app.py

import json
import os
from pathlib import Path

import gradio as gr
import numpy as np
import torch
from monai.bundle import ConfigParser

from utils import page_utils

with open("configs/inference.json") as f:
    inference_config = json.load(f)

device = torch.device('cpu')
if torch.cuda.is_available():
    device = torch.device('cuda:0')

# * NOTE: device must be hardcoded, config file won't affect the device selection
inference_config["device"] = device

parser = ConfigParser()
parser.read_config(f=inference_config)
parser.read_meta(f="configs/metadata.json")

inference = parser.get_parsed_content("inferer")
# loader = parser.get_parsed_content("dataloader")
network = parser.get_parsed_content("network_def")
preprocess = parser.get_parsed_content("preprocessing")
postprocess = parser.get_parsed_content("postprocessing")

use_fp16 = os.environ.get('USE_FP16', False)

state_dict = torch.load("models/model.pt")
network.load_state_dict(state_dict, strict=True)

network = network.to(device)
network.eval()

if use_fp16 and torch.cuda.is_available():
    network = network.half()

label2color = {0: (0, 0, 0),
             1: (225, 24, 69), # RED
             2: (135, 233, 17), # GREEN
             3: (0, 87, 233), # BLUE
             4: (242, 202, 25), # YELLOW
             5: (137, 49, 239),} # PURPLE

example_files = list(Path("sample_data").glob("*.png"))

def visualize_instance_seg_mask(mask):
    image = np.zeros((mask.shape[0], mask.shape[1], 3))
    labels = np.unique(mask)
    for i in range(image.shape[0]):
      for j in range(image.shape[1]):
        image[i, j, :] = label2color[mask[i, j]]
    image = image / 255
    return image

def query_image(img):
    data = {"image": img}
    batch = preprocess(data)
    batch['image'] = batch['image'].to(device)

    if use_fp16 and torch.cuda.is_available():
        batch['image'] = batch['image'].half()

    with torch.no_grad():
        pred = inference(batch['image'].unsqueeze(dim=0), network)

    batch["pred"] = pred
    for k,v in batch["pred"].items():
        batch["pred"][k] = v.squeeze(dim=0)

    batch = postprocess(batch)

    result = visualize_instance_seg_mask(batch["type_map"].squeeze())

    # Combine image
    result = batch["image"].permute(1, 2, 0).cpu().numpy() * 0.5 + result * 0.5

    # Solve rotating problem
    result = np.fliplr(result)
    result = np.rot90(result, k=1)

    return result

# load Markdown file
with open('index.html', encoding='utf-8') as f:
    html_content = f.read()

demo = gr.Interface(
    query_image,
    inputs=[gr.Image(type="filepath")],
    outputs="image",
    theme=gr.themes.Default(primary_hue=page_utils.KALBE_THEME_COLOR, secondary_hue=page_utils.KALBE_THEME_COLOR).set(
        button_primary_background_fill="*primary_600",
        button_primary_background_fill_hover="*primary_500",
        button_primary_text_color="white",
    ),
    description = html_content,
    examples=example_files,
)

demo.queue(concurrency_count=20).launch()