Augmentations per model#
Following table shows details of augmentations that used for each model.
Task |
Model |
Train |
Val |
Test |
|---|---|---|---|---|
Multi Class Classification
Multi Label Classification
H-Label Classification
|
Efficientnet-B0
Efficientnet-V2-S
MV3-Large
DeiT
|
- RandomResizedCrop (size=224)
- RandomFlip (flip_prob=0.5, direction=”horizontal”)
- Normalize
|
- Resize (size=224)
- Normalize
|
- Resize (size=224)
- Normalize
|
Detection
|
Yolox_l
Yolox_s
|
- Mosaic (img_scale=640, pad_val=114.0)
- RandomAffine
- MixUp (img_scale=640, ratio_range=(0.8, 1.6), pad_val=114.0)
|
- MultiScaleFlipAug (img_scale=(640, 640))
- Resize
- RandomFlip (flip_prob=0.5)
|
- MultiScaleFlipAug (img_scale=(640, 640))
- Resize
- RandomFlip (flip_prob=0.5)
|
Yolox_x
|
- YOLOXHSVRandomAug
- RandomFlip (flip_prob=0.5)
- Resize (img_scale=640)
- Pad
- Normalize
|
- Pad (size_divisor=32)
- Normalize
|
- Pad (size_divisor=32)
- Normalize
|
|
Yolox_tiny
|
- Mosaic (img_scale=640, pad_val=114.0)
- RandomAffine
- PhotoMetricDistortion
- RandomFlip (flip_prob=0.5)
- Resize (img_scale=640)
- Pad
- Normalize
|
- Resize (img_scale=(416, 416))
- MultiScaleFlipAug (img_scale=(416, 416))
- RandomFlip
- Pad
- Normalize
|
- MultiScaleFlipAug (img_scale=(416, 416))
- Resize
- RandomFlip
- Pad
- Normalize
|
|
Mobilenetv2_atss
Resnext101_atss
|
- MinIoURandomCrop
- Resize (img_scale=[(992, 736), (896, 736), (1088, 736), (992, 672), (992, 800)])
- RandomFlip (flip_prob=0.5)
- Normalize
|
- Resize (img_scale=(992, 736))
- MultiScaleFlipAug (img_scale=(992, 736))
- RandomFlip
- Normalize
|
- Resize (img_scale=(992, 736))
- MultiScaleFlipAug (img_scale=(992, 736))
- RandomFlip
- Normalize
|
|
Mobilenetv2_ssd
|
- PhotoMetricDistortion
- MinIoURandomCrop
- Resize (img_scale=(864, 864))
- Normalize
- RandomFlip (flip_prob=0.5)
|
- Resize (img_scale=(864, 864))
- MultiScaleFlipAug (img_scale=(864, 864))
- Normalize
|
- MultiScaleFlipAug (img_scale=(864, 864))
- Resize
- Normalize
|
|
Resnet50_Detr
Resnet50_dino
|
- RandomFlip (flip_prob=0.5)
- AutoAugment
- Resize
- RandomCrop
- Resize
- Normalize
- Pad (size_divisor=1)
|
- MultiScaleFlipAug (img_scale=(1333, 800)
- Resize
- RandomFlip
- Normalize
- Pad (size_divisor=32)
|
- MultiScaleFlipAug (img_scale=(1333, 800)
- Resize
- RandomFlip
- Normalize
- Pad (size_divisor=32)
|
|
Instance-segmentation
|
Convnext_maskrcnn
Efficientnetb2b_maskrcnn
Resnet50_maskrcnn
|
- Resize (img_scale=1024)
- RandomFlip (flip_prob=0.5)
- Normalize
- Pad (size_divisor=32)
|
- Resize (img_scale=1024)
- MultiScaleFlipAug
- RandomFlip (flip_prob=0.5)
- Normalize
- Pad (size_divisor=32)
|
- MultiScaleFlipAug (img_scale=1024)
- Resize
- RandomFlip (flip_prob=0.5)
- Normalize
- Pad (size_divisor=32)
|
Maskrcnn_swin_t
|
- Resize (img_scale=1344)
- RandomFlip (flip_prob=0.5)
- Normalize
- Pad (size_divisor=32)
- Pad (size_divisor=32)
|
- Resize (img_scale=1344)
- MultiScaleFlipAug
- RandomFlip (flip_prob=0.5)
- Normalize
- Pad (size_divisor=32)
|
- Resize (img_scale=1344)
- MultiScaleFlipAug
- RandomFlip (flip_prob=0.5)
- Normalize
|
|
Segmentation
|
Segnext_b
Segnext_s
Segnext_t
Lite_hrnet_18
Lite_hrnet_18_mod2
Lite_hrnet_s_mod2
Lite_hrnet_x_mod3
|
- Resize (img_scale=544)
- RandomCrop (crop_size=512, cat_max_ratio=0.75)
- RandomFlip (flip_prob=0.5, direction=”horizontal”)
- Normalize
- Pad (size=512, pad_val=0, seg_pad_val=255)
|
- Resize (img_scale=544)
- MultiScaleFlipAug
- RandomFlip
- Normalize
|
- Resize (img_scale=544)
- MultiScaleFlipAug
- RandomFlip
- Normalize
|