import torch from transformers import PerceiverForImageClassificationLearned, PerceiverImageProcessorPil from PIL import Image import requests

url = "http://images.cocodataset.org/val2017/000000039769.jpg" image = Image.open(requests.get(url, stream=True).raw) image_processor = PerceiverImageProcessorPil(size={"height": 384, "width": 384}) model = PerceiverForImageClassificationLearned.from_pretrained("deepmind/vision-perceiver-learned") model.eval() inputs = image_processor(image, return_tensors="pt").pixel_values try: with torch.no_grad(): outputs = model(inputs=inputs, interpolate_pos_encoding=True) print("Logits shape:", outputs.logits.shape) predicted_class = outputs.logits.argmax(-1).item() print("Predicted class:", predicted_class) except Exception as e: print(e)

Fix Plan

To fix the issue, we need to properly resize the input image to the expected size before passing it to the model. We can achieve this by using the image_processor to resize the image.

Step-by-Step Solution

• Update the image_processor to resize the image to the expected size (384x384) using the size parameter.
• Pass the resized image to the model.

Example Code

import torch
from transformers import PerceiverForImageClassificationLearned, PerceiverImageProcessorPil
from PIL import Image
import requests

url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)

# Update the image processor to resize the image
image_processor = PerceiverImageProcessorPil(
    size={"height": 384, "width": 384}, 
    resample=Image.BICUBIC  # Use bicubic resampling for better quality
)

model = PerceiverForImageClassificationLearned.from_pretrained("deepmind/vision-perceiver-learned")
model.eval()

# Preprocess the image
inputs = image_processor(image, return_tensors="pt").pixel_values

try:
    with torch.no_grad():
        outputs = model(inputs=inputs, interpolate_pos_encoding=True)
    print("Logits shape:", outputs.logits.shape)
    predicted_class = outputs.logits.argmax(-1).item()
    print("Predicted class:", predicted_class)
except Exception as e:
    print(e)

Verification

To verify that the fix worked, run the updated code and check if the inference completes successfully. The output should be a tensor with shape (1, 1000), representing the logits for the 1000 classes.

Extra Tips

• Make sure to use the correct resampling filter when resizing images to avoid losing important details.
• If you're working with large images, consider using a more efficient resizing method or a library like OpenCV for better performance.
• Always verify the input shape and size before passing it to the model to avoid any potential issues.