kaggle-hpa

Code for 46th place solution in Kaggle Human Protein Atlas - Single Cell Classification

Read this in other languages: English, 한국어

Summary

"Simple Image Level Multilabel Classifier"

The label was determined by applying a classifier to the single cell mask obtained by HPA-Cell-Segmentation.

Classifier was trained using the full dataset.

Tools

• Colab Pro, GCE, Tesla V100 16GB single GPU
• GCS
• Pytorch Lightning
• Neptune
• Kaggle API

Dataset

I used both the Competitions default dataset and the extra dataset.

HPA 512 PNG Dataset by @phalanx

HPA 768 PNG Dataset by @phalanx

HPA 1024 PNG Dataset

HPA Public Data 768x768 "rare classes" dataset by @Alexander Riedel

The extra dataset was downloaded by referring to the public note. Images saved to 768px png. The size is approximately 200 GB.

HPA public data download and HPACellSeg

Validation

MultilabelStratifiedKFold, 5-fold split was used.

The performance of Multilabel Classifier was verified with Macro-F1, Micro-F1 Score.

iterative-stratification

Model training

3-channel RGB images

The image size is 1024px, and trained with the following dataset.

• 1024px Competition default dataset + 768px rare classes dataset(1024 resized)
• 1024px Competition default dataset + 768px extra dataset(1024 resized)
• AdamW
• CosineAnnealingLR
• epochs = 5 for full, 10 for rare

bce, focal loss was used.

model	dataset	folds	loss	batch_size	init_lr	weight_decay	macro F1	micro F1	public LB	private LB
efficientnet_b0	full	2 of 5	bce	16	6.0e-4	1.0e-5	0.7663	0.8171	0.454	0.429
efficientnet_b0	rare classes	single	bce	16	6.0e-4	1.0e-5	0.8154	0.8368	0.394	0.360
seresnext26d_32x4d	full	single	alpha=0.75, gamma=0.0	14	6.5e-5	1.0e-5	0.7317	0.7956	0.381	0.335
final ensemble									0.471	0.433

Segmentation

HPA-Cell-Segmentation was used, and the speed was improved by referring to @deoxy's notebook.

The input image was resized by 1/4, and the CellSegmentator scale_factor=1.0.

The related values of the label_cell function have been adjusted to 1/4.

HPA-Cell-Segmentation

Faster HPA Cell Segmentation by @deoxy

Augmentation

A.Compose(
    [
        A.Resize(height=resize_height, width=resize_width),
        A.RandomScale(scale_limit=(-0.2, 0.2), p=1.0),
        A.PadIfNeeded(
            min_height=resize_height,
            min_width=resize_width,
            border_mode=cv2.BORDER_CONSTANT,
            value=0,
            p=1.0,
        ),
        A.RandomCrop(height=resize_height, width=resize_width, p=1.0),
        A.RandomBrightnessContrast(p=0.8),
        A.HorizontalFlip(p=0.5),
        A.VerticalFlip(p=0.5),
        A.Rotate(border_mode=cv2.BORDER_CONSTANT, value=0, p=0.5),
        A.Normalize(mean=norm_mean, std=norm_std),
        ToTensorV2(),
    ]
)

TTA 4x

HorizontalFlip, VerticalFlip, Resize 0.8, Resize 1.2

What did not work

• Label Smoothing

• pos/neg balanced weighted loss

  X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, and R. M. Summers. (Dec. 2017). "ChestX-ray8: Hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases.", (p. 5) https://arxiv.org/abs/1705.02315