|
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]} |