AI_CUP_2022_Orchid

先將 orchid_private_set.zip 和 orchid_public_set.zip 解壓縮至 ./AI_CUP_2022_Orchid_TEAM_511/AI_CUP_2022_test/orchid_set/內做存放

File Structure

├── orchid_set
    	├──orchid_public_set
	│     ├──*.jpg
	│     │
	│     │  ...
	│     │
	│     ├──*.jpg
	│    
	└── orchid_private_set
		  ├──*.jpg
		  │
		  │  ...
		  │
		  ├──*.jpg

1.在Colab上掛載Google Drive

from google.colab import drive
drive.mount("/content/drive")

2.進入到文件所在的目錄

import os
path = "/content/drive/My Drive/AI_CUP_2022_Orchid_TEAM_511"
os.chdir(path)
os.listdir(path)

3.需安裝套件:

pip install timm==0.6.2.dev0

pip install timm==0.6.2.dev0

4.執行 Final_Test.py 即可輸出答案，答案會儲存在(./Ans)

!python Final_Test.py

＃注意 本隊伍在實驗階段時有發現到Pillow的版本若非當初訓練時的版本(8.3.1)，在輸出結果上會有些許誤差。

!python Final_Test.py

5.如何執行 train.py

5-1 進入pytorch-image-models 目錄下

%cd pytorch-image-models

模型訓練的權重皆存放在 (./AI_CUP_2022_Orchid_TEAM_511/pytorch-image-models/output)

5-2 訓練 swin large

--experiment 儲存訓練實驗的檔案夾名稱

!CUDA_LAUNCH_BLOCKING=1 python train.py ../Datasets/training --train-split train --val-split val --model swin_large_patch4_window12_384_in22k --pretrained --num-classes 219 --mean 0.5075337404927281 0.45864544276917535 0.4169235386212412 --std 0.2125643051799512 0.2385082849964861 0.22386483801695406 -b 4 -vb 4 --opt adamw --weight-decay 0.01 --layer-decay 0.65 --sched cosine --lr 0.0001 --lr-cycle-limit 1 --warmup-lr 1e-5 --min-lr 1e-5 --epochs 300 --warmup-epochs 5 --color-jitter 0.5 --reprob 0.5 --recount 1 --remode pixel --cutmix 0.4 --train-interpolation bicubic --drop-path 0.1 -j 4 --save-images --output output --experiment 0

!CUDA_LAUNCH_BLOCKING=1 python train.py ../Datasets/training --train-split train --val-split val --model swin_large_patch4_window12_384_in22k --pretrained --num-classes 219 --mean 0.5075337404927281 0.45864544276917535 0.4169235386212412 --std 0.2125643051799512 0.2385082849964861 0.22386483801695406 -b 4 -vb 4 --opt adamw --weight-decay 0.01 --layer-decay 0.65 --sched cosine --lr 0.0001 --lr-cycle-limit 1 --warmup-lr 1e-5 --min-lr 1e-5 --epochs 300 --warmup-epochs 5 --color-jitter 0.5 --reprob 0.5 --recount 1 --remode pixel --cutmix 0.4 --train-interpolation bicubic --drop-path 0.1 -j 4 --save-images --output output --experiment 0

5-2-1 Fine tune swin large

!CUDA_LAUNCH_BLOCKING=1 python train.py ../Datasets/training --train-split train --val-split val --model swin_large_patch4_window12_384_in22k --initial-checkpoint ./model/0/model_best.pth.tar --num-classes 219 --mean 0.5075337404927281 0.45864544276917535 0.4169235386212412 --std 0.2125643051799512 0.2385082849964861 0.22386483801695406 -b 2 -vb 2 --opt adamw --weight-decay 0.01 --sched step --lr 0.00001 --decay-rate 1.0 --no-aug --epochs 100 --warmup-epochs 0 --color-jitter 0.0 --reprob 0.0 --recount 1 --remode pixel --cutmix 0.0 --train-interpolation bicubic --drop-path 0.0 --log-interval 30 -j 4 --save-images --output output --experiment 0_no_aug_no_weight_decay

5-3-1 訓練Swin v2 base (A-fold)

!CUDA_LAUNCH_BLOCKING=1 python train.py ../Datasets/training --train-split train --val-split val --model swinv2_base_window12to24_192to384_22kft1k --pretrained --num-classes 219 --img-size 384 --mean 0.5075337404927281 0.45864544276917535 0.4169235386212412 --std 0.2125643051799512 0.2385082849964861 0.22386483801695406 -b 2 -vb 2 --opt adamw --weight-decay 0.01 --layer-decay 0.65 --sched cosine --lr 0.0001 --lr-cycle-limit 1 --warmup-lr 1e-5 --min-lr 1e-5 --epochs 300 --warmup-epochs 5 --color-jitter 0.5 --reprob 0.5 --recount 1 --remode pixel --cutmix 0.4 --train-interpolation bicubic --drop-path 0.1 -j 4 --save-images --output output --experiment 0531_0310_swinv2_base_window12to24_192to384_22kft1k_official_aug_v0_new

5-3-2 訓練Swin v2 base (B-fold)

!CUDA_LAUNCH_BLOCKING=1 python train.py ../Datasets/AI_CUP_fold_complement --train-split train --val-split val --model swinv2_base_window12to24_192to384_22kft1k --pretrained --num-classes 219 --mean 0.45625534556274666 0.4220624936173144 0.3649616601198825 --std 0.2143212828816861 0.2210437745632625 0.2062174242104951 -b 2 -vb 2 --opt adamw --weight-decay 0.01 --layer-decay 0.65 --sched cosine --lr 0.0001 --lr-cycle-limit 2 --warmup-lr 1e-5 --min-lr 1e-5 --epochs 300 --warmup-epochs 5 --color-jitter 0.5 --reprob 0.5 --recount 1 --remode pixel --cutmix 0.4 --train-interpolation bicubic --drop-path 0.1 -j 4 --save-images --output output --experiment 0603_0131_swinv2_base_window12to24_192to384_22kft1k_complement_final_bacbbone_new --mixup-off-epoch 300

備註-指定單張GPU

CUDA_VISIBLE_DEVICES=1 bash ./distributed_train.sh 1

備註-呼叫標籤引數

from timm.data.parsers.parser_image_folder import find_images_and_targets
samples, class_to_idx=find_images_and_targets(folder='./',class_to_idx=None)