Instructions to use alakxender/trocr-dv-diet-base-bert with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use alakxender/trocr-dv-diet-base-bert with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="alakxender/trocr-dv-diet-base-bert")

# Load model directly
from transformers import AutoTokenizer, AutoModelForMultimodalLM

tokenizer = AutoTokenizer.from_pretrained("alakxender/trocr-dv-diet-base-bert")
model = AutoModelForMultimodalLM.from_pretrained("alakxender/trocr-dv-diet-base-bert")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use alakxender/trocr-dv-diet-base-bert with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "alakxender/trocr-dv-diet-base-bert"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "alakxender/trocr-dv-diet-base-bert",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/alakxender/trocr-dv-diet-base-bert

SGLang

How to use alakxender/trocr-dv-diet-base-bert with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "alakxender/trocr-dv-diet-base-bert" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "alakxender/trocr-dv-diet-base-bert",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "alakxender/trocr-dv-diet-base-bert" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "alakxender/trocr-dv-diet-base-bert",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use alakxender/trocr-dv-diet-base-bert with Docker Model Runner:
```
docker model run hf.co/alakxender/trocr-dv-diet-base-bert
```

TrOCR Finetuned for Dhivehi Text Recognition

A TrOCR model finetuned for Dhivehi (Divehi/Maldivian) text recognition using DeiT base encoder and BERT decoder.

Model Details

Base models:
- Encoder: facebook/deit-base-distilled-patch16-384
- Decoder: alakxender/bert-base-dv
Training data: 10k samples with 90/10 train/test split
Input size: 384x384 pixels
Beam search parameters:
- max_length: 64
- num_beams: 4
- early_stopping: True
- length_penalty: 2.0
- no_repeat_ngram_size: 3

Training

The model was trained with:

7 epochs
Batch size: 8
Learning rate: 4e-5
FP16 mixed precision
Training augmentations:
- Elastic transform (α=8.0, σ=5.0)
- Gaussian blur (kernel size=(5,9), σ=(0.1,5))
- Resize (384x384)
- Normalization ([0.5,0.5,0.5], [0.5,0.5,0.5])

Usage

from PIL import Image
import torch
from torchvision import transforms
from transformers import (
   DeiTImageProcessor,
   TrOCRProcessor, 
   VisionEncoderDecoderModel,
   AutoTokenizer
)

class OCRPredictor:
   def __init__(self, model_name="alakxender/trocr-dv-diet-base-bert"):
       self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
       self.model = self._load_model(model_name)
       self.processor = self._load_processor()
       self.transform = self._get_transforms()
       
   def _load_model(self, model_name):
       model = VisionEncoderDecoderModel.from_pretrained(model_name)
       return model.to(self.device)
   
   def _load_processor(self):
       tokenizer = AutoTokenizer.from_pretrained("alakxender/trocr-dv-diet-base-bert")
       image_processor = DeiTImageProcessor.from_pretrained("facebook/deit-base-distilled-patch16-384")
       return TrOCRProcessor(image_processor=image_processor, tokenizer=tokenizer)
       
   def _get_transforms(self):
       return transforms.Compose([
           transforms.Resize((384, 384)),
           transforms.ToTensor(),
           transforms.Normalize([0.5] * 3, [0.5] * 3)
       ])
       
   def predict(self, image_path):
       image = Image.open(image_path).convert("RGB")
       pixel_values = self.transform(image).unsqueeze(0).to(self.device)
       
       outputs = self.model.generate(
           pixel_values,
           max_length=64,
           num_beams=4,
           early_stopping=True,
           length_penalty=2.0,
           no_repeat_ngram_size=3
       )
       
       return self.processor.decode(outputs[0], skip_special_tokens=True)
       
# Usage
predictor = OCRPredictor()
text = predictor.predict("ocr2.png")
print(text) # ތިން މިނިސްޓްރީއެއް ހިންގާ މ.ގްރީން ބިލްޑިންގުގައި މިދިޔަ ބުރާސްފަތި ދުވަހު ހިނގި ބޮޑު އަލިފާނުގެ.

Evaluation Results

[
  {
    "file_name": "data/images/DV01-04/DV01-04_140.jpg",
    "predicted_text": "ޤާނޫނުގެ 42 ވަނަ މާއްދާގައި ލާޒިމްކުރާ މި ރިޕޯޓު ތައްޔާރުކޮށް ފޮނުވުމުގެ ޒިންމާއަކީ ޤާނޫނުން އިދާރާގެ އިންފޮމޭޝަން އޮފިސަރު ކުރައްވަންޖެހޭ ކަމެކެވެ .",
    "true_text": "ޤާނޫނުގެ 42 ވަނަ މާއްދާގައި ލާޒިމްކުރާ މި ރިޕޯޓު ތައްޔާރުކޮށް ފޮނުވުމުގެ ޒިންމާއަކީ ޤާނޫނުން އިދާރާގެ އިންފޮމޭޝަން އޮފިސަރު ކުރައްވަންޖެހޭ ކަމެކެވެ."
  }
]