3rd Place Solution of Traffic4Cast 2021 Core Challenge

This is the code for our solution to the NeurIPS 2021 Traffic4Cast Core Challenge.

Paper

Our solution is described in the "Solving Traffic4Cast Competition with U-Net and Temporal Domain Adaptation" paper.

If you wish to cite this code, please do it as follows:

@misc{konyakhin2021solving,
      title={Solving Traffic4Cast Competition with U-Net and Temporal Domain Adaptation}, 
      author={Vsevolod Konyakhin and Nina Lukashina and Aleksei Shpilman},
      year={2021},
      eprint={2111.03421},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Competition and Demonstration Track @ NeurIPS 2021

Learnt parameters

The models' learnt parameters are available by the link: https://drive.google.com/file/d/1zD0CecX4P3v5ugxaHO2CQW9oX7_D4BCa/view?usp=sharing
Please download the archive and unzip it into the weights folder of the repository, so its structure looks like the following:

├── ...
├── traffic4cast
├── weights
│   ├── densenet                 
│   │   ├── BERLIN_1008_1430_densenet_unet_mse_best_val_loss_2019=78.4303.pth                     
│   │   ├── CHICAGO_1010_1730_densenet_unet_mse_best_val_loss_2019=41.1579.pth
│   │   └── MELBOURNE_1009_1619_densenet_unet_mse_best_val_loss_2019=25.7395.pth    
│   ├── effnetb5
│   │   ├── BERLIN_1008_1430_efficientnetb5_unet_mse_best_val_loss_2019=80.3510.pth    
│   │   ├── CHICAGO_1012_1035_efficientnetb5_unet_mse_best_val_loss_2019=41.6425.pth
│   │   ├── ISTANBUL_1012_2315_efficientnetb5_unet_mse_best_val_loss_2019=55.7918.pth    
│   │   └── MELBOURNE_1010_0058_efficientnetb5_unet_mse_best_val_loss_2019=26.0132.pth    
│   └── unet
│       ├── BERLIN_0806_1425_vanilla_unet_mse_best_val_loss_2019=0.0000_v5.pth    
│       ├── CHICAGO_0805_0038_vanilla_unet_mse_best_val_loss_2019=42.6634.pth
│       ├── ISTANBUL_0805_2317_vanilla_unet_mse_best_val_loss_2019=0.0000_v4.pth
│       └── MELBOURNE_0804_1942_vanilla_unet_mse_best_val_loss_2019=26.7588.pth
├── ...

Submission reproduction

To generate the submission file, please run the following script:

# $1 - absolute path to the dataset, $2 device to run inference
sh submission.sh {absolute path to dataset} {cpu, cuda}
# Launch example
sh submission.sh /root/data/traffic4cast cuda

The above sctipt generates the submission file submission/submission_all_unets_da_none_mpcpm1_mean_temporal_{date}.zip, which gave us the best MSE of 49.379068541527 on the final leaderboard.

3rd Place Solution of the Traffic4Cast Core Challenge @ NeurIPS 2021

Related tags

Overview

3rd Place Solution of Traffic4Cast 2021 Core Challenge

Paper

Learnt parameters

Submission reproduction

Owner

Official code for: A Probabilistic Hard Attention Model For Sequentially Observed Scenes

StyleSwin: Transformer-based GAN for High-resolution Image Generation

A reimplementation of DCGAN in PyTorch

Tensorflow implementation of "Learning Deep Features for Discriminative Localization"

Implementation for "Seamless Manga Inpainting with Semantics Awareness" (SIGGRAPH 2021 issue)

Functional deep learning

An Agnostic Computer Vision Framework - Pluggable to any Training Library: Fastai, Pytorch-Lightning with more to come

Code for paper: Group-CAM: Group Score-Weighted Visual Explanations for Deep Convolutional Networks

A static analysis library for computing graph representations of Python programs suitable for use with graph neural networks.

Self-Supervised Deep Blind Video Super-Resolution

PyTorch implementations of the paper: "Learning Independent Instance Maps for Crowd Localization"

Repo for EMNLP 2021 paper "Beyond Preserved Accuracy: Evaluating Loyalty and Robustness of BERT Compression"

Template repository for managing machine learning research projects built with PyTorch-Lightning

GB-CosFace: Rethinking Softmax-based Face Recognition from the Perspective of Open Set Classification

Fuzzing JavaScript Engines with Aspect-preserving Mutation

Utility tools for the "Divide and Remaster" dataset, introduced as part of the Cocktail Fork problem paper

An ever-growing playground of notebooks showcasing CLIP's impressive zero-shot capabilities.

Simplified interface for TensorFlow (mimicking Scikit Learn) for Deep Learning

The codes I made while I practiced various TensorFlow examples

This is the face keypoint train code of project face-detection-project