Useful Tip: Transforms supported for point regression, `TensorPoint`

muellerzr · January 30, 2020, 3:40pm

I’ve always been confused as to which transforms work for keypoints (or anything else) as our y's get transformed depending on the x. In v2 you can look at the type dispatch for which transforms can be applied for what. I thought I’d make a quick list for those that need to know for their problems. This is especially useful as in most of the recent papers, the only augmentation I’ve seen is a flip left right or a rotation, and fastai offers a bit more than that!

General Augmentation (Push to the GPU/Batch transforms):

Rotate, FlipItem, DihedralItem, Zoom, Warp, FlipBatch
RandomResizedCropGPU

Cropping (Initial CPU transforms):

CropPad, RandomCrop, Resize, RandomResizedCrop

Hope this helps someone

For those that are wanting to see the source code to understand what I mean, see here:

github.com

fastai/fastai2/blob/master/fastai2/vision/augment.py

# AUTOGENERATED! DO NOT EDIT! File to edit: nbs/09_vision.augment.ipynb (unless otherwise specified).

__all__ = ['RandTransform', 'TensorTypes', 'FlipItem', 'DihedralItem', 'PadMode', 'CropPad', 'RandomCrop',
           'OldRandomCrop', 'ResizeMethod', 'Resize', 'RandomResizedCrop', 'AffineCoordTfm', 'RandomResizedCropGPU',
           'affine_mat', 'mask_tensor', 'flip_mat', 'Flip', 'DeterministicDraw', 'DeterministicFlip', 'dihedral_mat',
           'Dihedral', 'DeterministicDihedral', 'rotate_mat', 'Rotate', 'zoom_mat', 'Zoom', 'find_coeffs',
           'apply_perspective', 'Warp', 'LightingTfm', 'Brightness', 'Contrast', 'cutout_gaussian', 'norm_apply_denorm',
           'RandomErasing', 'setup_aug_tfms', 'aug_transforms']

# Cell
from ..data.all import *
from .core import *
from .data import *

# Cell
from torch import stack, zeros_like as t0, ones_like as t1
from torch.distributions.bernoulli import Bernoulli

# Cell
class RandTransform(Transform):

This file has been truncated. show original

jeremy · January 30, 2020, 4:59pm

Best would be to use the GPU versions, since they’re quite a bit faster.

muellerzr · January 30, 2020, 5:00pm

Noted! I’ll edit them in Thanks. Actually @jeremy I can’t seem to find the GPU versions, only for RandomResizedCrop. Am I looking at the right spot? https://github.com/fastai/fastai2/blob/master/fastai2/vision/augment.py

jeremy · January 30, 2020, 7:22pm

IIRC all the affine tfms are GPU - and there’s also flip_batch. Best to check the notebook, rather than the auto-generated py file.

muellerzr · January 30, 2020, 7:23pm

Got it I thought I had heard something like that. Thanks for solidifying that for me!