Upload app.py with huggingface_hub

app.py

@@ -0,0 +1,97 @@
+"""
+HuggingFace Space: Turkish Diacritic Restoration
+Loads the CRF model from the emircanerol/turkish-diacritic-crf model repo.
+"""
+import gradio as gr
+import torch
+from huggingface_hub import hf_hub_download
+
+# ── Load model ────────────────────────────────────────────────────────────────
+
+def _load_model():
+    # Download inference code and weights from the model repo
+    vocab_path   = hf_hub_download("emircanerol/turkish-diacritic-crf", "ldgc/vocab.py")
+    crf_gpu_path = hf_hub_download("emircanerol/turkish-diacritic-crf", "crf_gpu.py")
+    weights_path = hf_hub_download("emircanerol/turkish-diacritic-crf", "crf_gpu.safetensors")
+
+    # Make ldgc.vocab importable from the cached path
+    import importlib.util, sys, os
+    ldgc_dir = os.path.dirname(vocab_path)
+    pkg_dir  = os.path.dirname(ldgc_dir)
+
+    # Install ldgc as a package if not already
+    if pkg_dir not in sys.path:
+        sys.path.insert(0, pkg_dir)
+    init = os.path.join(ldgc_dir, "__init__.py")
+    if not os.path.exists(init):
+        open(init, "w").close()
+
+    # Load crf_gpu module from cached path
+    spec = importlib.util.spec_from_file_location("crf_gpu", crf_gpu_path)
+    mod  = importlib.util.module_from_spec(spec)
+    sys.modules["crf_gpu"] = mod
+    spec.loader.exec_module(mod)
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    model  = mod.CRFGPUModel.from_pretrained(weights_path, device=device)
+    return model, mod.predict_stream
+
+
+MODEL, predict_stream = _load_model()
+
+# ── Gradio interface ──────────────────────────────────────────────────────────
+
+_EXAMPLES = [
+    "turkce dogal dil isleme cok onemlidir",
+    "bugun hava cok guzel, disari cikmak istiyorum",
+    "universite ogrencileri kütüphane de calisıyor",
+    "ruzgar bugun cok siddetli esiyor",
+    "turkiye'nin baskenti ankara'dir",
+]
+
+
+def restore(text: str) -> str:
+    if not text.strip():
+        return ""
+    lines  = [l for l in text.splitlines() if l.strip()]
+    preds  = MODEL.predict(lines, batch_size=128)
+    return "\n".join(preds)
+
+
+with gr.Blocks(title="Turkish Diacritic Restoration") as demo:
+    gr.Markdown(
+        """
+        # Turkish Diacritic Restoration
+        Restores missing diacritics in Turkish text
+        (ç, ğ, ı, ö, ş, ü and circumflex variants).
+
+        **Model**: Bidirectional CRF with ±2 character context and bigram features,
+        trained on 200k Wikipedia sentences.
+        · [Model repo](https://huggingface.co/emircanerol/turkish-diacritic-crf)
+        · [Code](https://github.com/emircanerol/tr-grammar)
+        """
+    )
+
+    with gr.Row():
+        with gr.Column():
+            inp = gr.Textbox(
+                label="Noisy Turkish text (one sentence per line)",
+                placeholder="turkce cok guzel…",
+                lines=6,
+            )
+            btn = gr.Button("Restore diacritics", variant="primary")
+        with gr.Column():
+            out = gr.Textbox(label="Restored text", lines=6, interactive=False)
+
+    gr.Examples(
+        examples=[[e] for e in _EXAMPLES],
+        inputs=inp,
+        outputs=out,
+        fn=restore,
+        cache_examples=True,
+    )
+
+    btn.click(restore, inputs=inp, outputs=out)
+    inp.submit(restore, inputs=inp, outputs=out)
+
+demo.launch()