Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -5,120 +5,165 @@ from PIL import Image
|
|
5 |
import json
|
6 |
from huggingface_hub import hf_hub_download
|
7 |
|
8 |
-
#
|
9 |
MODEL_REPO = "AngelBottomless/camie-tagger-onnxruntime"
|
10 |
MODEL_FILE = "camie_tagger_initial.onnx"
|
11 |
META_FILE = "metadata.json"
|
|
|
|
|
|
|
|
|
12 |
model_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE, cache_dir=".")
|
13 |
meta_path = hf_hub_download(repo_id=MODEL_REPO, filename=META_FILE, cache_dir=".")
|
|
|
|
|
14 |
session = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])
|
15 |
-
|
16 |
-
|
|
|
|
|
|
|
|
|
|
|
17 |
def preprocess_image(pil_image: Image.Image) -> np.ndarray:
|
18 |
-
|
|
|
19 |
arr = np.array(img).astype(np.float32) / 255.0
|
20 |
arr = np.transpose(arr, (2, 0, 1))
|
21 |
-
|
22 |
-
return arr
|
23 |
|
24 |
-
|
25 |
-
|
26 |
-
|
|
|
|
|
|
|
27 |
input_tensor = preprocess_image(pil_image)
|
28 |
input_name = session.get_inputs()[0].name
|
29 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
probs = 1 / (1 + np.exp(-refined_logits))
|
31 |
-
probs = probs[0]
|
32 |
idx_to_tag = metadata["idx_to_tag"]
|
33 |
tag_to_category = metadata.get("tag_to_category", {})
|
34 |
category_thresholds = metadata.get("category_thresholds", {})
|
35 |
-
|
36 |
-
results_by_cat = {}
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
all_artist_tags_probs = [] # Store all artist tags and their probabilities
|
41 |
-
|
42 |
-
# Collect tags above thresholds
|
43 |
for idx, prob in enumerate(probs):
|
44 |
tag = idx_to_tag[str(idx)]
|
45 |
cat = tag_to_category.get(tag, "unknown")
|
46 |
-
if cat == 'artist':
|
47 |
-
all_artist_tags_probs.append((tag, float(prob))) # Store all artist tags
|
48 |
thresh = category_thresholds.get(cat, default_threshold)
|
|
|
|
|
49 |
if float(prob) >= thresh:
|
50 |
-
# add to category dictionary
|
51 |
results_by_cat.setdefault(cat, []).append((tag, float(prob)))
|
52 |
-
if cat
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
59 |
if output_format == "Prompt-style Tags":
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
artist_prompt_tags = [tag.replace("_", " ").replace("(", r"\\(").replace(")", r"\\)") for tag, prob in artist_tags_with_probs]
|
65 |
-
character_prompt_tags = [tag.replace("_", " ").replace("(", r"\\(").replace(")", r"\\)") for tag, prob in character_tags_with_probs]
|
66 |
-
general_prompt_tags = [tag.replace("_", " ").replace("(", r"\\(").replace(")", r"\\)") for tag, prob in general_tags_with_probs]
|
67 |
-
|
68 |
-
prompt_tags = artist_prompt_tags + character_prompt_tags + general_prompt_tags
|
69 |
-
|
70 |
-
# Ensure at least one artist tag if any artist tags were predicted at all, even below threshold
|
71 |
-
if not artist_prompt_tags and all_artist_tags_probs:
|
72 |
-
best_artist_tag, best_artist_prob = max(all_artist_tags_probs, key=lambda item: item[1])
|
73 |
-
prompt_tags = [best_artist_tag.replace("_", " ").replace("(", r"\\(").replace(")", r"\\)")] + prompt_tags
|
74 |
-
|
75 |
-
|
76 |
-
if not prompt_tags:
|
77 |
-
return "No tags predicted."
|
78 |
-
return ", ".join(prompt_tags)
|
79 |
-
else: # Detailed output
|
80 |
-
if not results_by_cat:
|
81 |
-
return "No tags predicted for this image."
|
82 |
-
|
83 |
-
# Ensure artist tag in detailed output even if below threshold
|
84 |
-
if 'artist' not in results_by_cat and all_artist_tags_probs:
|
85 |
-
best_artist_tag, best_artist_prob = max(all_artist_tags_probs, key=lambda item: item[1])
|
86 |
-
results_by_cat['artist'] = [(best_artist_tag, best_artist_prob)]
|
87 |
-
|
88 |
-
|
89 |
-
lines = []
|
90 |
-
lines.append("**Predicted Tags by Category:** \n") # (Markdown newline: two spaces + newline)
|
91 |
-
for cat, tag_list in results_by_cat.items():
|
92 |
-
# sort tags in this category by probability descending
|
93 |
-
tag_list.sort(key=lambda x: x[1], reverse=True)
|
94 |
-
lines.append(f"**Category: {cat}** – {len(tag_list)} tags")
|
95 |
-
for tag, prob in tag_list:
|
96 |
-
tag_pretty = tag.replace("_", " ").replace("(", r"\\(").replace(")", r"\\)") # Escape parentheses here with raw string
|
97 |
-
lines.append(f"- {tag_pretty} (Prob: {prob:.3f})")
|
98 |
-
lines.append("") # blank line between categories
|
99 |
-
return "\n".join(lines)
|
100 |
|
101 |
# Build the Gradio Blocks UI
|
102 |
-
demo = gr.Blocks(theme="gradio/soft")
|
103 |
|
104 |
with demo:
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
|
|
|
|
|
|
|
|
|
|
109 |
with gr.Row():
|
110 |
-
# Left column: Image input and format selection
|
111 |
with gr.Column():
|
112 |
image_in = gr.Image(type="pil", label="Input Image")
|
113 |
-
format_choice = gr.Radio(
|
|
|
|
|
|
|
|
|
114 |
tag_button = gr.Button("🔍 Tag Image")
|
115 |
-
# Right column: Output display
|
116 |
with gr.Column():
|
117 |
-
output_box = gr.Markdown("") #
|
118 |
-
|
119 |
tag_button.click(fn=tag_image, inputs=[image_in, format_choice], outputs=output_box)
|
120 |
-
|
121 |
-
gr.Markdown(
|
|
|
|
|
|
|
|
|
|
|
122 |
|
123 |
-
|
124 |
-
demo.launch()
|
|
|
5 |
import json
|
6 |
from huggingface_hub import hf_hub_download
|
7 |
|
8 |
+
# Constants
|
9 |
MODEL_REPO = "AngelBottomless/camie-tagger-onnxruntime"
|
10 |
MODEL_FILE = "camie_tagger_initial.onnx"
|
11 |
META_FILE = "metadata.json"
|
12 |
+
IMAGE_SIZE = (512, 512)
|
13 |
+
DEFAULT_THRESHOLD = 0.35
|
14 |
+
|
15 |
+
# Download model and metadata from Hugging Face Hub
|
16 |
model_path = hf_hub_download(repo_id=MODEL_REPO, filename=MODEL_FILE, cache_dir=".")
|
17 |
meta_path = hf_hub_download(repo_id=MODEL_REPO, filename=META_FILE, cache_dir=".")
|
18 |
+
|
19 |
+
# Initialize ONNX Runtime session and load metadata
|
20 |
session = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])
|
21 |
+
with open(meta_path, "r", encoding="utf-8") as f:
|
22 |
+
metadata = json.load(f)
|
23 |
+
|
24 |
+
def escape_tag(tag: str) -> str:
|
25 |
+
"""Escape underscores and parentheses for Markdown."""
|
26 |
+
return tag.replace("_", " ").replace("(", r"\\(").replace(")", r"\\)")
|
27 |
+
|
28 |
def preprocess_image(pil_image: Image.Image) -> np.ndarray:
|
29 |
+
"""Convert image to RGB, resize, normalize, and rearrange dimensions."""
|
30 |
+
img = pil_image.convert("RGB").resize(IMAGE_SIZE)
|
31 |
arr = np.array(img).astype(np.float32) / 255.0
|
32 |
arr = np.transpose(arr, (2, 0, 1))
|
33 |
+
return np.expand_dims(arr, 0)
|
|
|
34 |
|
35 |
+
def run_inference(pil_image: Image.Image) -> np.ndarray:
|
36 |
+
"""
|
37 |
+
Preprocess the image and run the ONNX model inference.
|
38 |
+
|
39 |
+
Returns the refined logits as a numpy array.
|
40 |
+
"""
|
41 |
input_tensor = preprocess_image(pil_image)
|
42 |
input_name = session.get_inputs()[0].name
|
43 |
+
# Only refined_logits are used (initial_logits is ignored)
|
44 |
+
_, refined_logits = session.run(None, {input_name: input_tensor})
|
45 |
+
return refined_logits[0]
|
46 |
+
|
47 |
+
def get_tags(refined_logits: np.ndarray, metadata: dict, default_threshold: float = DEFAULT_THRESHOLD):
|
48 |
+
"""
|
49 |
+
Compute probabilities from logits and collect tag predictions.
|
50 |
+
|
51 |
+
Returns:
|
52 |
+
results_by_cat: Dictionary mapping each category to a list of (tag, probability) above its threshold.
|
53 |
+
prompt_tags_by_cat: Similar dictionary but only for prompt-style categories (artist, character, general).
|
54 |
+
all_artist_tags: All artist tags (with probabilities) regardless of threshold.
|
55 |
+
"""
|
56 |
probs = 1 / (1 + np.exp(-refined_logits))
|
|
|
57 |
idx_to_tag = metadata["idx_to_tag"]
|
58 |
tag_to_category = metadata.get("tag_to_category", {})
|
59 |
category_thresholds = metadata.get("category_thresholds", {})
|
60 |
+
|
61 |
+
results_by_cat = {}
|
62 |
+
prompt_tags_by_cat = {"artist": [], "character": [], "general": []}
|
63 |
+
all_artist_tags = []
|
64 |
+
|
|
|
|
|
|
|
65 |
for idx, prob in enumerate(probs):
|
66 |
tag = idx_to_tag[str(idx)]
|
67 |
cat = tag_to_category.get(tag, "unknown")
|
|
|
|
|
68 |
thresh = category_thresholds.get(cat, default_threshold)
|
69 |
+
if cat == "artist":
|
70 |
+
all_artist_tags.append((tag, float(prob)))
|
71 |
if float(prob) >= thresh:
|
|
|
72 |
results_by_cat.setdefault(cat, []).append((tag, float(prob)))
|
73 |
+
if cat in prompt_tags_by_cat:
|
74 |
+
prompt_tags_by_cat[cat].append((tag, float(prob)))
|
75 |
+
return results_by_cat, prompt_tags_by_cat, all_artist_tags
|
76 |
+
|
77 |
+
def format_prompt_tags(prompt_tags_by_cat: dict, all_artist_tags: list) -> str:
|
78 |
+
"""
|
79 |
+
Format the tags for prompt-style output.
|
80 |
+
|
81 |
+
Returns a comma-separated string of escaped tags.
|
82 |
+
"""
|
83 |
+
# Sort tags within each category by probability (descending)
|
84 |
+
for cat in prompt_tags_by_cat:
|
85 |
+
prompt_tags_by_cat[cat].sort(key=lambda x: x[1], reverse=True)
|
86 |
+
|
87 |
+
artist_tags = [escape_tag(tag) for tag, _ in prompt_tags_by_cat.get("artist", [])]
|
88 |
+
character_tags = [escape_tag(tag) for tag, _ in prompt_tags_by_cat.get("character", [])]
|
89 |
+
general_tags = [escape_tag(tag) for tag, _ in prompt_tags_by_cat.get("general", [])]
|
90 |
+
prompt_tags = artist_tags + character_tags + general_tags
|
91 |
+
|
92 |
+
# Ensure at least one artist tag appears if available, even if below threshold
|
93 |
+
if not artist_tags and all_artist_tags:
|
94 |
+
best_artist_tag, _ = max(all_artist_tags, key=lambda item: item[1])
|
95 |
+
prompt_tags.insert(0, escape_tag(best_artist_tag))
|
96 |
+
return ", ".join(prompt_tags) if prompt_tags else "No tags predicted."
|
97 |
+
|
98 |
+
def format_detailed_output(results_by_cat: dict, all_artist_tags: list) -> str:
|
99 |
+
"""
|
100 |
+
Format the tags for detailed output.
|
101 |
+
|
102 |
+
Returns a Markdown-formatted string listing tags by category.
|
103 |
+
"""
|
104 |
+
if not results_by_cat:
|
105 |
+
return "No tags predicted for this image."
|
106 |
|
107 |
+
# Include an artist tag even if below threshold
|
108 |
+
if "artist" not in results_by_cat and all_artist_tags:
|
109 |
+
best_artist_tag, best_artist_prob = max(all_artist_tags, key=lambda item: item[1])
|
110 |
+
results_by_cat["artist"] = [(best_artist_tag, best_artist_prob)]
|
111 |
+
|
112 |
+
lines = ["**Predicted Tags by Category:** \n"]
|
113 |
+
for cat, tag_list in results_by_cat.items():
|
114 |
+
tag_list.sort(key=lambda x: x[1], reverse=True)
|
115 |
+
lines.append(f"**Category: {cat}** – {len(tag_list)} tags")
|
116 |
+
for tag, prob in tag_list:
|
117 |
+
lines.append(f"- {escape_tag(tag)} (Prob: {prob:.3f})")
|
118 |
+
lines.append("") # blank line between categories
|
119 |
+
return "\n".join(lines)
|
120 |
+
|
121 |
+
def tag_image(pil_image: Image.Image, output_format: str) -> str:
|
122 |
+
"""Run inference on the image and return formatted tags based on the chosen output format."""
|
123 |
+
if pil_image is None:
|
124 |
+
return "Please upload an image."
|
125 |
+
|
126 |
+
refined_logits = run_inference(pil_image)
|
127 |
+
results_by_cat, prompt_tags_by_cat, all_artist_tags = get_tags(refined_logits, metadata)
|
128 |
+
|
129 |
if output_format == "Prompt-style Tags":
|
130 |
+
return format_prompt_tags(prompt_tags_by_cat, all_artist_tags)
|
131 |
+
else:
|
132 |
+
return format_detailed_output(results_by_cat, all_artist_tags)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
133 |
|
134 |
# Build the Gradio Blocks UI
|
135 |
+
demo = gr.Blocks(theme="gradio/soft")
|
136 |
|
137 |
with demo:
|
138 |
+
gr.Markdown(
|
139 |
+
"# 🏷️ Camie Tagger – Anime Image Tagging\n"
|
140 |
+
"This demo uses an ONNX model of Camie Tagger to label anime illustrations with tags. "
|
141 |
+
"Upload an image and click **Tag Image** to see predictions."
|
142 |
+
)
|
143 |
+
gr.Markdown(
|
144 |
+
"*(Note: The model will predict a large number of tags across categories like character, general, artist, etc. "
|
145 |
+
"You can choose a concise prompt-style output or a detailed category-wise breakdown.)*"
|
146 |
+
)
|
147 |
with gr.Row():
|
|
|
148 |
with gr.Column():
|
149 |
image_in = gr.Image(type="pil", label="Input Image")
|
150 |
+
format_choice = gr.Radio(
|
151 |
+
choices=["Prompt-style Tags", "Detailed Output"],
|
152 |
+
value="Prompt-style Tags",
|
153 |
+
label="Output Format"
|
154 |
+
)
|
155 |
tag_button = gr.Button("🔍 Tag Image")
|
|
|
156 |
with gr.Column():
|
157 |
+
output_box = gr.Markdown("") # Markdown output for formatted results
|
158 |
+
|
159 |
tag_button.click(fn=tag_image, inputs=[image_in, format_choice], outputs=output_box)
|
160 |
+
|
161 |
+
gr.Markdown(
|
162 |
+
"----\n"
|
163 |
+
"**Model:** [Camie Tagger ONNX](https://huggingface.co/AngelBottomless/camie-tagger-onnxruntime) • "
|
164 |
+
"**Base Model:** Camais03/camie-tagger (61% F1 on 70k tags) • **ONNX Runtime:** for efficient CPU inference • "
|
165 |
+
"*Demo built with Gradio Blocks.*"
|
166 |
+
)
|
167 |
|
168 |
+
if __name__ == "__main__":
|
169 |
+
demo.launch()
|