Boostcamp First CV Competition Abstract

Classify 18 agumented classes
- 3 age intervals
- mask, no mask, and incorrect mask
- male or female

File name and answer label in csv
How to create the answer label file?
- With python IDLE (the hosts will provide it)
- Jupyter Notebook

Many data fields.. aside from the image itself. how can I use it? -> The goal is to predict all labels

(Done) The faces in the photo are always in the middle. We can crop it for better performance
(Failed due to too much time) Better to have the result for each classes from different models or at least from a different vector. The suggested class of 0 from 17 with different classes merged into one vector is undesirable
(Works Great) Stratified Dataset according to personals. Same person, many photo -> could be in both the train and test dataset -> could make the validation set less meaningful
(Fun, easy to share) Transformation Cafe. Sharing sorts of transformations as cafe menus
(Stupid that we weren’t doing it) Validation and train set split
(Seemingly useful for now) learning rate scheduler -> CosineAnnealingLR
(Great) Tidying up is so important. Name the model weights and tensorboard running log with time and other recognizable variables
(Trying out) Shouldn’t the transformation on validation be the same one with the one on test data?
(Very Useful) Any errors from the dataset or data loader? Any wrong labels from the first place? -> YES

Imbalanced dataset problem
- Use focal loss
- Use data augumentation
Resize
- original size of dataset 500x384
- 400x384 -> no face crop
- 384x384 -> very little face crop, still decent
Model Form
- destination class: 3 X
- 3 heads 1 model X
- 1 head 3 models X
- For debugging and fast training purposes -> 1 model 1 head!

You may also enjoy