Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
|
@@ -8,14 +8,14 @@ from pypdf import PdfReader
|
|
| 8 |
# 1. 加载模型(CPU Basic 友好)
|
| 9 |
# =========================================================
|
| 10 |
|
| 11 |
-
# 图像 → 文本:BLIP,比 vit-gpt2
|
| 12 |
vision_pipe = pipeline(
|
| 13 |
"image-to-text",
|
| 14 |
model="salesforce/blip-image-captioning-base",
|
| 15 |
device=-1 # CPU
|
| 16 |
)
|
| 17 |
|
| 18 |
-
# 文本生成:
|
| 19 |
text_pipe = pipeline(
|
| 20 |
"text2text-generation",
|
| 21 |
model="google/flan-t5-small",
|
|
@@ -25,20 +25,14 @@ text_pipe = pipeline(
|
|
| 25 |
|
| 26 |
|
| 27 |
# =========================================================
|
| 28 |
-
# 2.
|
| 29 |
# =========================================================
|
| 30 |
|
| 31 |
def dedup_sentences(text: str) -> str:
|
| 32 |
-
"""
|
| 33 |
-
简单去重复:
|
| 34 |
-
- 按句号/问号/感叹号拆分
|
| 35 |
-
- 去掉重复句子(忽略大小写和多余空格)
|
| 36 |
-
- 防止 caption / summary 疯狂复读同一句话
|
| 37 |
-
"""
|
| 38 |
if not text:
|
| 39 |
return text
|
| 40 |
|
| 41 |
-
# 先粗略按中英文句号拆分
|
| 42 |
parts = re.split(r'(?<=[。!?!?\.])\s+', text.strip())
|
| 43 |
seen = set()
|
| 44 |
result = []
|
|
@@ -54,22 +48,22 @@ def dedup_sentences(text: str) -> str:
|
|
| 54 |
result.append(s_clean)
|
| 55 |
|
| 56 |
out = ' '.join(result)
|
| 57 |
-
# 太长就截断一点,避免刷屏
|
| 58 |
if len(out) > 1200:
|
| 59 |
out = out[:1200]
|
| 60 |
return out
|
| 61 |
|
| 62 |
|
| 63 |
-
def extract_pdf_context(
|
| 64 |
"""
|
| 65 |
从上传的 PDF 中抽取和关键词最相关的几段文字,
|
| 66 |
再作为 figure 的上下文摘要使用。
|
|
|
|
| 67 |
"""
|
| 68 |
-
if
|
| 69 |
return ""
|
| 70 |
|
| 71 |
try:
|
| 72 |
-
reader = PdfReader(
|
| 73 |
pages_text = []
|
| 74 |
for page in reader.pages:
|
| 75 |
txt = page.extract_text() or ""
|
|
@@ -81,14 +75,9 @@ def extract_pdf_context(pdf_file, keywords: str, max_chars: int = 1200) -> str:
|
|
| 81 |
if not full_text.strip():
|
| 82 |
return ""
|
| 83 |
|
| 84 |
-
# 简单按空行拆分段落
|
| 85 |
paragraphs = [p.strip() for p in re.split(r'\n\s*\n', full_text) if p.strip()]
|
| 86 |
|
| 87 |
-
# 关键词打分(非常朴素,但够用)
|
| 88 |
kw_list = [k.strip().lower() for k in keywords.split(",") if k.strip()]
|
| 89 |
-
if not kw_list:
|
| 90 |
-
kw_list = []
|
| 91 |
-
|
| 92 |
scored = []
|
| 93 |
for p in paragraphs:
|
| 94 |
pl = p.lower()
|
|
@@ -96,13 +85,11 @@ def extract_pdf_context(pdf_file, keywords: str, max_chars: int = 1200) -> str:
|
|
| 96 |
for kw in kw_list:
|
| 97 |
if kw in pl:
|
| 98 |
score += 1
|
| 99 |
-
# 略微偏向包含 "fig" 的段落
|
| 100 |
if "fig" in pl or "figure" in pl:
|
| 101 |
score += 1
|
| 102 |
if score > 0:
|
| 103 |
scored.append((score, p))
|
| 104 |
|
| 105 |
-
# 如果没匹配到关键词,就拿前几段凑合
|
| 106 |
if not scored:
|
| 107 |
top_text = "\n\n".join(paragraphs[:3])
|
| 108 |
else:
|
|
@@ -110,11 +97,9 @@ def extract_pdf_context(pdf_file, keywords: str, max_chars: int = 1200) -> str:
|
|
| 110 |
top_paras = [p for _, p in scored[:3]]
|
| 111 |
top_text = "\n\n".join(top_paras)
|
| 112 |
|
| 113 |
-
# 控制长度
|
| 114 |
if len(top_text) > max_chars:
|
| 115 |
top_text = top_text[:max_chars]
|
| 116 |
|
| 117 |
-
# 用 T5 生成一个 2–3 句的简短 summary
|
| 118 |
prompt = (
|
| 119 |
"You are summarizing background for one figure in a scientific paper.\n\n"
|
| 120 |
f"Keywords: {', '.join(kw_list) if kw_list else 'N/A'}\n\n"
|
|
@@ -128,10 +113,7 @@ def extract_pdf_context(pdf_file, keywords: str, max_chars: int = 1200) -> str:
|
|
| 128 |
|
| 129 |
|
| 130 |
def enhance_image_simple(image: Image.Image) -> Image.Image:
|
| 131 |
-
"""
|
| 132 |
-
简单地提升亮度和对比度,用来生成“增强版图表”。
|
| 133 |
-
不用 Diffusion,轻量、可在 CPU Basic 上跑。
|
| 134 |
-
"""
|
| 135 |
if image is None:
|
| 136 |
return None
|
| 137 |
img = image.convert("RGB")
|
|
@@ -141,25 +123,23 @@ def enhance_image_simple(image: Image.Image) -> Image.Image:
|
|
| 141 |
|
| 142 |
|
| 143 |
# =========================================================
|
| 144 |
-
# 3. 主函数
|
| 145 |
-
# Step 2:1 次 small T5
|
| 146 |
-
# Step 3 + Caption + Summary:合并成 1 次 small T5
|
| 147 |
# =========================================================
|
| 148 |
|
| 149 |
-
def analyze_and_explain(image, keywords, language,
|
| 150 |
if image is None:
|
| 151 |
return None, "", "", None, "", ""
|
| 152 |
|
| 153 |
-
# ----
|
| 154 |
try:
|
| 155 |
raw_caption = vision_pipe(image)[0]["generated_text"]
|
| 156 |
except Exception:
|
| 157 |
raw_caption = ""
|
| 158 |
|
| 159 |
-
# ----
|
| 160 |
-
context_summary = extract_pdf_context(
|
| 161 |
|
| 162 |
-
# ---- Step 2
|
| 163 |
step2_prompt = (
|
| 164 |
"You are helping to describe one bar chart in a scientific paper.\n\n"
|
| 165 |
f"Rough visual caption from an image model: {raw_caption}\n"
|
|
@@ -181,12 +161,10 @@ def analyze_and_explain(image, keywords, language, pdf_file):
|
|
| 181 |
)
|
| 182 |
step2_en = text_pipe(step2_prompt)[0]["generated_text"].strip()
|
| 183 |
step2_en = dedup_sentences(step2_en)
|
| 184 |
-
|
| 185 |
-
# 保底:强制以 A bar chart showing 开头
|
| 186 |
if not step2_en.lower().startswith("a bar chart showing"):
|
| 187 |
step2_en = "A bar chart showing " + step2_en.lstrip()
|
| 188 |
|
| 189 |
-
# ---- Step 3 + Step 5:
|
| 190 |
context_part = (
|
| 191 |
f"Short paper context: {context_summary}\n\n"
|
| 192 |
if context_summary else ""
|
|
@@ -217,7 +195,6 @@ def analyze_and_explain(image, keywords, language, pdf_file):
|
|
| 217 |
|
| 218 |
multi_out = text_pipe(multi_prompt)[0]["generated_text"]
|
| 219 |
|
| 220 |
-
# 解析三段
|
| 221 |
expl_part = ""
|
| 222 |
cap_part = ""
|
| 223 |
sum_part = ""
|
|
@@ -241,10 +218,10 @@ def analyze_and_explain(image, keywords, language, pdf_file):
|
|
| 241 |
caption_en = dedup_sentences(cap_part)
|
| 242 |
summary_en = dedup_sentences(sum_part)
|
| 243 |
|
| 244 |
-
# ---- Step 4
|
| 245 |
enhanced_img = enhance_image_simple(image)
|
| 246 |
|
| 247 |
-
# ----
|
| 248 |
if language == "中文":
|
| 249 |
def translate_to_zh(txt):
|
| 250 |
if not txt:
|
|
@@ -279,7 +256,7 @@ with gr.Blocks() as demo:
|
|
| 279 |
gr.Markdown("## ChartSmith – AI 论文图表生成助手 (v2)")
|
| 280 |
|
| 281 |
with gr.Row():
|
| 282 |
-
# 左侧:输入
|
| 283 |
with gr.Column():
|
| 284 |
img_in = gr.Image(
|
| 285 |
type="pil",
|
|
@@ -296,11 +273,11 @@ with gr.Blocks() as demo:
|
|
| 296 |
)
|
| 297 |
pdf_in = gr.File(
|
| 298 |
label="上传论文 PDF(可选,用于结合上下文解释图表)",
|
| 299 |
-
type="
|
| 300 |
)
|
| 301 |
run_btn = gr.Button("分析并美化图表", variant="primary")
|
| 302 |
|
| 303 |
-
# 右侧:输出
|
| 304 |
with gr.Column():
|
| 305 |
orig_img = gr.Image(label="原始图表预览", interactive=False)
|
| 306 |
|
|
|
|
| 8 |
# 1. 加载模型(CPU Basic 友好)
|
| 9 |
# =========================================================
|
| 10 |
|
| 11 |
+
# 图像 → 文本:BLIP,比 vit-gpt2 稳定
|
| 12 |
vision_pipe = pipeline(
|
| 13 |
"image-to-text",
|
| 14 |
model="salesforce/blip-image-captioning-base",
|
| 15 |
device=-1 # CPU
|
| 16 |
)
|
| 17 |
|
| 18 |
+
# 文本生成:small T5,用来做描述/解释/翻译
|
| 19 |
text_pipe = pipeline(
|
| 20 |
"text2text-generation",
|
| 21 |
model="google/flan-t5-small",
|
|
|
|
| 25 |
|
| 26 |
|
| 27 |
# =========================================================
|
| 28 |
+
# 2. 工具函数
|
| 29 |
# =========================================================
|
| 30 |
|
| 31 |
def dedup_sentences(text: str) -> str:
|
| 32 |
+
"""按句子粗糙去重复,避免疯狂复读。"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
if not text:
|
| 34 |
return text
|
| 35 |
|
|
|
|
| 36 |
parts = re.split(r'(?<=[。!?!?\.])\s+', text.strip())
|
| 37 |
seen = set()
|
| 38 |
result = []
|
|
|
|
| 48 |
result.append(s_clean)
|
| 49 |
|
| 50 |
out = ' '.join(result)
|
|
|
|
| 51 |
if len(out) > 1200:
|
| 52 |
out = out[:1200]
|
| 53 |
return out
|
| 54 |
|
| 55 |
|
| 56 |
+
def extract_pdf_context(pdf_path: str, keywords: str, max_chars: int = 1200) -> str:
|
| 57 |
"""
|
| 58 |
从上传的 PDF 中抽取和关键词最相关的几段文字,
|
| 59 |
再作为 figure 的上下文摘要使用。
|
| 60 |
+
pdf_path 是文件路径字符串(来自 gr.File, type="filepath")。
|
| 61 |
"""
|
| 62 |
+
if not pdf_path:
|
| 63 |
return ""
|
| 64 |
|
| 65 |
try:
|
| 66 |
+
reader = PdfReader(pdf_path)
|
| 67 |
pages_text = []
|
| 68 |
for page in reader.pages:
|
| 69 |
txt = page.extract_text() or ""
|
|
|
|
| 75 |
if not full_text.strip():
|
| 76 |
return ""
|
| 77 |
|
|
|
|
| 78 |
paragraphs = [p.strip() for p in re.split(r'\n\s*\n', full_text) if p.strip()]
|
| 79 |
|
|
|
|
| 80 |
kw_list = [k.strip().lower() for k in keywords.split(",") if k.strip()]
|
|
|
|
|
|
|
|
|
|
| 81 |
scored = []
|
| 82 |
for p in paragraphs:
|
| 83 |
pl = p.lower()
|
|
|
|
| 85 |
for kw in kw_list:
|
| 86 |
if kw in pl:
|
| 87 |
score += 1
|
|
|
|
| 88 |
if "fig" in pl or "figure" in pl:
|
| 89 |
score += 1
|
| 90 |
if score > 0:
|
| 91 |
scored.append((score, p))
|
| 92 |
|
|
|
|
| 93 |
if not scored:
|
| 94 |
top_text = "\n\n".join(paragraphs[:3])
|
| 95 |
else:
|
|
|
|
| 97 |
top_paras = [p for _, p in scored[:3]]
|
| 98 |
top_text = "\n\n".join(top_paras)
|
| 99 |
|
|
|
|
| 100 |
if len(top_text) > max_chars:
|
| 101 |
top_text = top_text[:max_chars]
|
| 102 |
|
|
|
|
| 103 |
prompt = (
|
| 104 |
"You are summarizing background for one figure in a scientific paper.\n\n"
|
| 105 |
f"Keywords: {', '.join(kw_list) if kw_list else 'N/A'}\n\n"
|
|
|
|
| 113 |
|
| 114 |
|
| 115 |
def enhance_image_simple(image: Image.Image) -> Image.Image:
|
| 116 |
+
"""简单增强亮度和对比度,生成 Step 4 图像。"""
|
|
|
|
|
|
|
|
|
|
| 117 |
if image is None:
|
| 118 |
return None
|
| 119 |
img = image.convert("RGB")
|
|
|
|
| 123 |
|
| 124 |
|
| 125 |
# =========================================================
|
| 126 |
+
# 3. 主逻辑函数
|
|
|
|
|
|
|
| 127 |
# =========================================================
|
| 128 |
|
| 129 |
+
def analyze_and_explain(image, keywords, language, pdf_path):
|
| 130 |
if image is None:
|
| 131 |
return None, "", "", None, "", ""
|
| 132 |
|
| 133 |
+
# ---- Vision 初步 caption ----
|
| 134 |
try:
|
| 135 |
raw_caption = vision_pipe(image)[0]["generated_text"]
|
| 136 |
except Exception:
|
| 137 |
raw_caption = ""
|
| 138 |
|
| 139 |
+
# ---- PDF 上下文 ----
|
| 140 |
+
context_summary = extract_pdf_context(pdf_path, keywords or "")
|
| 141 |
|
| 142 |
+
# ---- Step 2:自动图表描述(1 次 small T5)----
|
| 143 |
step2_prompt = (
|
| 144 |
"You are helping to describe one bar chart in a scientific paper.\n\n"
|
| 145 |
f"Rough visual caption from an image model: {raw_caption}\n"
|
|
|
|
| 161 |
)
|
| 162 |
step2_en = text_pipe(step2_prompt)[0]["generated_text"].strip()
|
| 163 |
step2_en = dedup_sentences(step2_en)
|
|
|
|
|
|
|
| 164 |
if not step2_en.lower().startswith("a bar chart showing"):
|
| 165 |
step2_en = "A bar chart showing " + step2_en.lstrip()
|
| 166 |
|
| 167 |
+
# ---- Step 3 + Step 5 合并:1 次 small T5 ----
|
| 168 |
context_part = (
|
| 169 |
f"Short paper context: {context_summary}\n\n"
|
| 170 |
if context_summary else ""
|
|
|
|
| 195 |
|
| 196 |
multi_out = text_pipe(multi_prompt)[0]["generated_text"]
|
| 197 |
|
|
|
|
| 198 |
expl_part = ""
|
| 199 |
cap_part = ""
|
| 200 |
sum_part = ""
|
|
|
|
| 218 |
caption_en = dedup_sentences(cap_part)
|
| 219 |
summary_en = dedup_sentences(sum_part)
|
| 220 |
|
| 221 |
+
# ---- Step 4:图像增强 ----
|
| 222 |
enhanced_img = enhance_image_simple(image)
|
| 223 |
|
| 224 |
+
# ---- 中英文切换 ----
|
| 225 |
if language == "中文":
|
| 226 |
def translate_to_zh(txt):
|
| 227 |
if not txt:
|
|
|
|
| 256 |
gr.Markdown("## ChartSmith – AI 论文图表生成助手 (v2)")
|
| 257 |
|
| 258 |
with gr.Row():
|
| 259 |
+
# 左侧:输入
|
| 260 |
with gr.Column():
|
| 261 |
img_in = gr.Image(
|
| 262 |
type="pil",
|
|
|
|
| 273 |
)
|
| 274 |
pdf_in = gr.File(
|
| 275 |
label="上传论文 PDF(可选,用于结合上下文解释图表)",
|
| 276 |
+
type="filepath" # ✅ 这里改成 filepath
|
| 277 |
)
|
| 278 |
run_btn = gr.Button("分析并美化图表", variant="primary")
|
| 279 |
|
| 280 |
+
# 右侧:输出
|
| 281 |
with gr.Column():
|
| 282 |
orig_img = gr.Image(label="原始图表预览", interactive=False)
|
| 283 |
|