A walk with fastai2 - Vision - Study Group and Online Lectures Megathread

fastai
945

Hi,

I finally spot the problems that the function has (YAY!). As a summary if someone would like to implement it:

  1. Transforms on GPU are on batch so the shape of the tensor is like 4x3x244x244. In this specific case, the function should change values in the x[;,channel,None,None]. This is easily spot if you run dblock.summary(path) and see the last part when batch is built.
  2. It is necessary to input and output TensorImage for the function to work with the rest of batch transforms.

The code if someone wants to use it:

@patch
def rgb_randomize(x:TensorImage, channel:int=None, thresh:float=0.3, p=0.5):
    "Randomize one of the channels of the input image"
    if channel is None: channel = np.random.randint(0, x.shape[1])
    x[:,channel,None,None] = torch.rand(x[:,channel,None,None].shape) * np.random.uniform(0, thresh)
    return TensorImage(x)

class rgb_transform(RandTransform):
    order=10
    def __init__(self, channel=None, thresh=0.3, p=0.5, **kwargs):
        super().__init__(p=p)
        self.channel,self.thresh,self.p = channel,thresh,p
    def encodes(self, x:TensorImage): return TensorImage(x).rgb_randomize(channel=self.channel,thresh=self.thresh, p=self.p )

Then you can instantiate it in your batch transforms like:
batch_tfms=[*aug_transforms(flip_vert=True), rgb_transform(thresh=0.99, p=1)]

And run Datablock as usual.

I would like to thank @sgugger for the infinite help on this. I will suggest a PR as soon as possible.

4 Likes