I am trying to port a version of this great notebook from @iafoss to fastai v2.
However, I am struggling to create an equivalent Dataloader in v2. What it mainly does is to create a DL by merging the random k images from a given sample. The specific code that I want to update is:
def open_image(fn:PathOrStr, div:bool=True, convert_mode:str='RGB', cls:type=Image,
after_open:Callable=None)->Image:
with warnings.catch_warnings():
warnings.simplefilter("ignore", UserWarning) # EXIF warning from TiffPlugin
x = PIL.Image.open(fn).convert(convert_mode)
if after_open: x = after_open(x)
x = pil2tensor(x,np.float32)
if div: x.div_(255)
return cls(1.0-x) #invert image for zero padding
class MImage(ItemBase):
def __init__(self, imgs):
self.obj, self.data = \
(imgs), [(imgs[i].data - mean[...,None,None])/std[...,None,None] for i in range(len(imgs))]
def apply_tfms(self, tfms,*args, **kwargs):
for i in range(len(self.obj)):
self.obj[i] = self.obj[i].apply_tfms(tfms, *args, **kwargs)
self.data[i] = (self.obj[i].data - mean[...,None,None])/std[...,None,None]
return self
def __repr__(self): return f'{self.__class__.__name__} {img.shape for img in self.obj}'
def to_one(self):
img = torch.stack(self.data,1)
img = img.view(3,-1,N,sz,sz).permute(0,1,3,2,4).contiguous().view(3,-1,sz*N)
return Image(1.0 - (mean[...,None,None]+img*std[...,None,None]))
class MImageItemList(ImageList):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
def __len__(self)->int: return len(self.items) or 1
def get(self, i):
fn = Path(self.items[i]).with_suffix('')
fnames = [Path(str(fn)+'_'+str(i)+'.png')for i in random.sample(range(0, 250), N)]
imgs = [open_image(fname, convert_mode=self.convert_mode, after_open=self.after_open)
for fname in fnames]
return MImage(imgs)
def reconstruct(self, t):
return MImage([mean[...,None,None]+_t*std[...,None,None] for _t in t])
def show_xys(self, xs, ys, figsize:Tuple[int,int]=(300,50), **kwargs):
rows = min(len(xs),8)
fig, axs = plt.subplots(rows,1,figsize=figsize)
for i, ax in enumerate(axs.flatten() if rows > 1 else [axs]):
xs[i].to_one().show(ax=ax, y=ys[i], **kwargs)
plt.tight_layout()
#collate function to combine multiple images into one tensor
def MImage_collate(batch:ItemsList)->Tensor:
result = torch.utils.data.dataloader.default_collate(to_data(batch))
if isinstance(result[0],list):
result = [torch.stack(result[0],1),result[1]]
return result
I am trying to follow Siamese tutorial. Porting open_image is not an issue (Siamese has the proper function) and I guess that MImage class should be more or less the same (am I right?). However, I am not smart enough to understand how should I manage to translate MImageList and MImage_collate to v2.
As a start, I see that in v2 I should label the images and I should get this from ColReader. How do I manage to do this?