Rotation with Pytorch Torchvision but from two specific angles

Question

I want to randomly rotate my images at 45 degrees or -45 degrees, but only these two, no in between. 50% chances for 45 and 50% for -45.

I know that I can use RandomRotation of torchvision, but if I do like in this example:

import torchvision.transforms as T

transforms = T.Compose([
  T.RandomRotation((-45, 45))
])

The rotation will be random between -45 and 45 degrees, so that's not what I want. How to be sure the rotation will be specifically these two angles?

Thanks in advance for your answers.

Don't use a RandomRotation, but "flip a coin" e.g. if rand()>0.5: then apply a fix angle rotation? — Julien
– Julien, Commented Sep 3 at 7:29

Ahmad Abdallah · Accepted Answer · 2025-09-03 08:11:01Z

1

Like Julien said, just make a 50/50 random choice, then do the fixed angle rotation. Below is an example

import torchvision.transforms as T

transforms = T.Compose([
    T.RandomChoice([
        T.RandomRotation((45, 45)),
        T.RandomRotation((-45, -45)),
    ], p=[0.5, 0.5])
])

answered Sep 3 at 8:11

Ahmad Abdallah

1308 bronze badges

Sign up to request clarification or add additional context in comments.

Comments

simon · Accepted Answer · 2025-09-03 13:17:55Z

A quick fix would be to override the implementation of RandomRotation's get_params(), which currently draws from the interval [degrees[0], degrees[1]] with uniform probability, so that it draws either degrees[0] or degrees[1] with equal probability instead (doing basically a "coin flip", as mentioned in Julien's comment).

Internally, get_params() is used in RandomRotation's forward() method to get a new random value for each call. Also, it is always ensured that degrees is a two-element sequence, even if the instance's degrees parameter is initialized with a scalar value. So, to put it differently: currently, degrees is used as the two interval bounds (lower, upper); in your case, it can be adapted to be used as the two possible choices instead, meaning it more or less accidentally fits your purposes with this slight tweak.

In code, this can be achieved, for example, as follows:

import torch
from torchvision.transforms import RandomRotation

class FixedRandomRotation(RandomRotation):
    
    @staticmethod
    def get_params(degrees: list[float]) -> float:
        # Draw element from `degrees` sequence with equal probability
        return float(degrees[torch.randint(0, len(degrees), size=(1,))])

A full code example, using SciPy's face image as a sample, for demonstration:

import matplotlib.pyplot as plt  # For plotting
from scipy.datasets import face  # For demo image
import torch
from torchvision.transforms import RandomRotation

class FixedRandomRotation(RandomRotation):
    
    @staticmethod
    def get_params(degrees: list[float]) -> float:
        # Draw element from `degrees` sequence with equal probability
        return float(degrees[torch.randint(0, len(degrees), size=(1,))])
    
# Create rotation instance and demo image (permute: channels first)
rot = FixedRandomRotation(degrees=45)
demo_img = torch.from_numpy(face()).permute(2, 0, 1)

# Draw 16 times and plot results
fig, axes = plt.subplots(4, 4)
axes = axes.flatten()

for i in range(len(axes)):
    transformed = rot(demo_img)
    axes[i].set_axis_off()
    axes[i].imshow(transformed.permute(1, 2, 0).cpu().numpy())
plt.tight_layout(); plt.show()

This produces:

Collectives™ on Stack Overflow

Rotation with Pytorch Torchvision but from two specific angles

2 Answers 2

Comments

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Related