Add files
Browse files- .gitmodules +3 -0
- app.py +184 -0
- bizarre-pose-estimator +1 -0
- requirements.txt +2 -0
.gitmodules
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
[submodule "bizarre-pose-estimator"]
|
2 |
+
path = bizarre-pose-estimator
|
3 |
+
url = https://github.com/ShuhongChen/bizarre-pose-estimator
|
app.py
ADDED
@@ -0,0 +1,184 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python
|
2 |
+
|
3 |
+
from __future__ import annotations
|
4 |
+
|
5 |
+
import argparse
|
6 |
+
import functools
|
7 |
+
import os
|
8 |
+
import pathlib
|
9 |
+
import subprocess
|
10 |
+
import sys
|
11 |
+
import tarfile
|
12 |
+
from typing import Callable
|
13 |
+
|
14 |
+
# workaround for https://github.com/gradio-app/gradio/issues/483
|
15 |
+
command = 'pip install -U gradio==2.7.0'
|
16 |
+
subprocess.call(command.split())
|
17 |
+
|
18 |
+
import gradio as gr
|
19 |
+
import huggingface_hub
|
20 |
+
import numpy as np
|
21 |
+
import PIL.Image
|
22 |
+
import torch
|
23 |
+
import torch.nn as nn
|
24 |
+
import torchvision
|
25 |
+
import torchvision.transforms as T
|
26 |
+
|
27 |
+
sys.path.insert(0, 'bizarre-pose-estimator')
|
28 |
+
|
29 |
+
from _util.twodee_v0 import I as ImageWrapper
|
30 |
+
|
31 |
+
TOKEN = os.environ['TOKEN']
|
32 |
+
|
33 |
+
MODEL_REPO = 'hysts/bizarre-pose-estimator-models'
|
34 |
+
MODEL_FILENAME = 'segmenter.pth'
|
35 |
+
|
36 |
+
|
37 |
+
def parse_args() -> argparse.Namespace:
|
38 |
+
parser = argparse.ArgumentParser()
|
39 |
+
parser.add_argument('--device', type=str, default='cpu')
|
40 |
+
parser.add_argument('--score-slider-step', type=float, default=0.05)
|
41 |
+
parser.add_argument('--score-threshold', type=float, default=0.5)
|
42 |
+
parser.add_argument('--theme', type=str)
|
43 |
+
parser.add_argument('--live', action='store_true')
|
44 |
+
parser.add_argument('--share', action='store_true')
|
45 |
+
parser.add_argument('--port', type=int)
|
46 |
+
parser.add_argument('--disable-queue',
|
47 |
+
dest='enable_queue',
|
48 |
+
action='store_false')
|
49 |
+
parser.add_argument('--allow-flagging', type=str, default='never')
|
50 |
+
parser.add_argument('--allow-screenshot', action='store_true')
|
51 |
+
return parser.parse_args()
|
52 |
+
|
53 |
+
|
54 |
+
def load_sample_image_paths() -> list[pathlib.Path]:
|
55 |
+
image_dir = pathlib.Path('images')
|
56 |
+
if not image_dir.exists():
|
57 |
+
dataset_repo = 'hysts/sample-images-TADNE'
|
58 |
+
path = huggingface_hub.hf_hub_download(dataset_repo,
|
59 |
+
'images.tar.gz',
|
60 |
+
repo_type='dataset',
|
61 |
+
use_auth_token=TOKEN)
|
62 |
+
with tarfile.open(path) as f:
|
63 |
+
f.extractall()
|
64 |
+
return sorted(image_dir.glob('*'))
|
65 |
+
|
66 |
+
|
67 |
+
def load_model(
|
68 |
+
device: torch.device) -> tuple[torch.nn.Module, torch.nn.Module]:
|
69 |
+
path = huggingface_hub.hf_hub_download(MODEL_REPO,
|
70 |
+
MODEL_FILENAME,
|
71 |
+
use_auth_token=TOKEN)
|
72 |
+
ckpt = torch.load(path)
|
73 |
+
|
74 |
+
model = torchvision.models.segmentation.deeplabv3_resnet101()
|
75 |
+
model.classifier = nn.Sequential(
|
76 |
+
torchvision.models.segmentation.deeplabv3.ASPP(2048, [12, 24, 36]),
|
77 |
+
nn.Conv2d(256, 64, kernel_size=3, stride=1, padding=1),
|
78 |
+
nn.BatchNorm2d(64),
|
79 |
+
nn.LeakyReLU(),
|
80 |
+
nn.Conv2d(64, 16, kernel_size=3, stride=1, padding=1),
|
81 |
+
nn.BatchNorm2d(16),
|
82 |
+
nn.LeakyReLU(),
|
83 |
+
)
|
84 |
+
final_head = nn.Sequential(
|
85 |
+
nn.Conv2d(16 + 3, 16, kernel_size=3, stride=1, padding=1),
|
86 |
+
nn.BatchNorm2d(16),
|
87 |
+
nn.LeakyReLU(),
|
88 |
+
nn.Conv2d(16, 8, kernel_size=3, stride=1, padding=1),
|
89 |
+
nn.BatchNorm2d(8),
|
90 |
+
nn.LeakyReLU(),
|
91 |
+
nn.Conv2d(8, 2, kernel_size=1, stride=1),
|
92 |
+
)
|
93 |
+
model.load_state_dict(ckpt['model'])
|
94 |
+
final_head.load_state_dict(ckpt['final_head'])
|
95 |
+
model.to(device)
|
96 |
+
model.eval()
|
97 |
+
final_head.to(device)
|
98 |
+
final_head.eval()
|
99 |
+
return model, final_head
|
100 |
+
|
101 |
+
|
102 |
+
@torch.inference_mode()
|
103 |
+
def predict(image: PIL.Image.Image, score_threshold: float,
|
104 |
+
transform: Callable, device: torch.device, model: torch.nn.Module,
|
105 |
+
final_head: torch.nn.Module) -> np.ndarray:
|
106 |
+
data = ImageWrapper(image).resize_min(256).convert('RGBA').alpha_bg(
|
107 |
+
1).convert('RGB').pil()
|
108 |
+
data = torchvision.transforms.functional.to_tensor(data)
|
109 |
+
data = transform(data)
|
110 |
+
data = data.to(device).unsqueeze(0)
|
111 |
+
|
112 |
+
out = model(data)['out']
|
113 |
+
out_fin = final_head(torch.cat([
|
114 |
+
out,
|
115 |
+
data,
|
116 |
+
], dim=1))
|
117 |
+
probs = torch.softmax(out_fin, dim=1)[0]
|
118 |
+
probs = probs[1] # foreground
|
119 |
+
probs = PIL.Image.fromarray(probs.cpu().numpy()).resize(image.size)
|
120 |
+
|
121 |
+
mask = np.asarray(probs)
|
122 |
+
mask[mask < score_threshold] = 0
|
123 |
+
mask[mask > 0] = 1
|
124 |
+
mask = mask.astype(bool)
|
125 |
+
|
126 |
+
res = np.asarray(image)
|
127 |
+
res[~mask] = 255
|
128 |
+
return res
|
129 |
+
|
130 |
+
|
131 |
+
def main():
|
132 |
+
gr.close_all()
|
133 |
+
|
134 |
+
args = parse_args()
|
135 |
+
device = torch.device(args.device)
|
136 |
+
|
137 |
+
image_paths = load_sample_image_paths()
|
138 |
+
examples = [[path.as_posix(), args.score_threshold]
|
139 |
+
for path in image_paths]
|
140 |
+
|
141 |
+
model, final_head = load_model(device)
|
142 |
+
transform = T.Normalize(mean=[0.485, 0.456, 0.406],
|
143 |
+
std=[0.229, 0.224, 0.225])
|
144 |
+
|
145 |
+
func = functools.partial(predict,
|
146 |
+
transform=transform,
|
147 |
+
device=device,
|
148 |
+
model=model,
|
149 |
+
final_head=final_head)
|
150 |
+
func = functools.update_wrapper(func, predict)
|
151 |
+
|
152 |
+
repo_url = 'https://github.com/ShuhongChen/bizarre-pose-estimator'
|
153 |
+
title = 'ShuhongChen/bizarre-pose-estimator (segmenter)'
|
154 |
+
description = f'A demo for {repo_url}'
|
155 |
+
article = None
|
156 |
+
|
157 |
+
gr.Interface(
|
158 |
+
func,
|
159 |
+
[
|
160 |
+
gr.inputs.Image(type='pil', label='Input'),
|
161 |
+
gr.inputs.Slider(0,
|
162 |
+
1,
|
163 |
+
step=args.score_slider_step,
|
164 |
+
default=args.score_threshold,
|
165 |
+
label='Score Threshold'),
|
166 |
+
],
|
167 |
+
gr.outputs.Image(label='Masked'),
|
168 |
+
theme=args.theme,
|
169 |
+
title=title,
|
170 |
+
description=description,
|
171 |
+
article=article,
|
172 |
+
examples=examples,
|
173 |
+
allow_screenshot=args.allow_screenshot,
|
174 |
+
allow_flagging=args.allow_flagging,
|
175 |
+
live=args.live,
|
176 |
+
).launch(
|
177 |
+
enable_queue=args.enable_queue,
|
178 |
+
server_port=args.port,
|
179 |
+
share=args.share,
|
180 |
+
)
|
181 |
+
|
182 |
+
|
183 |
+
if __name__ == '__main__':
|
184 |
+
main()
|
bizarre-pose-estimator
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
Subproject commit 7382ec234fa40cd8a6ec4a28b4639209199bc035
|
requirements.txt
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
torch>=1.10.1
|
2 |
+
torchvision>=0.11.2
|