File size: 5,566 Bytes
af898ba e3e2c6e af898ba e3e2c6e af898ba e3e2c6e af898ba |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 |
#!/usr/bin/env python
from __future__ import annotations
import argparse
import functools
import os
import pathlib
import sys
import tarfile
from typing import Callable
import gradio as gr
import huggingface_hub
import numpy as np
import PIL.Image
import torch
import torch.nn as nn
import torchvision
import torchvision.transforms as T
sys.path.insert(0, 'bizarre-pose-estimator')
from _util.twodee_v0 import I as ImageWrapper
ORIGINAL_REPO_URL = 'https://github.com/ShuhongChen/bizarre-pose-estimator'
TITLE = 'ShuhongChen/bizarre-pose-estimator (segmenter)'
DESCRIPTION = f'A demo for {ORIGINAL_REPO_URL}'
ARTICLE = None
TOKEN = os.environ['TOKEN']
MODEL_REPO = 'hysts/bizarre-pose-estimator-models'
MODEL_FILENAME = 'segmenter.pth'
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser()
parser.add_argument('--device', type=str, default='cpu')
parser.add_argument('--score-slider-step', type=float, default=0.05)
parser.add_argument('--score-threshold', type=float, default=0.5)
parser.add_argument('--theme', type=str)
parser.add_argument('--live', action='store_true')
parser.add_argument('--share', action='store_true')
parser.add_argument('--port', type=int)
parser.add_argument('--disable-queue',
dest='enable_queue',
action='store_false')
parser.add_argument('--allow-flagging', type=str, default='never')
return parser.parse_args()
def load_sample_image_paths() -> list[pathlib.Path]:
image_dir = pathlib.Path('images')
if not image_dir.exists():
dataset_repo = 'hysts/sample-images-TADNE'
path = huggingface_hub.hf_hub_download(dataset_repo,
'images.tar.gz',
repo_type='dataset',
use_auth_token=TOKEN)
with tarfile.open(path) as f:
f.extractall()
return sorted(image_dir.glob('*'))
def load_model(
device: torch.device) -> tuple[torch.nn.Module, torch.nn.Module]:
path = huggingface_hub.hf_hub_download(MODEL_REPO,
MODEL_FILENAME,
use_auth_token=TOKEN)
ckpt = torch.load(path)
model = torchvision.models.segmentation.deeplabv3_resnet101()
model.classifier = nn.Sequential(
torchvision.models.segmentation.deeplabv3.ASPP(2048, [12, 24, 36]),
nn.Conv2d(256, 64, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(64),
nn.LeakyReLU(),
nn.Conv2d(64, 16, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(16),
nn.LeakyReLU(),
)
final_head = nn.Sequential(
nn.Conv2d(16 + 3, 16, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(16),
nn.LeakyReLU(),
nn.Conv2d(16, 8, kernel_size=3, stride=1, padding=1),
nn.BatchNorm2d(8),
nn.LeakyReLU(),
nn.Conv2d(8, 2, kernel_size=1, stride=1),
)
model.load_state_dict(ckpt['model'])
final_head.load_state_dict(ckpt['final_head'])
model.to(device)
model.eval()
final_head.to(device)
final_head.eval()
return model, final_head
@torch.inference_mode()
def predict(image: PIL.Image.Image, score_threshold: float,
transform: Callable, device: torch.device, model: torch.nn.Module,
final_head: torch.nn.Module) -> np.ndarray:
data = ImageWrapper(image).resize_min(256).convert('RGBA').alpha_bg(
1).convert('RGB').pil()
data = torchvision.transforms.functional.to_tensor(data)
data = transform(data)
data = data.to(device).unsqueeze(0)
out = model(data)['out']
out_fin = final_head(torch.cat([
out,
data,
], dim=1))
probs = torch.softmax(out_fin, dim=1)[0]
probs = probs[1] # foreground
probs = PIL.Image.fromarray(probs.cpu().numpy()).resize(image.size)
mask = np.asarray(probs)
mask[mask < score_threshold] = 0
mask[mask > 0] = 1
mask = mask.astype(bool)
res = np.asarray(image)
res[~mask] = 255
return res
def main():
gr.close_all()
args = parse_args()
device = torch.device(args.device)
image_paths = load_sample_image_paths()
examples = [[path.as_posix(), args.score_threshold]
for path in image_paths]
model, final_head = load_model(device)
transform = T.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225])
func = functools.partial(predict,
transform=transform,
device=device,
model=model,
final_head=final_head)
func = functools.update_wrapper(func, predict)
gr.Interface(
func,
[
gr.inputs.Image(type='pil', label='Input'),
gr.inputs.Slider(0,
1,
step=args.score_slider_step,
default=args.score_threshold,
label='Score Threshold'),
],
gr.outputs.Image(label='Masked'),
examples=examples,
title=TITLE,
description=DESCRIPTION,
article=ARTICLE,
theme=args.theme,
allow_flagging=args.allow_flagging,
live=args.live,
).launch(
enable_queue=args.enable_queue,
server_port=args.port,
share=args.share,
)
if __name__ == '__main__':
main()
|