Awesome
Traffic4Cast2021-SwinUNet3D (AI4EX Team)
Table of Content
General Info
This resipository contains our code submitted to Traffic4cast2021 competition (https://www.iarai.ac.at/traffic4cast/2021-competition/challenge/#challenge) This work is made available under the attached license
Requirements
This resipository depends on the following packages availability
- Pytorch Lightning
- timm
- torch_optimizer
- pytorch_model_summary
- einops
Installation:
unzip folder.zip
cd folder
conda create --name swinencoder_env python=3.6
conda activate swinunet3d_env
conda install pytorch=1.9.0 cudatoolkit=10.2 -c pytorch
pip install -r requirements.txt
Usage
- a.1)train from scratch (together with inference predictions)
python Traffic4Cast2021/main1.py --nodes 1 --gpus 4 --precision 16 --batch-size 5 --epochs 100 --mlp_ratio 1 --stages 4 --patch_size 4 --dropout 0.0 --start_filters 192 --sampling-step 1 --decode_depth 1 --use_neck --lr 1e-4 --optimizer lamb --merge_type both --mix_features --city_category TEMPORAL --memory_efficient - a.2) fine tune a model from a checkpoint
python main.py --gpus 1 --city_category TEMPORAL --mode train --name TEMPORAL_real_swinunet3d_141848694 --time-code 20210913T135845 --initial-epoch 36``` - b) evaluate a trained model from a checkpoint (submitted inference)
python main.py --gpus 1 --city_category TEMPORAL --mode test --name TEMPORAL_real_swinunet3d_141848694 --time-code 20210913T135845 --initial-epoch 36
Inference
- a) To generate predictions using our trained model
python main.py --gpus 1 --city_category TEMPORAL --mode test --name TEMPORAL_real_swinunet3d_141848694 --time-code 20210913T135845 --initial-epoch 36
- b) To create submission in form of a zipped file from files generater in (a)
python create_submission.py --name TEMPORAL_real_swinunet3d_141848694 --time-code 20210913T135845 --epoch 36
Accessing the trained checkpoint
Our trained model can be downloaded from https://drive.google.com/file/d/10zM-oiEjRD1rDlDw1bnx06Dl8Z3K3tNQ/view?usp=sharing