/malawi_1st_place

Primary LanguageJupyter Notebook

First Place malawi solution

There are 2 parts of the solution - Regression in part_panshin and Object Detection in part_sheoran.

Note that these parts require different cuda versions, drivers, GPU architectures, etc. All the specifications are listed in the corresponding dirs (part_panshin and part_sheoran).

In case of any issues, simply reach out at ivan.panshin@protonmail.com and I will solve all the issues for you.

Reproduce inference

  1. Create submission by Regression model. Look at part_panshin/README.md/Install and part_panshin/README.md/Reproduce inference.
  2. Create submission by Object Detection model. Look at part_sheoran/README.md/Install and part_sheoran/README.md/Reproduce inference
  3. Combine submissions to generate the 1st place sub.
python create_final_sub.py

Submission will be saved at final_subs/first_place_sub.csv. This submission should give the following scores:

  • Public: 0.2611
  • Private: 0.2688

Which is 1st private place.

Reproduce training

  1. Train Regression model. Look at part_panshin/README.md/Install and part_panshin/README.md/Reproduce Training.
  2. Train Object Detection model. Look at part_sheoran/README.md/Install and part_sheoran/README.md/Reproduce Training.

Documentation

  1. Overview and objectives

The main idea is to combine a pure Regression model (that predict number of houses directly) and Object Detection models (that actually predict houses bboxes)

  1. Architecture diagram diagram

  2. ETL process

  • Split data into folds according to Stratification and Grouping. For Regression, Stratification is done based on number of bboxes on images. Grouping is done based on image_ids. For Object Detection, Stratification is done based on empty/non-empty images. Grouping is done based on image_ids.
  • In order to load images, use CV2. In order to augment them, use Albumentations.
  1. Data modeling
  • Regression part uses maxvit_base_tf_512.in21k_ft_in1k from timm library. It's a powerful transformer that is very suitable for counting objects due to its global context.
  • Object detection part uses Co-Dino with Swin-Large from mmdet and Co-Deformable DETR with Swin-Base from mmdet.
  • Both models are optimized using MAE as loss, AdamW as optimizer.
  • Regression model uses CosineAnnealing as scheduler, and Object Detection - StepLR. In both cases we decrease the original LR by 10 times.
  • Regression model is trained for 600 epochs, Object Detection models - 16 and 36.
  • For augmentations we use Flips, Rotate, BrightnessContact and CoarseDropout.
  • To evaluate regression performance, MAE is calculated on validation set. The regression model achieves 0.282 MAE.
  • To evaluate Object Detection performance, mean average prevision at 0.5 is calculated on validation set. The models have roughly 0.9 AP@0.5
  • Regression normalization is done by max value of image format. In this case, divide by 255.
  • Object Detection normalization is done by mean-std based on ImageNet values.
  • Regression uses EMA to increase training metrics stability and bfloat16 to increase numerical stability during training.
  • Fix the seeds at part_panshin/reg_malawi/helpers/model_helpers/seed_everything
  1. Inference
  • To inference the regression model on new data, create a new .csv file similar to data/annotations/test.csv that contains base names of files in data/Images. Then specify path to that .csv in configs/dataset/test/test.csv at ann_path. The predictions will be saved at the path specified at configs/trainer/test.yaml.
  • To inference the object detection model on new data, have a new SampleSubmission.csv that has the same format as competition one, and have all the images in data/test_images.
  • When it comes to re-training, we recommend training from scratch, since it's easier in terms of setting up LRs for each re-training.
  1. Run time

  • The regression training takes about 7 hours on 4 x A6000 Ada.

  • The regression inference takes about 1 minute on 1 x A6000 Ada.

  • The Object Detection training takes about 10 hours on 4 x RTX 3090.

  • The Object Detection inference takes about 20 minutes on 1 x RTX 3090.