Skip to content

Latest commit

 

History

History
58 lines (49 loc) · 2.85 KB

README.md

File metadata and controls

58 lines (49 loc) · 2.85 KB
Raw

Traffic4Cast-2020-TLab

The 2nd place solution to Neurips 2020 Traffic4Cast competition. Check traffic4cast_2020.pdf for more details about our solution.

How to run our solution.

  1. run tools/process_lmdb.py with three city:
INPUT_DIR='./';
python tools/process_lmdb.py --city BERLIN -i $INPUT_DIR -o ./processed_data/ --test-slots ./processed_data/test_slots.json    -m 50000000000;
python tools/process_lmdb.py --city ISTANBUL -i $INPUT_DIR -o ./processed_data/ --test-slots ./processed_data/test_slots.json  -m 100000000000;
python tools/process_lmdb.py --city MOSCOW -i $INPUT_DIR -o ./processed_data/ --test-slots ./processed_data/test_slots.json    -m 200000000000;

Here -m option decides the max size for lmdb. This value varies for different cities.

  1. run tools/agg_feat_lmdb.py with three cities.
for CITY in BERLIN ISTANBUL MOSCOW;
do
  python tools/agg_feat_lmdb.py --city $CITY;
done

  1. Download pre-trained model here in pre-trained weights directory and put it into processed_data dictionary or set config file PRETRAIN_MODEL: ''

  2. Install apex and other requirement libraries in requirements.txt

  3. run gen.py to generate single model submission. Our best single model use v4-hrnet-w48-geo-embed-include-valid.yaml, and the online score is 1.1761e-3. First download model wegihts here in weights directory and put it into local weights directory and run:

for CITY in BERLIN ISTANBUL MOSCOW;
do
python gen.py --city $CITY \
              --path ./config/v4-hrnet-w48-geo-embed-include-valid.yaml \
              --tag best \
              --test-slots ./processed_data/test_slots.json;
done
  1. or run gen_ensemble.py to run ensemble models with uint8 output. We provide a version with a score of around 1.1667e-3. In fact, only need to ensemble 4 models then we can reach top 2. Our best score needs float output, and some files in Moscow will exceed the 20M limitation. So we need also generate two versions and replace ~4 files in Moscow to reach the best score. First, download model weights here in weights directory and put it into local weights directory, and run:
for CITY in BERLIN ISTANBUL MOSCOW;
do
python gen_ensemble.py -c $CITY -w "4,3,1,3,1,2,1,4" \
--use-all "0,0,0,0,0,0,0,1" \
--tag "best,best,best,best,best,best,best,best" \
--config "./config/v4-hrnet-w48-geo-embed-include-valid.yaml,\
./config/v4-hrnet-w48-include-valid.yaml,\
./config/v4-hrnet-include-valid.yaml,\
./config/v6-hrnet-sun-fix-dst-include-valid.yaml,\
./config/v6-hrnet-3D.yaml,\
./config/v6-unet-include-valid.yaml,\
./config/v4-hrnet-w48-geo-add-0-day-embed-include-valid.yaml,\
./config/v4-hrnet-w48-include-valid-all.yaml" 
done