ed2k/comma10k-baseline

A baseline segmentation example using the comma10k dataset (WIP)

🚗 comma10k-baseline

A semantic segmentation baseline using @comma.ai's comma10k dataset.

Using U-Net with efficientnet encoder, this baseline reaches 0.045 validation loss.

Visualize

Here is an example (randomly from the validation set, no cherry picking)

Ground truth

Predicted

How to use

This baseline uses two stages (i) 448x576 (ii) 896x1184 (close to full resolution)

python3 train_lit_model.py --backbone efficientnet-b4 --version first-stage --gpus 2 --batch-size 28 --epochs 100 --height 448 --width 576
python3 train_lit_model.py --backbone efficientnet-b4 --version second-stage --gpus 2 --batch-size 7 --learning-rate 5e-5 --epochs 30 --height 896 --width 1184 --seed-from-checkpoint .../efficientnet-b4/first-stage/checkpoints/last.ckpt

WIP

• Update to pytorch lightning 1.0
• Use A.PadIfNeeded in the second stage instead of Resize
• Try more image augmentations

Dependecies

Python 3.5+, pytorch 1.6+ and dependencies listed in requirements.txt.