Spaces:
Running
on
Zero
Running
on
Zero
Commit
·
98923c8
1
Parent(s):
14a932e
Update
Browse files
app.py
CHANGED
@@ -31,94 +31,91 @@ def preprocess_image(image):
|
|
31 |
image = image.astype(np.float32)
|
32 |
return image
|
33 |
|
34 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
35 |
|
36 |
|
37 |
class webui:
|
38 |
def __init__(self):
|
39 |
self.demo = gr.Blocks()
|
40 |
|
41 |
-
@spaces.GPU
|
42 |
-
def process_image(self, image_path, input_name, ort_sess, rating_tags, character_tags, general_tags):
|
43 |
-
thresh = 0.35
|
44 |
-
try:
|
45 |
-
image = Image.open(image_path)
|
46 |
-
image = image.convert("RGB") if image.mode != "RGB" else image
|
47 |
-
image = preprocess_image(image)
|
48 |
-
except Exception as e:
|
49 |
-
print(f"画像を読み込めません: {image_path}, エラー: {e}")
|
50 |
-
return
|
51 |
-
|
52 |
-
img = np.array([image])
|
53 |
-
prob = ort_sess.run(None, {input_name: img})[0][0] # ONNXモデルからの出力
|
54 |
-
|
55 |
-
# NSFW/SFW判定
|
56 |
-
tag_confidences = {tag: prob[i] for i, tag in enumerate(rating_tags)}
|
57 |
-
max_nsfw_score = max(tag_confidences.get("questionable", 0), tag_confidences.get("explicit", 0))
|
58 |
-
max_sfw_score = tag_confidences.get("general", 0)
|
59 |
-
NSFW_flag = None
|
60 |
-
|
61 |
-
if max_nsfw_score > max_sfw_score:
|
62 |
-
NSFW_flag = "NSFWの可能性が高いです"
|
63 |
-
else:
|
64 |
-
NSFW_flag = "SFWの可能性が高いです"
|
65 |
-
|
66 |
-
# 版権キャラクターの可能性を評価
|
67 |
-
character_tags_with_probs = []
|
68 |
-
for i, p in enumerate(prob[4:]):
|
69 |
-
if p >= thresh and i >= len(general_tags):
|
70 |
-
tag_index = i - len(general_tags)
|
71 |
-
if tag_index < len(character_tags):
|
72 |
-
tag_name = character_tags[tag_index]
|
73 |
-
prob_percent = round(p * 100, 2) # 確率をパーセンテージに変換
|
74 |
-
character_tags_with_probs.append((tag_name, f"{prob_percent}%"))
|
75 |
-
|
76 |
-
IP_flag = None
|
77 |
-
if character_tags_with_probs:
|
78 |
-
IP_flag = f"版権キャラクター: {character_tags_with_probs}の可能性があります"
|
79 |
-
else:
|
80 |
-
IP_flag = "版権キャラクターの可能性が低いと思われます"
|
81 |
-
|
82 |
-
# タグを生成
|
83 |
-
tag_freq = {}
|
84 |
-
undesired_tags = []
|
85 |
-
combined_tags = []
|
86 |
-
general_tag_text = ""
|
87 |
-
character_tag_text = ""
|
88 |
-
remove_underscore = True
|
89 |
-
caption_separator = ", "
|
90 |
-
general_threshold = 0.35
|
91 |
-
character_threshold = 0.35
|
92 |
-
|
93 |
-
for i, p in enumerate(prob[4:]):
|
94 |
-
if i < len(general_tags) and p >= general_threshold:
|
95 |
-
tag_name = general_tags[i]
|
96 |
-
if remove_underscore and len(tag_name) > 3: # ignore emoji tags like >_< and ^_^
|
97 |
-
tag_name = tag_name.replace("_", " ")
|
98 |
-
|
99 |
-
if tag_name not in undesired_tags:
|
100 |
-
tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
|
101 |
-
general_tag_text += caption_separator + tag_name
|
102 |
-
combined_tags.append(tag_name)
|
103 |
-
elif i >= len(general_tags) and p >= character_threshold:
|
104 |
-
tag_name = character_tags[i - len(general_tags)]
|
105 |
-
if remove_underscore and len(tag_name) > 3:
|
106 |
-
tag_name = tag_name.replace("_", " ")
|
107 |
-
|
108 |
-
if tag_name not in undesired_tags:
|
109 |
-
tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
|
110 |
-
character_tag_text += caption_separator + tag_name
|
111 |
-
combined_tags.append(tag_name)
|
112 |
-
|
113 |
-
# 先頭のカンマを取る
|
114 |
-
if len(general_tag_text) > 0:
|
115 |
-
general_tag_text = general_tag_text[len(caption_separator) :]
|
116 |
-
if len(character_tag_text) > 0:
|
117 |
-
character_tag_text = character_tag_text[len(caption_separator) :]
|
118 |
-
tag_text = caption_separator.join(combined_tags)
|
119 |
-
|
120 |
-
return NSFW_flag, IP_flag, tag_text
|
121 |
-
|
122 |
@spaces.GPU
|
123 |
def main(self, image_path, model_id):
|
124 |
print("Hugging Faceからモデルをダウンロード中")
|
@@ -144,10 +141,6 @@ class webui:
|
|
144 |
return NSFW_flag, IP_flag, tag_text
|
145 |
|
146 |
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
def launch(self):
|
152 |
with self.demo:
|
153 |
with gr.Row():
|
|
|
31 |
image = image.astype(np.float32)
|
32 |
return image
|
33 |
|
34 |
+
def process_image(image_path, input_name, ort_sess, rating_tags, character_tags, general_tags):
|
35 |
+
thresh = 0.35
|
36 |
+
try:
|
37 |
+
image = Image.open(image_path)
|
38 |
+
image = image.convert("RGB") if image.mode != "RGB" else image
|
39 |
+
image = preprocess_image(image)
|
40 |
+
except Exception as e:
|
41 |
+
print(f"画像を読み込めません: {image_path}, エラー: {e}")
|
42 |
+
return
|
43 |
+
|
44 |
+
img = np.array([image])
|
45 |
+
prob = ort_sess.run(None, {input_name: img})[0][0] # ONNXモデルからの出力
|
46 |
+
|
47 |
+
# NSFW/SFW判定
|
48 |
+
tag_confidences = {tag: prob[i] for i, tag in enumerate(rating_tags)}
|
49 |
+
max_nsfw_score = max(tag_confidences.get("questionable", 0), tag_confidences.get("explicit", 0))
|
50 |
+
max_sfw_score = tag_confidences.get("general", 0)
|
51 |
+
NSFW_flag = None
|
52 |
+
|
53 |
+
if max_nsfw_score > max_sfw_score:
|
54 |
+
NSFW_flag = "NSFWの可能性が高いです"
|
55 |
+
else:
|
56 |
+
NSFW_flag = "SFWの可能性が高いです"
|
57 |
+
|
58 |
+
# 版権キャラクターの可能性を評価
|
59 |
+
character_tags_with_probs = []
|
60 |
+
for i, p in enumerate(prob[4:]):
|
61 |
+
if p >= thresh and i >= len(general_tags):
|
62 |
+
tag_index = i - len(general_tags)
|
63 |
+
if tag_index < len(character_tags):
|
64 |
+
tag_name = character_tags[tag_index]
|
65 |
+
prob_percent = round(p * 100, 2) # 確率をパーセンテージに変換
|
66 |
+
character_tags_with_probs.append((tag_name, f"{prob_percent}%"))
|
67 |
+
|
68 |
+
IP_flag = None
|
69 |
+
if character_tags_with_probs:
|
70 |
+
IP_flag = f"版権キャラクター: {character_tags_with_probs}の可能性があります"
|
71 |
+
else:
|
72 |
+
IP_flag = "版権キャラクターの可能性が低いと思われます"
|
73 |
+
|
74 |
+
# タグを生成
|
75 |
+
tag_freq = {}
|
76 |
+
undesired_tags = []
|
77 |
+
combined_tags = []
|
78 |
+
general_tag_text = ""
|
79 |
+
character_tag_text = ""
|
80 |
+
remove_underscore = True
|
81 |
+
caption_separator = ", "
|
82 |
+
general_threshold = 0.35
|
83 |
+
character_threshold = 0.35
|
84 |
+
|
85 |
+
for i, p in enumerate(prob[4:]):
|
86 |
+
if i < len(general_tags) and p >= general_threshold:
|
87 |
+
tag_name = general_tags[i]
|
88 |
+
if remove_underscore and len(tag_name) > 3: # ignore emoji tags like >_< and ^_^
|
89 |
+
tag_name = tag_name.replace("_", " ")
|
90 |
+
|
91 |
+
if tag_name not in undesired_tags:
|
92 |
+
tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
|
93 |
+
general_tag_text += caption_separator + tag_name
|
94 |
+
combined_tags.append(tag_name)
|
95 |
+
elif i >= len(general_tags) and p >= character_threshold:
|
96 |
+
tag_name = character_tags[i - len(general_tags)]
|
97 |
+
if remove_underscore and len(tag_name) > 3:
|
98 |
+
tag_name = tag_name.replace("_", " ")
|
99 |
+
|
100 |
+
if tag_name not in undesired_tags:
|
101 |
+
tag_freq[tag_name] = tag_freq.get(tag_name, 0) + 1
|
102 |
+
character_tag_text += caption_separator + tag_name
|
103 |
+
combined_tags.append(tag_name)
|
104 |
+
|
105 |
+
# 先頭のカンマを取る
|
106 |
+
if len(general_tag_text) > 0:
|
107 |
+
general_tag_text = general_tag_text[len(caption_separator) :]
|
108 |
+
if len(character_tag_text) > 0:
|
109 |
+
character_tag_text = character_tag_text[len(caption_separator) :]
|
110 |
+
tag_text = caption_separator.join(combined_tags)
|
111 |
+
|
112 |
+
return NSFW_flag, IP_flag, tag_text
|
113 |
|
114 |
|
115 |
class webui:
|
116 |
def __init__(self):
|
117 |
self.demo = gr.Blocks()
|
118 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
119 |
@spaces.GPU
|
120 |
def main(self, image_path, model_id):
|
121 |
print("Hugging Faceからモデルをダウンロード中")
|
|
|
141 |
return NSFW_flag, IP_flag, tag_text
|
142 |
|
143 |
|
|
|
|
|
|
|
|
|
144 |
def launch(self):
|
145 |
with self.demo:
|
146 |
with gr.Row():
|