import os os.system('pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cpu') import gradio as gr import numpy as np import tensorflow as tf import torch import json from datasets.features import ClassLabel from transformers import AutoProcessor from datasets import Features, Sequence, ClassLabel, Value, Array2D, Array3D from datasets import load_dataset # this dataset uses the new Image feature :) from transformers import LayoutLMv3ForTokenClassification from transformers.data.data_collator import default_data_collator from transformers import AutoModelForTokenClassification import cv2 from PIL import Image, ImageDraw, ImageFont #setting up the Huggingface env # pip install -q git+https://github.com/huggingface/transformers.git # !pip install h5py # It's useful for evaluation metrics such as F1 on sequence labeling tasks # !pip install -q datasets seqeval # this dataset uses the new Image feature :) dataset = load_dataset("nielsr/funsd-layoutlmv3") #dataset = load_dataset("G:\\BITS - MTECH\\Sem -4\\Final Report\\code\dataset") Image.open(dataset[2]["image_path"]).convert("RGB").save("example1.png") Image.open(dataset[1]["image_path"]).convert("RGB").save("example2.png") Image.open(dataset[0]["image_path"]).convert("RGB").save("example3.png") example = dataset["test"][0] words, boxes, ner_tags = example["tokens"], example["bboxes"], example["ner_tags"] features = dataset["test"].features column_names = dataset["test"].column_names image_column_name = "image" text_column_name = "tokens" boxes_column_name = "bboxes" label_column_name = "ner_tags" def get_label_list(labels): unique_labels = set() for label in labels: unique_labels = unique_labels | set(label) label_list = list(unique_labels) label_list.sort() return label_list if isinstance(features[label_column_name].feature, ClassLabel): label_list = features[label_column_name].feature.names # No need to convert the labels since they are already ints. id2label = {k: v for k,v in enumerate(label_list)} label2id = {v: k for k,v in enumerate(label_list)} else: label_list = get_label_list(dataset["train"][label_column_name]) id2label = {k: v for k,v in enumerate(label_list)} label2id = {v: k for k,v in enumerate(label_list)} num_labels = len(label_list) label2color = {'question':'blue', 'answer':'green', 'header':'orange', 'other':'violet'} def prepare_examples(examples): images = examples[image_column_name] words = examples[text_column_name] boxes = examples[boxes_column_name] word_labels = examples[label_column_name] encoding = processor(images, words, boxes=boxes, word_labels=word_labels, truncation=True, padding="max_length") return encoding processor = AutoProcessor.from_pretrained("microsoft/layoutlmv3-base", apply_ocr=False) #model = AutoModelForTokenClassification.from_pretrained("microsoft/layoutlmv3-base") model = LayoutLMv3ForTokenClassification.from_pretrained("microsoft/layoutlmv3-base", id2label=id2label, label2id=label2id) # we need to define custom features for `set_format` (used later on) to work properly features = Features({ 'pixel_values': Array3D(dtype="float32", shape=(3, 224, 224)), 'input_ids': Sequence(feature=Value(dtype='int64')), 'attention_mask': Sequence(Value(dtype='int64')), 'bbox': Array2D(dtype="int64", shape=(512, 4)), 'labels': Sequence(feature=Value(dtype='int64')), }) # train_dataset = dataset["train"].map( # prepare_examples, # batched=True, # remove_columns=column_names, # features=features, # ) eval_dataset = dataset["test"].map( prepare_examples, batched=True, remove_columns=column_names, features=features, ) def unnormalize_box(bbox, width, height): return [ width * (bbox[0] / 1000), height * (bbox[1] / 1000), width * (bbox[2] / 1000), height * (bbox[3] / 1000), ] def process_image(image): print(type(image)) width, height = image.size # encode #encoding = processor(image, truncation=True, return_offsets_mapping=True, return_tensors="pt") #offset_mapping = encoding.pop('offset_mapping') image = example["image"] words = example["tokens"] boxes = example["bboxes"] word_labels = example["ner_tags"] encoding = processor(image, words, truncation=True,boxes=boxes, word_labels=word_labels,return_offsets_mapping=True, return_tensors="pt") offset_mapping = encoding.pop('offset_mapping') for k,v in encoding.items(): print(k,v.shape) # forward pass with torch.no_grad(): outputs = model(**encoding) # get predictions # We take the highest score for each token, using argmax. # This serves as the predicted label for each token. logits = outputs.logits #logits.shape predictions = logits.argmax(-1).squeeze().tolist() labels = encoding.labels.squeeze().tolist() token_boxes = encoding.bbox.squeeze().tolist() width, height = image.size true_predictions = [model.config.id2label[pred] for pred, label in zip(predictions, labels) if label != - 100] true_labels = [model.config.id2label[label] for prediction, label in zip(predictions, labels) if label != -100] true_boxes = [unnormalize_box(box, width, height) for box, label in zip(token_boxes, labels) if label != -100] # only keep non-subword predictions is_subword = np.array(offset_mapping.squeeze().tolist())[:,0] != 0 true_predictions = [id2label[pred] for idx, pred in enumerate(predictions) if not is_subword[idx]] true_boxes = [unnormalize_box(box, width, height) for idx, box in enumerate(token_boxes) if not is_subword[idx]] # draw predictions over the image draw = ImageDraw.Draw(image) font = ImageFont.load_default() for prediction, box in zip(true_predictions, true_boxes): predicted_label = id2label(prediction) draw.rectangle(box, outline=label2color[predicted_label]) draw.text((box[0]+10, box[1]-10), text=predicted_label, fill=label2color[predicted_label], font=font) return image title = "DocumentAI - Extraction using LayoutLMv3 model" description = "Extraction of Form or Invoice Extraction - We use Microsoft's LayoutLMv3 trained on Invoice Dataset to predict the Biller Name, Biller Address, Biller post_code, Due_date, GST, Invoice_date, Invoice_number, Subtotal and Total. To use it, simply upload an image or use the example image below. Results will show up in a few seconds." article="References
[1] Y. Xu et al., “LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking.” 2022. Paper Link
[2] LayoutLMv3 training and inference" examples =[['example1.png'],['example2.png'],['example3.png']] css = """.output_image, .input_image {height: 600px !important}""" iface = gr.Interface(fn=process_image, inputs=gr.inputs.Image(type="pil"), outputs=gr.outputs.Image(type="pil", label="annotated image"), title=title, description=description, article=article, examples=examples, css=css, analytics_enabled = True, enable_queue=True) iface.launch(inline=False, share=False, debug=False)