Spaces:
Running
on
Zero
Running
on
Zero
bugfix for diverse search with zero diversity_penalty
Browse files
app.py
CHANGED
@@ -26,30 +26,19 @@ MODEL_LIST = [
|
|
26 |
]
|
27 |
|
28 |
def merge_common_prefixes(suggestions, min_len=2):
|
29 |
-
"""
|
30 |
-
合併具有共同前綴的建議:
|
31 |
-
- 找出所有長度 ≥ min_len 的共同前綴
|
32 |
-
- 將這些前綴作為新建議,移除原有被合併的項目
|
33 |
-
"""
|
34 |
prefixes = []
|
35 |
to_remove = set()
|
36 |
-
|
37 |
for i in range(len(suggestions)):
|
38 |
for j in range(i+1, len(suggestions)):
|
39 |
s1, s2 = suggestions[i], suggestions[j]
|
40 |
-
# 計算字元級共同前綴
|
41 |
common = ''.join(c1 for c1, c2 in zip(s1, s2) if c1 == c2)
|
42 |
if len(common) >= min_len:
|
43 |
prefixes.append(common)
|
44 |
to_remove.update([s1, s2])
|
45 |
-
|
46 |
-
# 去重前綴
|
47 |
unique_prefixes = []
|
48 |
for p in prefixes:
|
49 |
if p not in unique_prefixes:
|
50 |
unique_prefixes.append(p)
|
51 |
-
|
52 |
-
# 剩下沒有被合併的建議
|
53 |
remainder = [s for s in suggestions if s not in to_remove]
|
54 |
return unique_prefixes + remainder
|
55 |
|
@@ -64,28 +53,29 @@ def get_pipeline(model_name):
|
|
64 |
|
65 |
@spaces.GPU
|
66 |
def suggest_next(text, model_name, k, m, num_beam_groups, diversity_penalty):
|
67 |
-
"""
|
68 |
-
使用 Diverse Beam Search 產生 m 條候選:
|
69 |
-
- num_beams = m
|
70 |
-
- num_beam_groups, diversity_penalty 可調整多樣性
|
71 |
-
之後轉繁體、去重、合併共同前綴後回傳。
|
72 |
-
"""
|
73 |
gen_pipe = get_pipeline(model_name)
|
74 |
-
|
75 |
-
|
76 |
-
max_new_tokens
|
77 |
-
num_beams
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
unique_suggestions = []
|
90 |
for s in suggestions:
|
91 |
if s not in unique_suggestions:
|
@@ -96,88 +86,18 @@ def suggest_next(text, model_name, k, m, num_beam_groups, diversity_penalty):
|
|
96 |
|
97 |
return update(choices=final_suggestions, value=None)
|
98 |
|
|
|
|
|
99 |
|
100 |
-
|
101 |
-
|
102 |
-
return current
|
103 |
-
# 直接插入選中的候選文字
|
104 |
-
return current + choice
|
105 |
-
|
106 |
-
# 自訂 CSS:模擬經典中文輸入法候選欄樣式,並優化手機響應與自動高度
|
107 |
-
custom_css = """
|
108 |
-
#suggestions-bar {
|
109 |
-
width: 100%;
|
110 |
-
margin-bottom: 8px;
|
111 |
-
}
|
112 |
#suggestions-bar .candidate-list {
|
113 |
-
display: flex;
|
114 |
-
|
115 |
-
|
116 |
-
border: 1px solid #999;
|
117 |
-
border-radius: 4px;
|
118 |
-
padding: 6px;
|
119 |
-
overflow-x: auto;
|
120 |
-
white-space: nowrap;
|
121 |
-
}
|
122 |
-
#suggestions-bar .candidate-list label {
|
123 |
-
cursor: pointer;
|
124 |
-
padding: 6px 10px;
|
125 |
-
font-size: 16px;
|
126 |
}
|
127 |
-
#suggestions-bar .candidate-list label:
|
128 |
-
|
129 |
-
}
|
130 |
-
#suggestions-bar .candidate-list input[type=radio]:checked + label {
|
131 |
-
background: #e6f7ff;
|
132 |
-
border: 1px solid #1890ff;
|
133 |
-
}
|
134 |
-
#input-box textarea {
|
135 |
-
width: 100%;
|
136 |
-
font-size: 16px;
|
137 |
-
padding: 6px;
|
138 |
-
box-sizing: border-box;
|
139 |
-
overflow: hidden;
|
140 |
-
resize: none;
|
141 |
-
}
|
142 |
-
#predict-button {
|
143 |
-
margin-top: 8px;
|
144 |
-
width: 100%;
|
145 |
-
}
|
146 |
-
/* 手機響應式 */
|
147 |
-
@media only screen and (max-width: 600px) {
|
148 |
-
#suggestions-bar .candidate-list label {
|
149 |
-
padding: 8px;
|
150 |
-
font-size: 18px;
|
151 |
-
}
|
152 |
-
#predict-button {
|
153 |
-
font-size: 18px;
|
154 |
-
}
|
155 |
-
}
|
156 |
-
"""
|
157 |
-
|
158 |
-
# 自動增高腳本
|
159 |
-
auto_height_js = """
|
160 |
-
<script>
|
161 |
-
window.addEventListener('load', () => {
|
162 |
-
const textarea = document.querySelector('#input-box textarea');
|
163 |
-
if (!textarea) return;
|
164 |
-
textarea.style.height = 'auto';
|
165 |
-
textarea.addEventListener('input', function() {
|
166 |
-
this.style.height = 'auto';
|
167 |
-
this.style.height = this.scrollHeight + 'px';
|
168 |
-
});
|
169 |
-
});
|
170 |
-
</script>
|
171 |
-
"""
|
172 |
-
|
173 |
-
with gr.Blocks(css=custom_css) as demo:
|
174 |
-
gr.HTML(auto_height_js)
|
175 |
-
gr.Markdown(
|
176 |
-
"## 🇹🇼 繁體中文 IME 加速器 \
|
177 |
-
"
|
178 |
-
"結合小型語言模型與 ZeroGPU,提供即時輸入法風格候選欄。"
|
179 |
-
)
|
180 |
-
|
181 |
with gr.Column():
|
182 |
suggestions = gr.Radio(
|
183 |
[], label="", interactive=True, type="value",
|
@@ -188,14 +108,11 @@ with gr.Blocks(css=custom_css) as demo:
|
|
188 |
lines=1, max_lines=20, elem_id="input-box"
|
189 |
)
|
190 |
|
191 |
-
# 永遠顯示預測按鈕
|
192 |
with gr.Row():
|
193 |
auto_predict = gr.Checkbox(
|
194 |
value=True, label="自動預測(內容變更時觸發)", elem_id="auto-predict"
|
195 |
)
|
196 |
-
predict_button = gr.Button(
|
197 |
-
"預測", elem_id="predict-button"
|
198 |
-
)
|
199 |
|
200 |
with gr.Accordion("進階設定", open=False):
|
201 |
model_selector = gr.Dropdown(
|
@@ -216,7 +133,6 @@ with gr.Blocks(css=custom_css) as demo:
|
|
216 |
label="多樣性懲罰 (diversity_penalty)"
|
217 |
)
|
218 |
|
219 |
-
# 綁定事件
|
220 |
predict_button.click(
|
221 |
fn=suggest_next,
|
222 |
inputs=[
|
@@ -251,4 +167,4 @@ with gr.Blocks(css=custom_css) as demo:
|
|
251 |
outputs=input_text,
|
252 |
)
|
253 |
|
254 |
-
demo.launch()
|
|
|
26 |
]
|
27 |
|
28 |
def merge_common_prefixes(suggestions, min_len=2):
|
|
|
|
|
|
|
|
|
|
|
29 |
prefixes = []
|
30 |
to_remove = set()
|
|
|
31 |
for i in range(len(suggestions)):
|
32 |
for j in range(i+1, len(suggestions)):
|
33 |
s1, s2 = suggestions[i], suggestions[j]
|
|
|
34 |
common = ''.join(c1 for c1, c2 in zip(s1, s2) if c1 == c2)
|
35 |
if len(common) >= min_len:
|
36 |
prefixes.append(common)
|
37 |
to_remove.update([s1, s2])
|
|
|
|
|
38 |
unique_prefixes = []
|
39 |
for p in prefixes:
|
40 |
if p not in unique_prefixes:
|
41 |
unique_prefixes.append(p)
|
|
|
|
|
42 |
remainder = [s for s in suggestions if s not in to_remove]
|
43 |
return unique_prefixes + remainder
|
44 |
|
|
|
53 |
|
54 |
@spaces.GPU
|
55 |
def suggest_next(text, model_name, k, m, num_beam_groups, diversity_penalty):
|
|
|
|
|
|
|
|
|
|
|
|
|
56 |
gen_pipe = get_pipeline(model_name)
|
57 |
+
# 構造 generate 參數字典,僅在 penalty>0 時加入 diversity 相關
|
58 |
+
gen_kwargs = {
|
59 |
+
"max_new_tokens": k,
|
60 |
+
"num_beams": m,
|
61 |
+
"num_return_sequences": m,
|
62 |
+
"do_sample": False,
|
63 |
+
"early_stopping": True,
|
64 |
+
}
|
65 |
+
if diversity_penalty and diversity_penalty > 0:
|
66 |
+
gen_kwargs["num_beam_groups"] = num_beam_groups
|
67 |
+
gen_kwargs["diversity_penalty"] = diversity_penalty
|
68 |
+
|
69 |
+
outs = gen_pipe(text, **gen_kwargs)
|
70 |
+
|
71 |
+
# 提取純下文、過濾空字串、繁體化
|
72 |
+
suggestions = [
|
73 |
+
cc.convert(out["generated_text"][len(text):].strip())
|
74 |
+
for out in outs
|
75 |
+
if out["generated_text"][len(text):].strip()
|
76 |
+
]
|
77 |
+
|
78 |
+
# 去重
|
79 |
unique_suggestions = []
|
80 |
for s in suggestions:
|
81 |
if s not in unique_suggestions:
|
|
|
86 |
|
87 |
return update(choices=final_suggestions, value=None)
|
88 |
|
89 |
+
def append_suggestion(text, choice):
|
90 |
+
return text + choice
|
91 |
|
92 |
+
with gr.Blocks(css="""
|
93 |
+
#suggestions-bar { width: 100%; margin-bottom: 8px; }
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
94 |
#suggestions-bar .candidate-list {
|
95 |
+
display: flex; gap: 8px; background: #fff;
|
96 |
+
border: 1px solid #999; border-radius: 4px;
|
97 |
+
padding: 6px; overflow-x: auto; white-space: nowrap;
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
98 |
}
|
99 |
+
#suggestions-bar .candidate-list label { cursor: pointer; }
|
100 |
+
""") as demo:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
101 |
with gr.Column():
|
102 |
suggestions = gr.Radio(
|
103 |
[], label="", interactive=True, type="value",
|
|
|
108 |
lines=1, max_lines=20, elem_id="input-box"
|
109 |
)
|
110 |
|
|
|
111 |
with gr.Row():
|
112 |
auto_predict = gr.Checkbox(
|
113 |
value=True, label="自動預測(內容變更時觸發)", elem_id="auto-predict"
|
114 |
)
|
115 |
+
predict_button = gr.Button("預測", elem_id="predict-button")
|
|
|
|
|
116 |
|
117 |
with gr.Accordion("進階設定", open=False):
|
118 |
model_selector = gr.Dropdown(
|
|
|
133 |
label="多樣性懲罰 (diversity_penalty)"
|
134 |
)
|
135 |
|
|
|
136 |
predict_button.click(
|
137 |
fn=suggest_next,
|
138 |
inputs=[
|
|
|
167 |
outputs=input_text,
|
168 |
)
|
169 |
|
170 |
+
demo.launch()
|