Create app.py

app.py

@@ -0,0 +1,255 @@
+# ═══════════════════════════════════════════════════════════════════════════════
+# COMPATIBILITY PATCHES — must run before any other import
+# ═══════════════════════════════════════════════════════════════════════════════
+
+import sys, types
+
+# 1) Python 3.13 removed audioop; pydub needs it via gradio 4.44
+for _m in ("audioop", "pyaudioop"):
+    if _m not in sys.modules:
+        sys.modules[_m] = types.ModuleType(_m)
+
+# 2) gradio_client schema patches (bool/non-dict schema guard)
+import gradio_client.utils as _gcu
+
+_orig_get_type = _gcu.get_type
+_orig_inner    = _gcu._json_schema_to_python_type
+_orig_j2p      = _gcu.json_schema_to_python_type
+
+def _safe_get_type(schema):
+    if not isinstance(schema, dict): return {}
+    return _orig_get_type(schema)
+
+def _safe_inner(schema, defs=None):
+    if not isinstance(schema, dict): return "Any"
+    return _orig_inner(schema, defs)
+
+def _safe_j2p(schema):
+    if not isinstance(schema, dict): return "Any"
+    return _orig_j2p(schema)
+
+_gcu.get_type                    = _safe_get_type
+_gcu._json_schema_to_python_type = _safe_inner
+_gcu.json_schema_to_python_type  = _safe_j2p
+
+# 3) FIX: gradio 4.44.0 calls templates.TemplateResponse(dict, dict) — passes
+#    the context dict as the first arg (template name) instead of a string.
+#    Starlette's TemplateResponse signature is (name: str, context: dict, ...).
+#    We patch TemplateResponse to detect the swapped args and correct them.
+import starlette.templating as _st
+
+_OrigJinja2Templates = _st.Jinja2Templates
+
+class _PatchedJinja2Templates(_OrigJinja2Templates):
+    def TemplateResponse(self, *args, **kwargs):
+        # Gradio 4.44 bug: passes (dict_context, dict_context) — first arg is
+        # a dict instead of the template name string.
+        if args and isinstance(args[0], dict):
+            ctx = args[0]
+            # The real template name lives in the context under "request" scope;
+            # gradio always uses "index.html" as its main template.
+            name = ctx.pop("__template_name__", None) or "index.html"
+            # Reconstruct as (name, context, *rest)
+            args = (name, ctx) + args[1:]
+        return super().TemplateResponse(*args, **kwargs)
+
+# Patch into starlette.templating so any already-imported reference is updated
+_st.Jinja2Templates = _PatchedJinja2Templates
+
+# Also patch it into gradio.routes which imported Jinja2Templates at load time
+try:
+    import gradio.routes as _gr
+    if hasattr(_gr, 'templates') and _gr.templates is not None:
+        _gr.templates.__class__ = _PatchedJinja2Templates
+except Exception:
+    pass
+
+# ═══════════════════════════════════════════════════════════════════════════════
+# MAIN APP
+# ═══════════════════════════════════════════════════════════════════════════════
+
+import gradio as gr
+import re, os, tempfile
+from huggingface_hub import InferenceClient
+
+HF_TOKEN = os.environ.get("HF_TOKEN", "")
+MODEL_ID  = "google/gemma-2-9b-it"
+client    = InferenceClient(model=MODEL_ID, token=HF_TOKEN)
+
+# ── SRT helpers ────────────────────────────────────────────────────────────────
+
+def parse_srt(content: str) -> list:
+    blocks = []
+    for block in re.split(r'\n\s*\n', content.strip()):
+        lines = block.strip().splitlines()
+        if len(lines) < 2 or not lines[0].strip().isdigit() or '-->' not in lines[1]:
+            continue
+        blocks.append({
+            'number':   int(lines[0].strip()),
+            'timecode': lines[1].strip(),
+            'text':     '\n'.join(lines[2:]).strip(),
+        })
+    return blocks
+
+def blocks_to_srt(blocks: list) -> str:
+    return '\n\n'.join(
+        f"{i}\n{b['timecode']}\n{b['text']}"
+        for i, b in enumerate(blocks, 1)
+    ) + '\n'
+
+# ── Detection ──────────────────────────────────────────────────────────────────
+
+REMOVE_RE = re.compile(
+    r'https?://\S+|www\.\S+\.\S+'
+    r'|subscene|opensubtitles|addic7ed|podnapisi|subdownloader'
+    r'|subtitles?\s+by\s+\w+|translated\s+by\s+\w+'
+    r'|encoded\s+by|synced\s+by|ripped\s+by|corrected\s+by'
+    r'|\bcopyright\b|©|all\s+rights\s+reserved'
+    r'|downloaded\s+from|subtitle\s+group',
+    re.IGNORECASE
+)
+
+def is_metadata(text: str) -> bool:
+    return bool(REMOVE_RE.search(text))
+
+def gemma_classify(text: str) -> str:
+    if not text.strip():
+        return 'REMOVE'
+    prompt = (
+        "<start_of_turn>user\n"
+        "Is this subtitle text metadata (credits/copyright/website/promo) or real dialogue?\n"
+        "Reply ONLY 'REMOVE' or 'KEEP'.\n\nText:\n" + text +
+        "\n<end_of_turn>\n<start_of_turn>model\n"
+    )
+    try:
+        r = client.text_generation(prompt, max_new_tokens=10, temperature=0.1, do_sample=False)
+        return 'REMOVE' if 'REMOVE' in r.strip().upper() else 'KEEP'
+    except Exception as e:
+        print(f"[Gemma] {e}")
+        return 'KEEP'
+
+# ── Core ───────────────────────────────────────────────────────────────────────
+
+def process(content: str, use_gemma: bool):
+    blocks = parse_srt(content)
+    if not blocks:
+        return "", "⚠️ කිසිම subtitle block එකක් හමු නොවීය."
+
+    kept, log = [], []
+    for b in blocks:
+        t = b['text']
+        if is_metadata(t):
+            log.append(f"#{b['number']} [REGEX]: {t[:80]}")
+            continue
+        if use_gemma and t.strip() and gemma_classify(t) == 'REMOVE':
+            log.append(f"#{b['number']} [GEMMA]: {t[:80]}")
+            continue
+        kept.append(b)
+
+    report = [
+        f"✅ සම්පූර්ණ blocks: {len(blocks)}",
+        f"🗑️ ඉවත් කළ blocks: {len(blocks)-len(kept)}",
+        f"✔️ ඉතිරි blocks: {len(kept)}", "",
+    ]
+    if log:
+        report += ["ඉවත් කළ lines:"] + [f"  • {l}" for l in log]
+
+    return blocks_to_srt(kept), '\n'.join(report)
+
+def save_tmp(text: str) -> str:
+    f = tempfile.NamedTemporaryFile(mode='w', suffix='_cleaned.srt',
+                                    encoding='utf-8', delete=False)
+    f.write(text); f.close()
+    return f.name
+
+# ── Handlers ───────────────────────────────────────────────────────────────────
+
+def handle_file(path, use_gemma):
+    if not path:
+        return "", "ගොනුවක් තෝරන්න.", None
+    try:
+        with open(path, encoding='utf-8-sig', errors='replace') as f:
+            content = f.read()
+    except Exception as e:
+        return "", f"❌ {e}", None
+    cleaned, report = process(content, use_gemma)
+    return cleaned, report, save_tmp(cleaned) if cleaned else None
+
+def handle_text(text, use_gemma):
+    if not text.strip():
+        return "", "⚠️ SRT content ඇතුළු කරන්න.", None
+    cleaned, report = process(text, use_gemma)
+    return cleaned, report, save_tmp(cleaned) if cleaned else None
+
+# ── UI ─────────────────────────────────────────────────────────────────────────
+
+css = ".mono{font-family:monospace;font-size:13px} footer{display:none!important}"
+
+with gr.Blocks(title="🎬 Sinhala SRT Cleaner", theme=gr.themes.Soft(), css=css) as demo:
+
+    gr.Markdown("# 🎬 සිංහල SRT Subtitle Cleaner\n**Gemma 2 9B** භාවිතා කර author credits, website URLs, copyright notices ඉවත් කරයි.")
+
+    with gr.Tabs():
+
+        with gr.TabItem("📁 ගොනු Upload"):
+            with gr.Row():
+                with gr.Column(scale=1):
+                    file_in   = gr.File(label="SRT ගොනුව", file_types=[".srt"], type="filepath")
+                    gemma_cb1 = gr.Checkbox(label="🤖 Gemma AI", value=True,
+                                            info="Uncheck → Regex only (ඉක්මන්)")
+                    btn1      = gr.Button("🚀 Process", variant="primary", size="lg")
+                with gr.Column(scale=2):
+                    report1 = gr.Textbox(label="📊 Report", lines=8, interactive=False)
+            preview1 = gr.Textbox(label="✅ Cleaned Preview", lines=15,
+                                  interactive=False, elem_classes="mono")
+            dl1 = gr.File(label="⬇️ Download")
+            btn1.click(fn=handle_file, inputs=[file_in, gemma_cb1],
+                       outputs=[preview1, report1, dl1])
+
+        with gr.TabItem("📝 Text Paste"):
+            with gr.Row():
+                with gr.Column():
+                    text_in   = gr.Textbox(label="SRT Content", lines=15, elem_classes="mono",
+                                           placeholder="1\n00:00:01,000 --> 00:00:04,000\nExample...")
+                    gemma_cb2 = gr.Checkbox(label="🤖 Gemma AI", value=True)
+                    btn2      = gr.Button("🚀 Process", variant="primary")
+                with gr.Column():
+                    report2  = gr.Textbox(label="📊 Report", lines=6, interactive=False)
+                    preview2 = gr.Textbox(label="✅ Cleaned Output", lines=15,
+                                         interactive=False, elem_classes="mono")
+                    dl2      = gr.File(label="⬇️ Download")
+            btn2.click(fn=handle_text, inputs=[text_in, gemma_cb2],
+                       outputs=[preview2, report2, dl2])
+
+        with gr.TabItem("🔌 API"):
+            gr.Markdown("""
+## API Endpoint
+### Python — File
+```python
+from gradio_client import Client
+client = Client("your-username/srt-cleaner")
+cleaned, report, file = client.predict(
+    "path/to/file.srt", True, api_name="/handle_file"
+)
+```
+### Python — Text
+```python
+cleaned, report, file = client.predict(
+    "1\\n00:00:01,000 --> 00:00:04,000\\nwww.subscene.com\\n\\n2\\n...",
+    True, api_name="/handle_text"
+)
+```
+### cURL
+```bash
+curl -X POST "https://your-space.hf.space/run/predict" \\
+  -H "Content-Type: application/json" \\
+  -d '{"fn_index":1,"data":["SRT_CONTENT",true]}'
+```
+            """)
+
+    gr.Markdown("---\n**Setup:** Space Settings → Secrets → `HF_TOKEN` set කරන්න.")
+
+demo.queue(max_size=10)
+
+if __name__ == "__main__":
+    demo.launch(server_name="0.0.0.0", server_port=7860)