import os os.system('pip install torch==1.8.0+cpu torchvision==0.9.0+cpu -f https://download.pytorch.org/whl/torch_stable.html') os.system('pip install -q detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cpu/torch1.8/index.html') import gradio as gr import numpy as np from transformers import LayoutLMv2Processor, LayoutLMv2ForTokenClassification from datasets import load_dataset from PIL import Image, ImageDraw, ImageFont processor = LayoutLMv2Processor.from_pretrained("microsoft/layoutlmv2-base-uncased") model = LayoutLMv2ForTokenClassification.from_pretrained("Theivaprakasham/layoutlmv2-finetuned-sroie") # load image example dataset = load_dataset("darentang/sroie", split="test") Image.open(dataset[50]["image_path"]).convert("RGB").save("example1.png") Image.open(dataset[14]["image_path"]).convert("RGB").save("example2.png") Image.open(dataset[20]["image_path"]).convert("RGB").save("example3.png") # define id2label, label2color labels = dataset.features['ner_tags'].feature.names id2label = {v: k for v, k in enumerate(labels)} label2color = {'B-ADDRESS': 'blue', 'B-COMPANY': 'green', 'B-DATE': 'red', 'B-TOTAL': 'red', 'I-ADDRESS': "blue", 'I-COMPANY': 'green', 'I-DATE': 'red', 'I-TOTAL': 'red', 'O': 'green'} label2color = dict((k.lower(), v.lower()) for k,v in label2color.items()) def unnormalize_box(bbox, width, height): return [ width * (bbox[0] / 1000), height * (bbox[1] / 1000), width * (bbox[2] / 1000), height * (bbox[3] / 1000), ] def iob_to_label(label): return label def process_image(image): width, height = image.size # encode encoding = processor(image, truncation=True, return_offsets_mapping=True, return_tensors="pt") offset_mapping = encoding.pop('offset_mapping') # forward pass outputs = model(**encoding) # get predictions predictions = outputs.logits.argmax(-1).squeeze().tolist() token_boxes = encoding.bbox.squeeze().tolist() # 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 = iob_to_label(prediction).lower() 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 = "Bill Information extraction using LayoutLMv2 model" description = "Bill Information Extraction - We use Microsoft's LayoutLMv2 trained on SROIE Dataset to predict the Company Name, Address, Date, and Total Amount from Bills. 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., “LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding.” 2022. Paper Link
[2] LayoutLMv2 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=True, debug=False)