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.

scripts/cam_consistency_training/configs/eff-b2-focal-alpha1-cutmix-pubhpa-maskv2.yaml
scripts/cam_consistency_training/configs/eff-b5-focal-alpha1-cutmix-pubhpa-maskv2.yaml
scripts/cam_consistency_training/configs/eff-b7-focal-alpha1-cutmix-pubhpa-maskv2.yaml
scripts/cam_consistency_training/configs/eff-b2-cutmix-pubhpa-768-to-1536.yaml
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/.
scripts/cam_consistency_training/configs/eff-b2-focal-stage2-b2b2b5b7avg.yaml
scripts/cell_crop/configs/resnest50-bce-from768-cutmix-softpl.yaml
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/.
scripts/cam_consistency_training/configs/eff-b2-focal-stage3.yaml
scripts/cam_consistency_training/configs/eff-b2-focal-stage3-cos.yaml
scripts/cell_crop/configs/resnest50-bce-from768-stage3.yaml
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 ...

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

Related tags

Overview

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

Dataset Preparation

Resize Images

Download Public HPA

Create Cell Mask

Create Input for Cell-level Classifier

Training

Inference

Validation Set

Test Set

Use the code on Kaggle Notebook

Owner

The official implementation of the CVPR2021 paper: Decoupled Dynamic Filter Networks

Deep Learning applied to Integral data analysis

VGGFace2-HQ - A high resolution face dataset for face editing purpose

Info and sample codes for "NTU RGB+D Action Recognition Dataset"

This is an official implementation for the WTW Dataset in "Parsing Table Structures in the Wild " on table detection and table structure recognition.

iNAS: Integral NAS for Device-Aware Salient Object Detection

Collection of generative models in Pytorch version.

Tree LSTM implementation in PyTorch

Chainer Implementation of Semantic Segmentation using Adversarial Networks

Classification models 1D Zoo - Keras and TF.Keras

Official implementation of "GS-WGAN: A Gradient-Sanitized Approach for Learning Differentially Private Generators" (NeurIPS 2020)

The comma.ai Calibration Challenge!

Large-scale Hyperspectral Image Clustering Using Contrastive Learning, CIKM 21 Workshop

Tensorflow implementation of MIRNet for Low-light image enhancement

InterFaceGAN - Interpreting the Latent Space of GANs for Semantic Face Editing

Feedback is important: response-aware feedback mechanism for background based conversation

In this tutorial, you will perform inference across 10 well-known pre-trained object detectors and fine-tune on a custom dataset. Design and train your own object detector.

Link prediction using Multiple Order Local Information (MOLI)

Official PyTorch implementation of Segmenter: Transformer for Semantic Segmentation

This code provides a PyTorch implementation for OTTER (Optimal Transport distillation for Efficient zero-shot Recognition), as described in the paper.