The End of T.Scale: What Happened to torchvision.transforms.Scale and What to Use Instead

If you’ve ever upgraded a legacy PyTorch project and suddenly hit an AttributeError: module 'torchvision.transforms' has no attribute 'Scale', you’re not alone. This post explains exactly when T.Scale was removed, why it was removed, and how to migrate your code cleanly.

What Was T.Scale?

torchvision.transforms.Scale was an early image-resizing transform in the torchvision library. It accepted either an integer (resize the shorter side to that value while preserving aspect ratio) or a tuple (H, W) (resize to an exact size). In short, it did exactly what Resize does today.

import torchvision.transforms as T

# Old way (deprecated, then removed)
transform = T.Scale(256)
transform = T.Scale((224, 224))

When Was It Removed?

Milestone Version Notes
Deprecated <= 0.8 Scale marked as deprecated; Resize introduced as the replacement
Still present 0.11 Available but with deprecation warning
Removed 0.12 Scale fully deleted from the public API

torchvision 0.12 (released alongside PyTorch 1.11 in March 2022) performed a broad API cleanup, removing all transforms that had been deprecated since version 0.8 or earlier. Scale was one of those long-overdue removals.

Why Was It Removed?

Two reasons drove the decision:

  1. It was a redundant alias. Scale and Resize were functionally identical. Maintaining two names for the same operation added confusion and documentation overhead with no benefit.

  2. Scheduled deprecation cleanup. Starting from 0.12, the torchvision team enforced a formal deprecation policy: APIs deprecated in 0.8 or earlier were removed. Scale fell squarely into this category.

What to Use Instead

import torchvision.transforms as T

# Resize shorter side to 256, preserving aspect ratio
transform = T.Resize(256)

# Resize to an exact (H, W)
transform = T.Resize((224, 224))

Functional API

import torchvision.transforms.functional as F
from PIL import Image

img = Image.open("image.jpg")
img_resized = F.resize(img, 256)          # shorter side -> 256
img_resized = F.resize(img, (224, 224))   # exact size

Quick migration reference

Old code New code
T.Scale(256) T.Resize(256)
T.Scale((224, 224)) T.Resize((224, 224))
transforms.Scale(n) transforms.Resize(n)

Checking for Other Removed APIs

If you are upgrading from an old codebase, it is worth auditing for other deprecated transforms removed in 0.12. A quick grep can help:

grep -rn "T\.Scale\|transforms\.Scale\|T\.RandomSizedCrop\|T\.CenterCrop" your_project/

RandomSizedCrop is another common one that was removed in the same cleanup wave - its replacement is RandomResizedCrop.

Summary

  • torchvision.transforms.Scale was removed in torchvision 0.12 (March 2022).
  • It was deprecated long before that (<= 0.8) because Resize is the canonical, better-named equivalent.
  • The fix is a straight one-to-one replacement: T.Scale -> T.Resize.

Keeping dependencies up to date and replacing deprecated APIs early is always cheaper than debugging a broken pipeline after a major upgrade.