lilt-en-funsd

This model is a fine-tuned version of SCUT-DLVCLab/lilt-roberta-en-base on the funsd-layoutlmv3 dataset. It achieves the following results on the evaluation set:

• Loss: 1.6117
• Answer: {'precision': 0.8821428571428571, 'recall': 0.9069767441860465, 'f1': 0.8943874471937237, 'number': 817}
• Header: {'precision': 0.6126126126126126, 'recall': 0.5714285714285714, 'f1': 0.591304347826087, 'number': 119}
• Question: {'precision': 0.9045045045045045, 'recall': 0.9322191272051996, 'f1': 0.9181527206218564, 'number': 1077}
• Overall Precision: 0.8797
• Overall Recall: 0.9006
• Overall F1: 0.8900
• Overall Accuracy: 0.8204

Model Usage

from transformers import LiltForTokenClassification, LayoutLMv3Processor
from PIL import Image, ImageDraw, ImageFont
import torch

# load model and processor from huggingface hub
model = LiltForTokenClassification.from_pretrained("philschmid/lilt-en-funsd")
processor = LayoutLMv3Processor.from_pretrained("philschmid/lilt-en-funsd")


# helper function to unnormalize bboxes for drawing onto the image
def unnormalize_box(bbox, width, height):
    return [
        width * (bbox[0] / 1000),
        height * (bbox[1] / 1000),
        width * (bbox[2] / 1000),
        height * (bbox[3] / 1000),
    ]


label2color = {
    "B-HEADER": "blue",
    "B-QUESTION": "red",
    "B-ANSWER": "green",
    "I-HEADER": "blue",
    "I-QUESTION": "red",
    "I-ANSWER": "green",
}
# draw results onto the image
def draw_boxes(image, boxes, predictions):
    width, height = image.size
    normalizes_boxes = [unnormalize_box(box, width, height) for box in boxes]

    # draw predictions over the image
    draw = ImageDraw.Draw(image)
    font = ImageFont.load_default()
    for prediction, box in zip(predictions, normalizes_boxes):
        if prediction == "O":
            continue
        draw.rectangle(box, outline="black")
        draw.rectangle(box, outline=label2color[prediction])
        draw.text((box[0] + 10, box[1] - 10), text=prediction, fill=label2color[prediction], font=font)
    return image


# run inference
def run_inference(image, model=model, processor=processor, output_image=True):
    # create model input
    encoding = processor(image, return_tensors="pt")
    del encoding["pixel_values"]
    # run inference
    outputs = model(**encoding)
    predictions = outputs.logits.argmax(-1).squeeze().tolist()
    # get labels
    labels = [model.config.id2label[prediction] for prediction in predictions]
    if output_image:
        return draw_boxes(image, encoding["bbox"][0], labels)
    else:
        return labels


run_inference(dataset["test"][34]["image"])

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• training_steps: 2500
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.24.0
• Pytorch 1.12.1+cu113
• Datasets 2.7.0
• Tokenizers 0.12.1