decay by 0.5 every 50 epochs for first 200 epochs
  • 37s/epoch on single 1080TI
  • 0.870 on private LB / 0.852 on public LB
• +Topology-aware loss
  • Cosine annealing with restarts LR for 400 epochs (8 cycles)
  • 50s/epoch on single 1080TI
  • 0.877 on private LB / 0.859 on public LB
• +Spatial and channel squeeze & excitation module
  • Cosine annealing with restarts LR for 1400 epochs (14 cycles)
  • 54s/epoch on single 1080TI
  • 0.886 on private LB / 0.871 on public LB
• +Guided upsampling module
  • Cosine annealing with restarts LR for 200 epochs (1 cycle)
  • 76s/epoch on single 1080TI
  • 0.887 on private LB / 0.872 on public LB

Requirements

• pytorch 0.4.1
• torchvision 0.2.0
• numpy
• opencv
• tqdm
• pandas

pip install -r requirements.txt

Usage

Data

• Download data from Kaggle TGS competition
• Extract train/test.zip, duplicate train directory for val (validation) directory
• Modify the path appropriately in config.json

To train the model

k=10
for((i=1;i<=$k;i=i+1))
do
python train.py --arch pspnet --dataset tgs \
                --img_rows 101 --img_cols 101 --r_pad 14 \
                --n_epoch ${n_epoch} --batch_size 40 --seed 1234 \
                --l_rate 1e-3 --feature_scale 2 --weight_decay 1e-4 \
                --num_k_split $i --max_k_split $k --num_cycles ${num_cycles} --lambda_top 5e-2
done

python train.py -h for more details

To test the model

k=10
for((i=1;i<=$k;i=i+1))
do
python test.py --model_path checkpoints/pspnet_tgs_${n_epoch}_$i-$k_model.pth --dataset tgs \
               --img_rows 101 --img_cols 101 --r_pad 14 --seed 1234 \
               --batch_size 1 --feature_scale 2 --split test --num_k_split $i --max_k_split $k --pred_thr 0.5
done

python test.py -h for more details

To create final submission

python merge.py --dataset tgs --img_rows 101 --img_cols 101 \
                --batch_size 1 --feature_scale 2 --split test --max_k_split 10 --pred_thr 0.5

python merge.py -h for more details

Reference

[1] U-Net: Convolutional Networks for Biomedical Image Segmentation (U-Net)

[2] Pyramid Scene Parsing Network (PSPNet)

[3] Concurrent Spatial and Channel Squeeze & Excitation in Fully Convolutional Networks (scSE)

[4] Guided Upsampling Network for Real-Time Semantic Segmentation (GUN)

[5] The Lovász-Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks

[6] Beyond the Pixel-Wise Loss for Topology-Aware Delineation

[7] Part of code adapted from meetshah1995/pytorch-semseg

[8] Official implementation for Lovász-Softmax loss