date : 2022-11-16
Using Kaggle
- Tidy Kaggle competitions that are similar to current task you need to solve.
- Organize appliable ideas from many informations like Code and Discussions from corresponding competitions.
For example, when doing object detection task for classifying recycling items, NFL Health & Safety-Helmet Assignment, VinBigData Chest X-ray Abnormalities Detection, and TensorFlow-Help Protect the Great Barrier Reef competitions can be useful. Also, writings from Code and Discussions section of the corresponding competitions which have lots of votes can be useful.
Discussions cases
- References
- Articles, Papers, …
- Background
- Previous competition
- Competitions
- Top solutions
- Best cv(cross validation) scores
- Outlines
- Top solution
- Closed competition
- Idea
- CV strategy
Code cases
- EDA
- Sort by most votes
- Baseline
- Sort by most votes
Use EDA info when preprocessing
Use EDA info when applying augmentation
ex)
When we need to preserve as many info about original image:
Flip, Rotate90, Cutout, Cutmix … can be used.
ColorJitter, Interpolation(Resize, Rotate30, …), Noise, … cannot be used.
Applying EDA
- Pre-processing
- Augmentations
- CV Strategy
- Custom Model
- Custom Loss
- Post-processing
- …
Other Tips
- Most important parts
CV Strategy, CV LB Correlation
-
Data that can be used to training stably
SWA(Stochastic Weight Averaging), More Ensemble, More TTA …
-
Scheduler
For image data, scheduler that LR decreases by every epoch.
At the beginning, mainly use ReduceLR or Cycle Scheduler to find appropriate epochs and learning rate.
At last, use CosineWarmup to raise the score until the very end.
-
Cooperating method
share mainly ideas. build baseline code for each. -> maximize the ensemble effects
-
When score does not enhance
-
Go back to EDA
-
OOF analysis
identify if there exists common feature of the inputs that model does not predict correctly.
If this suceed, it can be a key to enhance the score!
-