The kaggle_global_wheat_detection from kiccho1101

Kaggle Profile: https://www.kaggle.com/kiccho11

kaggle_global_wheat_detection's People

Contributors

Stargazers

Watchers

kaggle_global_wheat_detection's Issues

TTA

Try TTA
Reference Kernels
- https://www.kaggle.com/shonenkov/wbf-over-tta-single-model-efficientdet
- https://www.kaggle.com/whurobin/wbf-over-tta-single-model-fasterrcnn-resnest

お勉強

今まで見たことなかった画像系の動画も、このコンペのおかげで何となく分かるようになってきたので、見ていく。

Create CV runner

Create runner.py for cross-validation.
We'd better do it by pair-programming.
Understand competition metrics
Implement calc_metrics function
mlflow
hydra

Try many augmentations

現実の範囲内であり得るAugmentationをやりまくる
↓良さげなNotebook

https://www.kaggle.com/kaushal2896/data-augmentation-tutorial-basic-cutout-mixup

Implement calc_metric

評価関数実装するよ。
https://www.kaggle.com/pestipeti/competition-metric-details-script

mlflow server

ちなみに今GCPでインスタンス借りてそこで回してます
GPUはTeslaのT4 (速いのか遅いのかわからん)

Try pytorch-lightning

https://www.kaggle.com/artgor/object-detection-with-pytorch-lightning
This notebook seems interesting.

OOF

Try OOF
Reference Kernels
- https://www.kaggle.com/shonenkov/oof-evaluation-mixup-efficientdet
- https://www.kaggle.com/hawkey/yolov5-pseudo-labeling-oof-evaluation

Mixup

https://www.kaggle.com/shonenkov/oof-evaluation-mixup-efficientdet
今の実装だとたぶんMixup内蔵されてないので、これ見て実装する

Solve effdet error

effdetのバージョンをアップデートすると、eval時の引数が変更されているのでエラーになる件。
https://www.kaggle.com/angqx95/training-updated-version-efficientdet
このノートブックに対処法が書いてありそう

Debug model forwarding

When running notebook/cv.ipynb, we got a following error.

Fitter prepared. Device is cpu

2020-07-12T22:45:16.055624
LR: 0.0002
Train Step 0/675, summary_loss: 0.00000, time: 0.26056
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-4-42e71780afd2> in <module>
     20 )
     21 
---> 22 fitter.fit(train_loader, valid_loader)

~/Documents/src/github.com/kiccho1101/kaggle_global_wheat_detection/src/factories/fitter.py in fit(self, train_loader, valid_loader)
     59 
     60             start = time.time()
---> 61             summary_loss = self._train_one_epoch(train_loader)
     62 
     63             self.log(

~/Documents/src/github.com/kiccho1101/kaggle_global_wheat_detection/src/factories/fitter.py in _train_one_epoch(self, train_loader)
    109             self.optimizer.zero_grad()
    110 
--> 111             loss, _, _ = self.model(images, bboxes, labels)
    112             loss.backward()
    113             summary_loss.update(loss.detach().item(), batch_size)

~/.local/share/virtualenvs/kaggle_global_wheat_detection-Hxp-F21z/lib/python3.8/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    548             result = self._slow_forward(*input, **kwargs)
    549         else:
--> 550             result = self.forward(*input, **kwargs)
    551         for hook in self._forward_hooks.values():
    552             hook_result = hook(self, input, result)

TypeError: forward() takes 3 positional arguments but 4 were given

Train effdet like yolov5

https://www.kaggle.com/raininbox/train-efficientdet-like-yolo-v5
I don't understand this title, but I'll try to understand it and implement it if possible.

WBF

FasterRCNN Pseudo Labeling

https://www.kaggle.com/nvnnghia/fasterrcnn-pseudo-labeling

Want to understand what pseudo labeling is.

For license issue

ライセンス問題、結果よくわからんのでDiscussionで読んだスクショを貼ってく

Something is wrong with the evaluation function

今のプログラムだと、precisionが0.01という明らかに低すぎる数字になってしまう。
おそらくプログラムにバグありなので、解明して修正しないと。。

Ideas

Backbone

EfficientNet
https://github.com/zhanghang1989/ResNeSt

Pre-trained model

better pre-trained model than ImageNet-1.2M pre-trained model
- Weakly supervised Instagram-1G (tag)
- https://github.com/Cadene/pretrained-models.pytorch

Data Augmentation

CutMix [2] = Mixup + Cutout
- Is this really implemented?
AutoAugment [5] or faster model

How to use test data

"Temporal Ensembling for Semi-Supervised Learning"
training with noisy student [4]
semi-supervised learning for object detection [3] instead of pseudo labeling
- consisntency learning is a basic idea for semi-supervised learning for classification (e.g. Temporal Ensembling, Mean Teacher, MixMatch, FixMatch)