File size: 3,644 Bytes
ff28913 4b1d1b4 ff28913 4b1d1b4 ff28913 4b1d1b4 ff28913 4b1d1b4 468627d 4b1d1b4 468627d 4b1d1b4 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 |
---
library_name: transformers
license: apache-2.0
datasets:
- alakxender/dhivehi-image-text
language:
- dv
base_model:
- facebook/deit-base-distilled-patch16-384
---
# 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
```python
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
```json
[
{
"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."
}
]
``` |