Gabriel's picture
Update handler.py
6e7afc8 verified
import base64
import io
from typing import Any, Dict, List
import requests
import torch
from PIL import Image
from transformers import TrOCRProcessor, VisionEncoderDecoderModel
device = "cuda" if torch.cuda.is_available() else "cpu"
class EndpointHandler:
def __init__(self, path=""):
self.processor = TrOCRProcessor.from_pretrained(path)
self.model = VisionEncoderDecoderModel.from_pretrained(path)
self.model.to(device)
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
inputs = data.pop("inputs", data)
image_input = inputs.get("image")
if not image_input:
return {"error": "No image provided."}
try:
if image_input.startswith("http"):
response = requests.get(image_input, stream=True)
if response.status_code == 200:
image = Image.open(response.raw).convert("RGB")
else:
return {
"error": f"Failed to fetch image. Status code: {response.status_code}"
}
else:
image_data = base64.b64decode(image_input)
image = Image.open(io.BytesIO(image_data)).convert("RGB")
except Exception as e:
return {"error": f"Failed to process the image. Details: {str(e)}"}
pixel_values = self.processor(images=image, return_tensors="pt").pixel_values
generated_ids = self.model.generate(pixel_values.to(device))
prediction = self.processor.batch_decode(
generated_ids, skip_special_tokens=True
)
return {"text": prediction[0]}