Spaces:
Running
Running
Upload 2 files
#2
by
John6666
- opened
README.md
CHANGED
@@ -1,13 +1,13 @@
|
|
1 |
-
---
|
2 |
-
title: SigLIP Tagger
|
3 |
-
emoji: 🧷
|
4 |
-
colorFrom: green
|
5 |
-
colorTo: blue
|
6 |
-
sdk: gradio
|
7 |
-
sdk_version: 4.
|
8 |
-
app_file: app.py
|
9 |
-
pinned: true
|
10 |
-
license: apache-2.0
|
11 |
-
---
|
12 |
-
|
13 |
-
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
|
|
1 |
+
---
|
2 |
+
title: SigLIP Tagger
|
3 |
+
emoji: 🧷
|
4 |
+
colorFrom: green
|
5 |
+
colorTo: blue
|
6 |
+
sdk: gradio
|
7 |
+
sdk_version: 4.43.0
|
8 |
+
app_file: app.py
|
9 |
+
pinned: true
|
10 |
+
license: apache-2.0
|
11 |
+
---
|
12 |
+
|
13 |
+
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
app.py
CHANGED
@@ -1,164 +1,164 @@
|
|
1 |
-
import os
|
2 |
-
from PIL import Image
|
3 |
-
|
4 |
-
import numpy as np
|
5 |
-
import torch
|
6 |
-
|
7 |
-
from transformers import (
|
8 |
-
AutoImageProcessor,
|
9 |
-
)
|
10 |
-
|
11 |
-
import gradio as gr
|
12 |
-
|
13 |
-
from modeling_siglip import SiglipForImageClassification
|
14 |
-
|
15 |
-
|
16 |
-
HF_TOKEN = os.environ
|
17 |
-
|
18 |
-
EXAMPLES = [["./images/sample.jpg"], ["./images/sample2.webp"]]
|
19 |
-
|
20 |
-
model_maps: dict[str, dict] = {
|
21 |
-
"test2": {
|
22 |
-
"repo": "p1atdev/siglip-tagger-test-2",
|
23 |
-
},
|
24 |
-
"test3": {
|
25 |
-
"repo": "p1atdev/siglip-tagger-test-3",
|
26 |
-
},
|
27 |
-
# "test4": {
|
28 |
-
# "repo": "p1atdev/siglip-tagger-test-4",
|
29 |
-
# },
|
30 |
-
}
|
31 |
-
|
32 |
-
for key in model_maps.keys():
|
33 |
-
model_maps[key]["model"] = SiglipForImageClassification.from_pretrained(
|
34 |
-
model_maps[key]["repo"], torch_dtype=torch.bfloat16, token=HF_TOKEN
|
35 |
-
)
|
36 |
-
model_maps[key]["processor"] = AutoImageProcessor.from_pretrained(
|
37 |
-
model_maps[key]["repo"], token=HF_TOKEN
|
38 |
-
)
|
39 |
-
|
40 |
-
README_MD = (
|
41 |
-
f"""\
|
42 |
-
## SigLIP Tagger Test 3
|
43 |
-
An experimental model for tagging danbooru tags of images using SigLIP.
|
44 |
-
|
45 |
-
Model(s):
|
46 |
-
"""
|
47 |
-
+ "\n".join(
|
48 |
-
f"- [{value['repo']}](https://huggingface.co/{value['repo']})"
|
49 |
-
for value in model_maps.values()
|
50 |
-
)
|
51 |
-
+ "\n"
|
52 |
-
+ """
|
53 |
-
Example images by NovelAI and niji・journey.
|
54 |
-
"""
|
55 |
-
)
|
56 |
-
|
57 |
-
|
58 |
-
def compose_text(results: dict[str, float], threshold: float = 0.3):
|
59 |
-
return ", ".join(
|
60 |
-
[
|
61 |
-
key
|
62 |
-
for key, value in sorted(results.items(), key=lambda x: x[1], reverse=True)
|
63 |
-
if value > threshold
|
64 |
-
]
|
65 |
-
)
|
66 |
-
|
67 |
-
|
68 |
-
@torch.no_grad()
|
69 |
-
def predict_tags(image: Image.Image, model_name: str, threshold: float):
|
70 |
-
if image is None:
|
71 |
-
return None, None
|
72 |
-
|
73 |
-
inputs = model_maps[model_name]["processor"](image, return_tensors="pt")
|
74 |
-
|
75 |
-
logits = (
|
76 |
-
model_maps[model_name]["model"](
|
77 |
-
**inputs.to(
|
78 |
-
model_maps[model_name]["model"].device,
|
79 |
-
model_maps[model_name]["model"].dtype,
|
80 |
-
)
|
81 |
-
)
|
82 |
-
.logits.detach()
|
83 |
-
.cpu()
|
84 |
-
.float()
|
85 |
-
)
|
86 |
-
|
87 |
-
logits = np.clip(logits, 0.0, 1.0)
|
88 |
-
|
89 |
-
results = {}
|
90 |
-
|
91 |
-
for prediction in logits:
|
92 |
-
for i, prob in enumerate(prediction):
|
93 |
-
if prob.item() > 0:
|
94 |
-
results[model_maps[model_name]["model"].config.id2label[i]] = (
|
95 |
-
prob.item()
|
96 |
-
)
|
97 |
-
|
98 |
-
return compose_text(results, threshold), results
|
99 |
-
|
100 |
-
|
101 |
-
css = """\
|
102 |
-
.sticky {
|
103 |
-
position: sticky;
|
104 |
-
top: 16px;
|
105 |
-
}
|
106 |
-
|
107 |
-
.gradio-container {
|
108 |
-
overflow: clip;
|
109 |
-
}
|
110 |
-
"""
|
111 |
-
|
112 |
-
|
113 |
-
def demo():
|
114 |
-
with gr.Blocks(css=css) as ui:
|
115 |
-
gr.Markdown(README_MD)
|
116 |
-
|
117 |
-
with gr.Row():
|
118 |
-
with gr.Column():
|
119 |
-
with gr.Row(elem_classes="sticky"):
|
120 |
-
with gr.Column():
|
121 |
-
input_img = gr.Image(
|
122 |
-
label="Input image", type="pil", height=480
|
123 |
-
)
|
124 |
-
|
125 |
-
with gr.Group():
|
126 |
-
model_name_radio = gr.Radio(
|
127 |
-
label="Model",
|
128 |
-
choices=list(model_maps.keys()),
|
129 |
-
value="test3",
|
130 |
-
)
|
131 |
-
tag_threshold_slider = gr.Slider(
|
132 |
-
label="Tags threshold",
|
133 |
-
minimum=0.0,
|
134 |
-
maximum=1.0,
|
135 |
-
value=0.3,
|
136 |
-
step=0.01,
|
137 |
-
)
|
138 |
-
|
139 |
-
start_btn = gr.Button(value="Start", variant="primary")
|
140 |
-
|
141 |
-
gr.Examples(
|
142 |
-
examples=EXAMPLES,
|
143 |
-
inputs=[input_img],
|
144 |
-
cache_examples=False,
|
145 |
-
)
|
146 |
-
|
147 |
-
with gr.Column():
|
148 |
-
output_tags = gr.Text(label="Output text", interactive=False)
|
149 |
-
output_label = gr.Label(label="Output tags")
|
150 |
-
|
151 |
-
start_btn.click(
|
152 |
-
fn=predict_tags,
|
153 |
-
inputs=[input_img, model_name_radio, tag_threshold_slider],
|
154 |
-
outputs=[output_tags, output_label],
|
155 |
-
)
|
156 |
-
|
157 |
-
ui.launch(
|
158 |
-
debug=True,
|
159 |
-
# share=True
|
160 |
-
)
|
161 |
-
|
162 |
-
|
163 |
-
if __name__ == "__main__":
|
164 |
-
demo()
|
|
|
1 |
+
import os
|
2 |
+
from PIL import Image
|
3 |
+
|
4 |
+
import numpy as np
|
5 |
+
import torch
|
6 |
+
|
7 |
+
from transformers import (
|
8 |
+
AutoImageProcessor,
|
9 |
+
)
|
10 |
+
|
11 |
+
import gradio as gr
|
12 |
+
|
13 |
+
from modeling_siglip import SiglipForImageClassification
|
14 |
+
|
15 |
+
|
16 |
+
HF_TOKEN = os.environ.get("HF_READ_TOKEN")
|
17 |
+
|
18 |
+
EXAMPLES = [["./images/sample.jpg"], ["./images/sample2.webp"]]
|
19 |
+
|
20 |
+
model_maps: dict[str, dict] = {
|
21 |
+
"test2": {
|
22 |
+
"repo": "p1atdev/siglip-tagger-test-2",
|
23 |
+
},
|
24 |
+
"test3": {
|
25 |
+
"repo": "p1atdev/siglip-tagger-test-3",
|
26 |
+
},
|
27 |
+
# "test4": {
|
28 |
+
# "repo": "p1atdev/siglip-tagger-test-4",
|
29 |
+
# },
|
30 |
+
}
|
31 |
+
|
32 |
+
for key in model_maps.keys():
|
33 |
+
model_maps[key]["model"] = SiglipForImageClassification.from_pretrained(
|
34 |
+
model_maps[key]["repo"], torch_dtype=torch.bfloat16, token=HF_TOKEN
|
35 |
+
)
|
36 |
+
model_maps[key]["processor"] = AutoImageProcessor.from_pretrained(
|
37 |
+
model_maps[key]["repo"], token=HF_TOKEN
|
38 |
+
)
|
39 |
+
|
40 |
+
README_MD = (
|
41 |
+
f"""\
|
42 |
+
## SigLIP Tagger Test 3
|
43 |
+
An experimental model for tagging danbooru tags of images using SigLIP.
|
44 |
+
|
45 |
+
Model(s):
|
46 |
+
"""
|
47 |
+
+ "\n".join(
|
48 |
+
f"- [{value['repo']}](https://huggingface.co/{value['repo']})"
|
49 |
+
for value in model_maps.values()
|
50 |
+
)
|
51 |
+
+ "\n"
|
52 |
+
+ """
|
53 |
+
Example images by NovelAI and niji・journey.
|
54 |
+
"""
|
55 |
+
)
|
56 |
+
|
57 |
+
|
58 |
+
def compose_text(results: dict[str, float], threshold: float = 0.3):
|
59 |
+
return ", ".join(
|
60 |
+
[
|
61 |
+
key
|
62 |
+
for key, value in sorted(results.items(), key=lambda x: x[1], reverse=True)
|
63 |
+
if value > threshold
|
64 |
+
]
|
65 |
+
)
|
66 |
+
|
67 |
+
|
68 |
+
@torch.no_grad()
|
69 |
+
def predict_tags(image: Image.Image, model_name: str, threshold: float):
|
70 |
+
if image is None:
|
71 |
+
return None, None
|
72 |
+
|
73 |
+
inputs = model_maps[model_name]["processor"](image, return_tensors="pt")
|
74 |
+
|
75 |
+
logits = (
|
76 |
+
model_maps[model_name]["model"](
|
77 |
+
**inputs.to(
|
78 |
+
model_maps[model_name]["model"].device,
|
79 |
+
model_maps[model_name]["model"].dtype,
|
80 |
+
)
|
81 |
+
)
|
82 |
+
.logits.detach()
|
83 |
+
.cpu()
|
84 |
+
.float()
|
85 |
+
)
|
86 |
+
|
87 |
+
logits = np.clip(logits, 0.0, 1.0)
|
88 |
+
|
89 |
+
results = {}
|
90 |
+
|
91 |
+
for prediction in logits:
|
92 |
+
for i, prob in enumerate(prediction):
|
93 |
+
if prob.item() > 0:
|
94 |
+
results[model_maps[model_name]["model"].config.id2label[i]] = (
|
95 |
+
prob.item()
|
96 |
+
)
|
97 |
+
|
98 |
+
return compose_text(results, threshold), results
|
99 |
+
|
100 |
+
|
101 |
+
css = """\
|
102 |
+
.sticky {
|
103 |
+
position: sticky;
|
104 |
+
top: 16px;
|
105 |
+
}
|
106 |
+
|
107 |
+
.gradio-container {
|
108 |
+
overflow: clip;
|
109 |
+
}
|
110 |
+
"""
|
111 |
+
|
112 |
+
|
113 |
+
def demo():
|
114 |
+
with gr.Blocks(css=css) as ui:
|
115 |
+
gr.Markdown(README_MD)
|
116 |
+
|
117 |
+
with gr.Row():
|
118 |
+
with gr.Column():
|
119 |
+
with gr.Row(elem_classes="sticky"):
|
120 |
+
with gr.Column():
|
121 |
+
input_img = gr.Image(
|
122 |
+
label="Input image", type="pil", height=480
|
123 |
+
)
|
124 |
+
|
125 |
+
with gr.Group():
|
126 |
+
model_name_radio = gr.Radio(
|
127 |
+
label="Model",
|
128 |
+
choices=list(model_maps.keys()),
|
129 |
+
value="test3",
|
130 |
+
)
|
131 |
+
tag_threshold_slider = gr.Slider(
|
132 |
+
label="Tags threshold",
|
133 |
+
minimum=0.0,
|
134 |
+
maximum=1.0,
|
135 |
+
value=0.3,
|
136 |
+
step=0.01,
|
137 |
+
)
|
138 |
+
|
139 |
+
start_btn = gr.Button(value="Start", variant="primary")
|
140 |
+
|
141 |
+
gr.Examples(
|
142 |
+
examples=EXAMPLES,
|
143 |
+
inputs=[input_img],
|
144 |
+
cache_examples=False,
|
145 |
+
)
|
146 |
+
|
147 |
+
with gr.Column():
|
148 |
+
output_tags = gr.Text(label="Output text", interactive=False)
|
149 |
+
output_label = gr.Label(label="Output tags")
|
150 |
+
|
151 |
+
start_btn.click(
|
152 |
+
fn=predict_tags,
|
153 |
+
inputs=[input_img, model_name_radio, tag_threshold_slider],
|
154 |
+
outputs=[output_tags, output_label],
|
155 |
+
)
|
156 |
+
|
157 |
+
ui.launch(
|
158 |
+
debug=True,
|
159 |
+
# share=True
|
160 |
+
)
|
161 |
+
|
162 |
+
|
163 |
+
if __name__ == "__main__":
|
164 |
+
demo()
|