jeanflop/ocr_correcteur-v1

Ocr Correcteur v1

This model lora weight has been finetune on french OCR dataset. The architecture used is Flan T large. On a sample of 1000. More stong model is under cooks.

• Install dependencies

!pip install -q transformers accelerate peft diffusers
!pip install -U bitsandbytes

• Load and merge adaptaters in 8Bit (recommanded)

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer,BitsAndBytesConfig

# Load peft config for pre-trained checkpoint etc.
peft_model_id = "jeanflop/ocr_correcteur-v1"
config = PeftConfig.from_pretrained(peft_model_id)

# load base LLM model and tokenizer
peft_model = AutoModelForSeq2SeqLM.from_pretrained(config.base_model_name_or_path,  load_in_8bit=True,  device_map={"":1})
peft_tokenizer = AutoTokenizer.from_pretrained('google/flan-t5-large')

# Load the Lora model
peft_model = PeftModel.from_pretrained(peft_model, peft_model_id, device_map={"":1})
# model.eval()

print("Peft model loaded")

• Run inference (recommanded)

Add your text

inputs=f"""
Fix text : {text}"""

Run

peft_model.config.max_length=512
peft_tokenizer.model_max_length=512
inputs = peft_tokenizer(inputs, return_tensors="pt")
outputs = peft_model.generate(**inputs,max_length=512)
answer = peft_tokenizer.decode(outputs[0])
from textwrap import fill

print(fill(answer, width=80))