Posted on Apr 19 • Edited on May 16

Resize in PyTorch

#python #pytorch #resize #v2

Buy Me a Coffee☕

*Memos:

My post explains RandomResizedCrop() about size argument (1).
My post explains Pad().
My post explains OxfordIIITPet().

Resize() can resize an image as shown below:

*Memos:

The 1st argument for initialization is size(Required-Type:int, tuple/list(int) or size()): *Memos:
- It's [height, width].
- It must be 1 <= x.
- None can be explicitly set to it only if max_size isn't None.
- A tuple/list must be the 1D with 1 or 2 elements.
- A single value(int or tuple/list(int)) is applied to a smaller image's width or height edge, then the other larger width or height edge is also resized: *Memos:
- If an image's width is smaller than its height, it's [size * height / width, size].
- If an image width is larger than its height, it's [size, size * width / height].
- If an image width is equal to its height, it's [size, size].
The 2nd argument for initialization is interpolation(Optional-Default:InterpolationMode.BILINEAR-Type:InterpolationMode): *Memos:
- NEAREST, NEAREST_EXACT, BILINEAR, BICUBIC, BOX, HAMMING and LANCZOS modes can be used.
- My post explains InterpolationMode with and without anti-aliasing.
The 3rd argument for initialization is max_size(Optional-Default:None-Type:int): *Memos:
- It's only supported if size is a single value(int or tuple/list(int)).
- After size is applied if a larger image's width or height edge exceeds it, it's applied to a larger image's width or height edge to limit the image size, then the other smaller image's width or height edge also becomes smaller than before.
The 4th argument for initialization is antialias(Optional-Default:True-Type:bool): *Memos:
- If it's True(Default) and interpolation is BILINEAR or BICUBIC, anti-aliasing is applied for both a PIL image and tensor.
- If it's False or None and interpolation is BILINEAR or BICUBIC, anti-aliasing is still applied for a PIL image while anti-aliasing isn't applied for a tensor.
The 1st argument is img(Required-Type:PIL Image or tensor(int/float/complex/bool)): *Memos:
- A tensor must be 3D or more D.
- Don't use img=.
v2 is recommended to use according to V1 or V2? Which one should I use?.

from torchvision.datasets import OxfordIIITPet from torchvision.transforms.v2 import Resize from torchvision.transforms.functional import InterpolationMode r = Resize(size=100) r = Resize(size=100, interpolation=InterpolationMode.BILINEAR, max_size=None, antialias=True) r # Resize(size=[100], interpolation=InterpolationMode.BILINEAR, # antialias=True)  r.size # [100]  r.interpolation # <InterpolationMode.BILINEAR: 'bilinear'>  print(r.max_size) # None  r.antialias # True  origin_data = OxfordIIITPet( root="data", transform=None ) s1000_data = OxfordIIITPet( # `s` is size.  root="data", transform=Resize(size=1000) # transform=Resize(size=[1000])  # transform=Resize(size=[1000, 1000]) ) s500_data = OxfordIIITPet( root="data", transform=Resize(size=500) ) s100_data = OxfordIIITPet( root="data", transform=Resize(size=100) ) s50_data = OxfordIIITPet( root="data", transform=Resize(size=50) ) s10_data = OxfordIIITPet( root="data", transform=Resize(size=10) ) s1_data = OxfordIIITPet( root="data", transform=Resize(size=1) ) s600_900_data = OxfordIIITPet( root="data", transform=Resize(size=[600, 900]) ) s900_600_data = OxfordIIITPet( root="data", transform=Resize(size=[900, 600]) ) s200_300_data = OxfordIIITPet( root="data", transform=Resize(size=[200, 300]) ) s300_200_data = OxfordIIITPet( root="data", transform=Resize(size=[300, 200]) ) s1000origin_data = OxfordIIITPet( root="data", transform=Resize(size=1000) ) s1000ms1100_data = OxfordIIITPet( # `ms` is max_size.  root="data", transform=Resize(size=1000, max_size=1100) ) sNonems1100_data = OxfordIIITPet( root="data", transform=Resize(size=None, max_size=1100) ) s100origin_data = OxfordIIITPet( root="data", transform=Resize(size=100, max_size=110) ) s100ms110_data = OxfordIIITPet( root="data", transform=Resize(size=100, max_size=110) ) sNonems110_data = OxfordIIITPet( root="data", transform=Resize(size=None, max_size=110) ) import matplotlib.pyplot as plt def show_images1(data, main_title=None): plt.figure(figsize=[10, 5]) plt.suptitle(t=main_title, y=0.8, fontsize=14) for i, (im, _) in zip(range(1, 6), data): plt.subplot(1, 5, i) plt.imshow(X=im) plt.tight_layout() plt.show() show_images1(data=origin_data, main_title="origin_data") show_images1(data=s1000_data, main_title="s1000_data") show_images1(data=s500_data, main_title="s500_data") show_images1(data=s100_data, main_title="s100_data") show_images1(data=s50_data, main_title="s50_data") show_images1(data=s10_data, main_title="s10_data") show_images1(data=s1_data, main_title="s1_data") print() show_images1(data=origin_data, main_title="origin_data") show_images1(data=s600_900_data, main_title="s600_900_data") show_images1(data=s900_600_data, main_title="s900_600_data") show_images1(data=s600_900_data, main_title="s200_300_data") show_images1(data=s900_600_data, main_title="s300_200_data") print() show_images1(data=s1000origin_data, main_title="s1000origin_data") show_images1(data=s1000ms1100_data, main_title="s1000ms1100_data") show_images1(data=sNonems1100_data, main_title="sNonems1100_data") print() show_images1(data=s100origin_data, main_title="s100origin_data") show_images1(data=s100ms110_data, main_title="s100ms110_data") show_images1(data=sNonems110_data, main_title="sNonems110_data") # ↓ ↓ ↓ ↓ ↓ ↓ The code below is identical to the code above. ↓ ↓ ↓ ↓ ↓ ↓ def show_images2(data, main_title=None, s=None, ip=InterpolationMode.BILINEAR, ms=None, a=True): plt.figure(figsize=[10, 5]) plt.suptitle(t=main_title, y=0.8, fontsize=14) if main_title != "origin_data": for i, (im, _) in zip(range(1, 6), data): plt.subplot(1, 5, i) r = Resize(size=s, interpolation=ip, max_size=ms, antialias=a) plt.imshow(X=r(im)) else: for i, (im, _) in zip(range(1, 6), data): plt.subplot(1, 5, i) plt.imshow(X=im) plt.tight_layout() plt.show() show_images2(data=origin_data, main_title="origin_data") show_images2(data=origin_data, main_title="s1000_data", s=1000) show_images2(data=origin_data, main_title="s500_data", s=500) show_images2(data=origin_data, main_title="s100_data", s=100) show_images2(data=origin_data, main_title="s50_data", s=50) show_images2(data=origin_data, main_title="s10_data", s=10) show_images2(data=origin_data, main_title="s1_data", s=1) print() show_images2(data=origin_data, main_title="origin_data") show_images2(data=origin_data, main_title="s600_900_data", s=[600, 900]) show_images2(data=origin_data, main_title="s900_600_data", s=[900, 600]) show_images2(data=origin_data, main_title="s200_300_data", s=[200, 300]) show_images2(data=origin_data, main_title="s300_200_data", s=[300, 200]) print() show_images2(data=origin_data, main_title="s1000origin_data", s=1000) show_images2(data=origin_data, main_title="s1000ms1100_data", s=1000, ms=1100) show_images2(data=origin_data, main_title="sNonems1100_data", ms=1100) print() show_images2(data=origin_data, main_title="s100origin_data", s=100) show_images2(data=origin_data, main_title="s100ms110_data", s=100, ms=110) show_images2(data=origin_data, main_title="sNonems110_data", ms=110)

DEV Community

Resize in PyTorch

Top comments (0)