disable all data augmentation results in unexpected cropping

[shunmian]

Describe the bug
When disable all data augmentation, it results in unexpected cropping on original image and lose the key feature of the image.

To Reproduce

I have disabled all data augmentation by setting the below parameters in create_transform, especially the scale, ratio, crop_pct

import matplotlib.pyplot as plt
from PIL import Image
from torchvision.transforms.functional import to_pil_image, resize

image_path = 'input_image.jpg'
img = Image.open(image_path).convert("L")

# Define the data augmentation transform by closing all augmentation
transform = create_transform(
    input_size=224,       
    hflip=0.0,            
    vflip=0.0,            
    is_training=False,    
    auto_augment=None,    
    interpolation='bilinear',  
    mean=(0.5), 
    std=(0.5),   
    scale=[1.0, 1.0],  # Disable scaling
    ratio=[1.0, 1.0],  # Disable aspect ratio changes
    crop_pct=1       # Disable cropping
)

# Function to visualize the augmentations
def visualize_augmentations(image, transform, n_samples=5):
    plt.figure(figsize=(18, 5))
    
    # Plot the original image (original aspect ratio)
    plt.subplot(1, n_samples + 2, 1)
    plt.imshow(image, cmap='gray')  # Use 'gray' colormap for grayscale
    plt.axis("off")
    plt.title("Original Image (Original Aspect)")
    
    # Resize the original image to the input_size
    resized_image = resize(image, (224, 224))
    
    # Plot the resized original image
    plt.subplot(1, n_samples + 2, 2)
    plt.imshow(resized_image, cmap='gray')  # Use 'gray' colormap for grayscale
    plt.axis("off")
    plt.title("Original Image (Resized)")
    
    # Plot the augmented images
    for i in range(n_samples):
        # Apply the transformation
        augmented_image = transform(image)
        
        # Convert to PIL image for visualization
        augmented_image = to_pil_image(augmented_image, mode='L')
        
        # Plot the image
        plt.subplot(1, n_samples + 2, i + 3)
        plt.imshow(augmented_image, cmap='gray')  # Use 'gray' colormap for grayscale
        plt.axis("off")
        plt.title(f"Augmented {i + 1}")
    
    plt.suptitle("Original and Augmented Images", fontsize=16)
    plt.show()

# Visualize the augmentations
visualize_augmentations(img, transform)

Expected behavior
The Augmented 1-5 image should be same as Original Image (Resized), especially should include "A", which is an important feature of the input image that should not being cropped off.

Screenshots
input_image.jpg

output_image.jpg

Desktop (please complete the following information):

• OS: MacOS Sonoma 14.1
• This repository version: timm version 1.0.12
• PyTorch version: 2.5.1

[rwightman]

@shunmian that's not augmentation, that's expected (default) center cropping. The transforms output a constant image size, it defaults to square unless you provide a different size tuple. You input a non-square image it resizes the shortest size to 224 and then crops center to output 224 x 224. This is the default preprocessing expected for most models. Note the scale/ratio args are for training random resize & crop and don't impact things when is_training=False.

If you want the whole image you can specificy crop_mode='squash', but it will distort aspect, or crop_mode='border' and it will add borders (letterboxing) to achieve target size.

[rwightman]

Also, that fn is documented, https://huggingface.co/docs/timm/en/reference/data#timm.data.create_transform