app.py

import os
os.system('!python -m pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu113/torch1.10/index.html')
os.system('!git clone -b add_dit_inference_bis https://github.com/NielsRogge/unilm.git')

import argparse
import cv2
from ditod import add_vit_config
from detectron2.config import get_cfg
from detectron2.utils.visualizer import ColorMode, Visualizer
from detectron2.data import MetadataCatalog
from detectron2.engine import DefaultPredictor


# Step 1: instantiate config
cfg = get_cfg()
add_vit_config(cfg)
cfg.merge_from_file("cascade_dit_base.yaml")

# Step 2: add model weights URL to config
cfg.MODEL.WEIGHTS = https://layoutlm.blob.core.windows.net/dit/dit-fts/publaynet_dit-b_mrcnn.pth

# Step 3: set device
# TODO also support GPU
cfg.MODEL.DEVICE='cpu'

# Step 4: define model
predictor = DefaultPredictor(cfg)


def analyze_image(img):
    md = MetadataCatalog.get(cfg.DATASETS.TEST[0])
    if cfg.DATASETS.TEST[0]=='icdar2019_test':
        md.set(thing_classes=["table"])
    else:
        md.set(thing_classes=["text","title","list","table","figure"])
    
    output = predictor(img)["instances"]
    v = Visualizer(img[:, :, ::-1],
                    md,
                    scale=1.0,
                    instance_mode=ColorMode.SEGMENTATION)
    result = v.draw_instance_predictions(output.to("cpu"))
    result_image = result.get_image()[:, :, ::-1]
    
    return result_image
    
title = "Interactive demo: Document Layout Analysis with DiT"
description = "This is a demo for Microsoft's Document Image Transformer (DiT)."
examples =[['publaynet_example.jpeg']]

iface = gr.Interface(fn=analyze_image, 
                     inputs=gr.inputs.Image(type="numpy"), 
                     outputs=gr.outputs.Image(type="numpy", label="analyzed image"),
                     title=title,
                     description=description,
                     article=article,
                     examples=examples,
                     enable_queue=True)
iface.launch(debug=True)