demo.py

# -*- coding: utf-8 -*-
"""Demo.ipynb

Automatically generated by Colaboratory.

Original file is located at
    https://colab.research.google.com/drive/1Icb8zeoaudyTDOKM1QySNay1cXzltRAp
"""

!pip install -q -U gradio peft

import gradio as gr
from tqdm.notebook import tqdm
from PIL import Image
import re

import torch
import torch.nn as nn
from warnings import simplefilter

simplefilter('ignore')
device = 'cuda' if torch.cuda.is_available() else 'cpu'

# Seting up the model
from peft import PeftConfig, PeftModel
numeric_lora_config = PeftConfig.from_pretrained("Edgar404/donut-sroie-lora-r8-x3")

from transformers import VisionEncoderDecoderConfig

image_size = [720,960]
max_length = 512

config = VisionEncoderDecoderConfig.from_pretrained(numeric_lora_config.base_model_name_or_path)
config.encoder.image_size = image_size
config.decoder.max_length = max_length

from transformers import DonutProcessor, VisionEncoderDecoderModel
model = VisionEncoderDecoderModel.from_pretrained(numeric_lora_config.base_model_name_or_path ,config = config )
numeric_processor = DonutProcessor.from_pretrained("Edgar404/donut-sroie-lora-r8-x3")

model.config.pad_token_id = numeric_processor.tokenizer.pad_token_id
model.config.decoder_start_token_id = numeric_processor.tokenizer.convert_tokens_to_ids(['<s_cord-v2>'])[0]
model.decoder.resize_token_embeddings(len(numeric_processor.tokenizer))

model = PeftModel.from_pretrained(model, model_id = "Edgar404/donut-sroie-lora-r8-x3", adapter_name = 'numeric')
model.to(device)

# Handwritten setting

hand_processor = DonutProcessor.from_pretrained("Edgar404/donut-lora-r8-x2")


def resize_token_handwritten():

  try :
    model.load_adapter("Edgar404/donut-lora-r8-x2" ,'handwritten')

  except Exception :
    # resizing the handwritten embedding layer
    embedding_layer  = model.decoder.model.decoder.embed_tokens.modules_to_save.handwritten
    old_num_tokens, old_embedding_dim = embedding_layer.weight.shape

    new_embeddings = nn.Embedding(
            len(hand_processor.tokenizer), old_embedding_dim
    )

    new_embeddings.to(
        embedding_layer.weight.device,
        dtype=embedding_layer.weight.dtype,
    )

    model.decoder.model.decoder.embed_tokens.modules_to_save.handwritten = new_embeddings


    # Resizing the handwritten lm_head layer
    lm_layer = model.decoder.lm_head.modules_to_save.handwritten

    old_num_tokens, old_input_dim = lm_layer.weight.shape

    new_lm_head = nn.Linear(
            old_input_dim, len(hand_processor.tokenizer),
            bias = False
    )

    new_lm_head.to(
        lm_layer.weight.device,
        dtype=lm_layer.weight.dtype,
    )
    model.decoder.lm_head.modules_to_save.handwritten = new_lm_head


resize_token_handwritten()

model.load_adapter("Edgar404/donut-lora-r8-x2" ,'handwritten')


def process_image(image , mode = 'numeric' ):
    """ Function that takes an image and perform an OCR using the model DonUT via the task document
    parsing

    parameters
    __________
    image : a machine readable image of class PIL or numpy"""

    model.set_adapter(mode)
    processor = numeric_processor if mode == 'numeric' else hand_processor



    task_prompt = "<s_cord-v2>"
    decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids

    pixel_values = processor(image, return_tensors="pt").pixel_values

    outputs = model.generate(
        pixel_values.to(device),
        decoder_input_ids=decoder_input_ids.to(device),
        max_length=model.decoder.config.max_position_embeddings,
        pad_token_id=processor.tokenizer.pad_token_id,
        eos_token_id=processor.tokenizer.eos_token_id,
        use_cache=True,
        bad_words_ids=[[processor.tokenizer.unk_token_id]],
        return_dict_in_generate=True,
    )

    sequence = processor.batch_decode(outputs.sequences)[0]
    sequence = sequence.replace(processor.tokenizer.eos_token, "").replace(processor.tokenizer.pad_token, "")
    sequence = re.sub(r"<.*?>", "", sequence, count=1).strip()
    output = processor.token2json(sequence)

    return output

import gradio as gr

def image_classifier(image , mode):
    return process_image(image , mode)



examples_list = [['./test_images/TRAIN_00001.jpg' ,"handwritten"] ,
                 ['./test_images/001.jpg','numeric'],
                 ['./test_images/TEST_0019.jpg' ,"handwritten"],
                 ['./test_images/005.jpg','numeric'],
                 ['./test_images/007.jpg','numeric'],
                 ['./test_images/VALIDATION_0011.jpg' ,"handwritten"],
                 ['./test_images/VALIDATION_0022.jpg' ,"handwritten"],
                 ['./test_images/062.jpg','numeric'],
                 ['./test_images/119.jpg','numeric'],
                 ['./test_images/150.jpg','numeric']
                 ]

demo = gr.Interface(fn=image_classifier, inputs=["image",
                                                 gr.Radio(["handwritten", "numeric"], label="mode")],
                     outputs="text",
                    examples = examples_list )

demo.launch(share = True)