image_classification.py

import requests
import torch
import torch.nn.functional as F
from transformers import ImageClassificationPipeline, AutoConfig, AutoTokenizer
from PIL import Image
from io import BytesIO

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model_name_or_path = "Akazi/resnet_c_s_redwood_finetuned"
config = AutoConfig.from_pretrained(model_name_or_path)
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)

class CustomImageClassificationPipeline(ImageClassificationPipeline):
    def _sanitize_parameters(self, **pipeline_parameters):
        return {}, {}, {}

    def preprocess(self, inputs):
        if isinstance(inputs, str):
            if inputs.startswith("http://") or inputs.startswith("https://"):
                response = requests.get(inputs)
                image = Image.open(BytesIO(response.content))
            else:
                image = Image.open(inputs)
        elif isinstance(inputs, bytes):
            image = Image.open(BytesIO(inputs))
        else:
            image = inputs

        inputs = tokenizer(image, return_tensors="pt")
        inputs.to(device)

        return inputs

    def _forward(self, model_inputs):
        return self.model(**model_inputs)

    def postprocess(self, model_outputs):
        logits = model_outputs.logits
        probabilities = F.softmax(logits, dim=1).detach().cpu().numpy()[0]
        labels = [config.id2label[i] for i in range(len(probabilities))]
        outputs = [{"label": label, "score": round(score, 5)} for label, score in zip(labels, probabilities)]
        return outputs

# Create the pipeline
pipeline = CustomImageClassificationPipeline(model=model_name_or_path, device=device)

# Example usage
image_path = "path/to/your/image.jpg"
results = pipeline(image_path)

# Print the results
for result in results:
    print(f"Label: {result['label']}, Score: {result['score']}")