7th place solution of Human Protein Atlas - Single Cell Classification on Kaggle

Related tags

Deep Learningdeep-learningkagglekaggle-competitionhpa-challenge-2021
Overview

kaggle-hpa-2021-7th-place-solution

Code for 7th place solution of Human Protein Atlas - Single Cell Classification on Kaggle.

A description of the method can be found in this post in the kaggle discussion.

Dataset Preparation

Resize Images

# Resize train images to 768x768
python scripts/hap_segmenter/create_cell_mask.py resize_image \
    --input_directory data/input/hpa-single-cell-image-classification.zip/train \
    --output_directory data/input/hpa-768768.zip \
    --image_size 768
# Resize train images to 1536x1536
python scripts/hap_segmenter/create_cell_mask.py resize_image \
    --input_directory data/input/hpa-single-cell-image-classification.zip/train \
    --output_directory data/input/hpa-1536.zip \
    --image_size 1536

# Resize test images to 768x768
python scripts/hpa_segmenter/create_cell_mask.py resize_image \
    --input_directory /kaggle/input/hpa-single-cell-image-classification/test \
    --output_directory data/input/hpa-768-test.zip \
    --image_size 768
# Resize test images to 1536x1536
python scripts/hpa_segmenter/create_cell_mask.py resize_image \
    --input_directory /kaggle/input/hpa-single-cell-image-classification/test \
    --output_directory data/input/hpa-1536-test.zip \
    --image_size 1536

You can specify a directory in a zip file in the same way as a normal directory.

Download Public HPA

Download all images in kaggle_2021.tsv in this dataset, resize them into 768x768 and 1536x1536, and archive them as data/input/hpa-public-768.zip and data/input/hpa-public-1536.zip.

Create Cell Mask

# Create cell masks for the Kaggle train set with 1536x1536
python scripts/hpa_segmenter/create_cell_mask.py create_cell_mask \
    --input_directory data/input/hpa-1536.zip \
    --output_directory data/input/hpa-1536-mask-v2.zip \
    --label_cell_scale_factor 1.0

# Resize the masks to 768x768
python scripts/hpa_segmenter/create_cell_mask.py resize_cell_mask \
    --input_directory data/input/hpa-1536-mask-v2.zip \
    --output_directory data/input/hpa-768-mask-v2-from-1536.zip \
    --image_size 768

# Create cell masks for the Public HPA dataset with 1536x1536
python scripts/hpa_segmenter/create_cell_mask.py create_cell_mask \
    --input_directory data/input/hpa-public-1536.zip/hpa-public-1536 \
    --output_directory data/input/hpa-public-1536-mask-v2.zip \
    --label_cell_scale_factor 1.0

# Resize the masks to 768x768
python scripts/hpa_segmenter/create_cell_mask.py resize_cell_mask \
    --input_directory data/input/hpa-public-1536-mask-v2.zip \
    --output_directory data/input/hpa-public-768-mask-v2-from-1536.zip \
    --image_size 768

# Create cell masks for the test set with the original resolution
# Run with `--label_cell_scale_factor = 0.5` to save inference time
python scripts/hpa_segmenter/create_cell_mask.py create_cell_mask \
    --input_directory /kaggle/input/hpa-single-cell-image-classification/test \
    --output_directory data/input/hpa-test-mask-v2.zip \
    --label_cell_scale_factor 0.5

# Resize the masks to 1536x1536
python scripts/hpa_segmenter/create_cell_mask.py resize_cell_mask \
    --input_directory data/input/hpa-test-mask-v2.zip \
    --output_directory data/input/hpa-test-mask-v2-1536.zip \
    --image_size 1536

# Resize the masks to 768x768
python scripts/hpa_segmenter/create_cell_mask.py resize_cell_mask \
    --input_directory data/input/hpa-test-mask-v2.zip \
    --output_directory data/input/hpa-test-mask-v2-768.zip \
    --image_size 768

Create Input for Cell-level Classifier

# Create cell-level inputs for the Kaggle train set using 768x768 images as fixed scale image.
python scripts/hap_segmenter/create_cell_mask.py crop_and_resize_cell \
    --image_directory data/input/hpa-768768.zip \
    --cell_mask_directory data/input/hpa-768-mask-v2-from-1536.zip \
    --output_directory data/input/hpa-cell-crop-v2-192-from-768.zip \
    --image_size 192

# Create cell-level inputs for the Public HPA dataset using 768x768 images as fixed scale image.
python scripts/hap_segmenter/create_cell_mask.py crop_and_resize_cell \
    --image_directory data/input/hpa-public-768.zip \
    --cell_mask_directory data/input/hpa-public-768-mask-v2-from-1536.zip \
    --output_directory data/input/hpa-public-cell-crop-v2-192-from-768.zip \
    --image_size 192

# Create cell-level inputs for the Kaggle train set using 1536x1536 images as fixed scale image.
python scripts/hap_segmenter/create_cell_mask.py crop_and_resize_cell \
    --image_directory data/input/hpa-1536.zip \
    --cell_mask_directory data/input/hpa-1536-mask-v2.zip \
    --output_directory data/input/hpa-cell-crop-v2-192-from-1536.zip \
    --image_size 192

# Create cell-level inputs for the Public HPA dataset using 1536x1536 images as fixed scale image.
python scripts/hap_segmenter/create_cell_mask.py crop_and_resize_cell \
    --image_directory data/input/hpa-public-1536.zip \
    --cell_mask_directory data/input/hpa-public-1536-mask-v2.zip \
    --output_directory data/input/hpa-public-cell-crop-v2-192-from-1536.zip \
    --image_size 192

# Create cell-level inputs for the test set using 768x768 images as fixed scale image.
python scripts/hpa_segmenter/create_cell_mask.py crop_and_resize_cell \
    --image_directory data/input/hpa-768768-test.zip \
    --cell_mask_directory data/input/hpa-test-mask-v2-768.zip \
    --output_directory data/input/hpa-test-cell-crop-v2-192-from-768.zip \
    --image_size 192

# Create cell-level inputs for the test set using 1536x1536 images as fixed scale image.
python scripts/hpa_segmenter/create_cell_mask.py crop_and_resize_cell \
    --image_directory data/input/hpa-1536-test.zip \
    --cell_mask_directory data/input/hpa-test-mask-v2-1536.zip \
    --output_directory data/input/hpa-test-cell-crop-v2-192-from-1536.zip \
    --image_size 192

Training

# Train image-level classifier
python scripts/cam_consistency_training/run.py train \
    --config_path scripts/cam_consistency_training/configs/${CONFIG_NAME}.yaml

# Train cell-level classifier
python scripts/cell_crop/run.py train \
    --config_path scripts/cell_crop/configs/${CONFIG_NAME}.yaml

If you want to train on multiple GPUs, use a launcher like torch.distributed.launch and pass --local_rank option. You can override the fields in the config by passing an argument like field_name=${value} (e.g. fold_index=1). We trained 5 folds for all models used in the final submission pipeline. The config files are located in scripts/cam_consistency_training/configs and scripts/cell_crop/configs. We trained the models in the following order.

  1. scripts/cam_consistency_training/configs/eff-b2-focal-alpha1-cutmix-pubhpa-maskv2.yaml
  2. scripts/cam_consistency_training/configs/eff-b5-focal-alpha1-cutmix-pubhpa-maskv2.yaml
  3. scripts/cam_consistency_training/configs/eff-b7-focal-alpha1-cutmix-pubhpa-maskv2.yaml
  4. scripts/cam_consistency_training/configs/eff-b2-cutmix-pubhpa-768-to-1536.yaml
  5. Do predict_valid and concat_valid_predictions (described below) for each model and save the average of the output files under data/working/consistency_training/b2-1536-b2-b5-b7-768-avg/.
  6. scripts/cam_consistency_training/configs/eff-b2-focal-stage2-b2b2b5b7avg.yaml
  7. scripts/cell_crop/configs/resnest50-bce-from768-cutmix-softpl.yaml
  8. Do predict_valid and concat_valid_predictions for each model and save the average of the output files under data/working/image-level-and-cell-crop-both-5folds/.
  9. scripts/cam_consistency_training/configs/eff-b2-focal-stage3.yaml
  10. scripts/cam_consistency_training/configs/eff-b2-focal-stage3-cos.yaml
  11. scripts/cell_crop/configs/resnest50-bce-from768-stage3.yaml
  12. scripts/cell_crop/configs/resnest50-bce-from1536-stage3-cos.yaml

Inference

Validation Set

# Image-level classifier inference
python scripts/cam_consistency_training/run.py predict_valid \
    --config_path scripts/cam_consistency_training/configs/${CONFIG_NAME}.yaml

# Cell-level classifier inference
python scripts/cell_crop/run.py predict_valid \
    --config_path scripts/cell_crop/configs/${CONFIG_NAME}.yaml

# Concatenate the predictions for each fold to obtain the OOF prediction for the entire training data
python scripts/cam_consistency_training/run.py concat_valid_predictions \
    --config_path scripts/cam_consistency_training/configs/${CONFIG_NAME}.yaml
python scripts/cell_crop/run.py concat_valid_predictions \
    --config_path scripts/cell_crop/configs/${CONFIG_NAME}.yaml

Test Set

# Image-level classifier inference
python scripts/cam_consistency_training/run.py predict_test \
    --config_path scripts/cam_consistency_training/configs/${CONFIG_NAME}.yaml

# Cell-level classifier inference
python scripts/cell_crop/run.py predict_test \
    --config_path scripts/cell_crop/configs/${CONFIG_NAME}.yaml

# Make our final submission with post-processing
python scripts/average_predictions.py \
    --orig_size_cell_mask_directory data/input/hpa-test-mask-v2.zip \
    "data/working/consistency_training/eff-b2-focal-stage3/0" \
    "data/working/consistency_training/eff-b2-focal-stage3/1" \
    "data/working/consistency_training/eff-b2-focal-stage3/2" \
    "data/working/consistency_training/eff-b2-focal-stage3/3" \
    "data/working/consistency_training/eff-b2-focal-stage3/4" \
    "data/working/consistency_training/eff-b2-focal-stage3-cos/0" \
    "data/working/consistency_training/eff-b2-focal-stage3-cos/1" \
    "data/working/consistency_training/eff-b2-focal-stage3-cos/2" \
    "data/working/consistency_training/eff-b2-focal-stage3-cos/3" \
    "data/working/consistency_training/eff-b2-focal-stage3-cos/4" \
    "data/working/cell_crop/resnest50-bce-from768-stage3/0" \
    "data/working/cell_crop/resnest50-bce-from768-stage3/1" \
    "data/working/cell_crop/resnest50-bce-from768-stage3/2" \
    "data/working/cell_crop/resnest50-bce-from768-stage3/3" \
    "data/working/cell_crop/resnest50-bce-from768-stage3/4" \
    "data/working/cell_crop/resnest50-bce-from1536-stage3-cos/0" \
    "data/working/cell_crop/resnest50-bce-from1536-stage3-cos/1" \
    "data/working/cell_crop/resnest50-bce-from1536-stage3-cos/2" \
    "data/working/cell_crop/resnest50-bce-from1536-stage3-cos/3" \
    "data/working/cell_crop/resnest50-bce-from1536-stage3-cos/4" \
    --edge_area_threshold 80000 --center_area_threshold 32000

Use the code on Kaggle Notebook

Use docker to zip the source code and the wheels of the dependencies and upload them as a dataset.

docker run --rm -it -v /path/to/this/repo:/tmp/workspace -w /tmp/workspace/ gcr.io/kaggle-images/python bash ./build_zip.sh

In Kaggle Notebook, when you copy the code as shown below, you can run it the same way as your local environment.

# Make a working directory
!mkdir -p /kaggle/tmp

# Change the current directory
cd /kaggle/tmp

# Copy source code from the uploaded dataset
!cp -r /kaggle/input/<your-dataset-name>/* .

# You can use it as well as local environment
!python scripts/hpa_segmenter/create_cell_mask.py create_cell_mask ...
Owner
GitHub Repository
Council-GAN - Implementation for our paper Breaking the Cycle - Colleagues are all you need (CVPR 2020)

Council-GAN Implementation of our paper Breaking the Cycle - Colleagues are all you need (CVPR 2020) Paper Ori Nizan , Ayellet Tal, Breaking the Cycle

ori nizan 260 Nov 16, 2022
Implementation of TimeSformer, a pure attention-based solution for video classification

TimeSformer - Pytorch Implementation of TimeSformer, a pure and simple attention-based solution for reaching SOTA on video classification.

Phil Wang 602 Jan 03, 2023
Tianshou - An elegant PyTorch deep reinforcement learning library.

Tianshou (天授) is a reinforcement learning platform based on pure PyTorch. Unlike existing reinforcement learning libraries, which are mainly based on

Tsinghua Machine Learning Group 5.5k Jan 05, 2023
Constructing interpretable quadratic accuracy predictors to serve as an objective function for an IQCQP problem that represents NAS under latency constraints and solve it with efficient algorithms.

IQNAS: Interpretable Integer Quadratic programming Neural Architecture Search Realistic use of neural networks often requires adhering to multiple con

0 Oct 24, 2021
Visual Adversarial Imitation Learning using Variational Models (VMAIL)

Visual Adversarial Imitation Learning using Variational Models (VMAIL) This is the official implementation of the NeurIPS 2021 paper. Project website

14 Nov 18, 2022
A curated list of awesome resources related to Semantic Search🔎 and Semantic Similarity tasks.

A curated list of awesome resources related to Semantic Search🔎 and Semantic Similarity tasks.

224 Jan 04, 2023
Example repository for custom C++/CUDA operators for TorchScript

Custom TorchScript Operators Example This repository contains examples for writing, compiling and using custom TorchScript operators. See here for the

106 Dec 14, 2022
Reinforcement Learning Theory Book (rus)

Reinforcement Learning Theory Book (rus)

qbrick 206 Nov 27, 2022
A compendium of useful, interesting, inspirational usage of pandas functions, each example will be an ipynb file

Pandas_by_examples A compendium of useful/interesting/inspirational usage of pandas functions, each example will be an ipynb file What is this reposit

Guangyuan(Frank) Li 32 Nov 20, 2022
A toolkit for Lagrangian-based constrained optimization in Pytorch

Cooper About Cooper is a toolkit for Lagrangian-based constrained optimization in Pytorch. This library aims to encourage and facilitate the study of

Cooper 34 Jan 01, 2023
Project NII pytorch scripts

project-NII-pytorch-scripts By Xin Wang, National Institute of Informatics, since 2021 I am a new pytorch user. If you have any suggestions or questio

Yamagishi and Echizen Laboratories, National Institute of Informatics 184 Dec 23, 2022
A python library for face detection and features extraction based on mediapipe library

FaceAnalyzer A python library for face detection and features extraction based on mediapipe library Introduction FaceAnalyzer is a library based on me

Saifeddine ALOUI 14 Dec 30, 2022
[AAAI22] Reliable Propagation-Correction Modulation for Video Object Segmentation

Reliable Propagation-Correction Modulation for Video Object Segmentation (AAAI22) Preview version paper of this work is available at: https://arxiv.or

Xiaohao Xu 70 Dec 04, 2022
A little software to generate and save Julia or Mandelbrot's Fractals.

Julia-Mandelbrot-s-Fractals A little software to generate and save Julia or Mandelbrot's Fractals. Dependencies : Python 3.7 or more. (Also possible t

Olivier 0 Jul 09, 2022
Commonality in Natural Images Rescues GANs: Pretraining GANs with Generic and Privacy-free Synthetic Data - Official PyTorch Implementation (CVPR 2022)

Commonality in Natural Images Rescues GANs: Pretraining GANs with Generic and Privacy-free Synthetic Data (CVPR 2022) Potentials of primitive shapes f

31 Sep 27, 2022
Visual Question Answering in Pytorch

Visual Question Answering in pytorch /!\ New version of pytorch for VQA available here: https://github.com/Cadene/block.bootstrap.pytorch This repo wa

Remi 672 Jan 01, 2023
TensorFlow Tutorials with YouTube Videos

TensorFlow Tutorials Original repository on GitHub Original author is Magnus Erik Hvass Pedersen Introduction These tutorials are intended for beginne

9.1k Jan 02, 2023
DeRF: Decomposed Radiance Fields

DeRF: Decomposed Radiance Fields Daniel Rebain, Wei Jiang, Soroosh Yazdani, Ke Li, Kwang Moo Yi, Andrea Tagliasacchi Links Paper Project Page Abstract

UBC Computer Vision Group 24 Dec 02, 2022
OneFlow is a performance-centered and open-source deep learning framework.

OneFlow OneFlow is a performance-centered and open-source deep learning framework. Latest News Version 0.5.0 is out! First class support for eager exe

OneFlow 4.2k Jan 07, 2023
Hyperopt for solving CIFAR-100 with a convolutional neural network (CNN) built with Keras and TensorFlow, GPU backend

Hyperopt for solving CIFAR-100 with a convolutional neural network (CNN) built with Keras and TensorFlow, GPU backend This project acts as both a tuto

Guillaume Chevalier 103 Jul 22, 2022