Bengali-kaggle
best result
liner5 CV 0.9905 LB 0.9810 (train with cutmix randomshiftrotate and input size is 128x128x1) Liner1 CV 0.9913 LB 0.9810 (train with cutmix randomshiftrotate and input size is 128x128x1)
CMD RUN: python train.py --model senet50 --outdir YOUR_OUT_DIR --gpu_ids 2,3 --width 128 --height 128 --feather_data_path BengaliData/feather_resize128/ --mixup 1 --image_mode gray --patience 3 --LR_SCHEDULER REDUCED --optimizer RADAM --image_mode gray --lr 1e-3 --lr_ratio 0.9 --batch_size 512
model: which model to use outdir: model and log save dir gpu_ids: which gpu will use, gpu index start from 0 width: input image width height: input image height feather_data_path: data location for train and val, generate by offline with parquet2feather in data.py mixup: use cutmix or not image_mode: input image mode rgb or gray patience: ReduceLROnPlateau patience LR_SCHEDULER : which one schedular will use optimizer: which optimizer will use lr : learning rate lr_ratio: ReduceLROnPlateau factor batch_size: batch_size
After i check our submission in kaggle, i find the liner5 is the best model now
liner5 head + cutmix + rotate inputsize 128x128x1 global_max_recall CV 0.9905 LB 0.9810 liner1 head + cutmix + rotate inputsize 128x128x1 global_max_recall CV 0.9913 LB 0.9810
So the liner5's gap between CV an LB is small, so it should be better model