initial deploy

.gitignore

@@ -0,0 +1,5 @@
+__pycache__/
+*.pyc
+.DS_Store
+.venv/
+MuScriptor-repo/

README.md

@@ -1,13 +1,14 @@
 ---
 title: MuScriptor
-emoji: 🏆
-colorFrom: indigo
-colorTo: green
+emoji: 📜
+colorFrom: blue
+colorTo: purple
 sdk: gradio
-sdk_version: 6.20.0
-python_version: '3.12'
+sdk_version: 5.28.0
+python_version: '3.13'
 app_file: app.py
 pinned: false
+license: cc-by-nc-4.0
 ---
 
 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

SOURCES.md

@@ -0,0 +1,4 @@
+# Sources — MuScriptor
+
+- Source repo: https://github.com/muscriptor/muscriptor
+- Paper: https://arxiv.org/html/2607.08168v1

app.py

@@ -0,0 +1,137 @@
+import sys
+
+sys.stdout.reconfigure(line_buffering=True)
+
+try:
+    import spaces
+except ImportError:
+    # keep @spaces.GPU usable as a no-op; ZeroGPU requires this exact name.
+    class spaces:
+        class GPU:
+            def __init__(self, func=None, duration=60):
+                self.func = func
+
+            def __call__(self, *args, **kwargs):
+                if self.func is not None:
+                    return self.func(*args, **kwargs)
+                func = args[0]
+                return func
+
+import tempfile
+
+import gradio as gr
+import torch
+from pyharp import ModelCard, build_endpoint
+
+from muscriptor import TranscriptionModel
+from muscriptor.tokenizer.mt3 import resolve_instrument_names
+
+model_card = ModelCard(
+    name="MuScriptor",
+    description=(
+        "Multi-instrument audio-to-MIDI transcription, trained on 170k songs "
+        "from classical music to heavy metal."
+    ),
+    author="Simon Rouard, Michael Krause, Axel Roebel, Carl-Johann Simon-Gabriel, Alexandre Défossez (Kyutai x Mirelo)",
+    tags=["transcription", "midi", "multi-instrument"],
+)
+
+_models: dict[str, TranscriptionModel] = {}
+
+
+def _get_model(variant: str) -> TranscriptionModel:
+    """Load and cache a TranscriptionModel for a given size, one per variant.
+
+    small/medium run on CPU; large runs on GPU. Built directly on its target
+    device inside this GPU-decorated call, per ZeroGPU rules.
+    """
+    if variant not in _models:
+        device = "cuda" if variant == "large" else "cpu"
+        _models[variant] = TranscriptionModel.load_model(variant, device=device)
+    return _models[variant]
+
+
+def _parse_instruments(text: str) -> list[str] | None:
+    """Turn the comma-separated instruments box into exact group names.
+
+    Accepts unambiguous abbreviations, same as the CLI's --instruments.
+    Blank input means "no restriction, let the model decide".
+    """
+    tokens = [t for t in text.split(",") if t.strip()]
+    if not tokens:
+        return None
+    try:
+        return resolve_instrument_names(tokens)
+    except ValueError as e:
+        raise gr.Error(str(e))
+
+
+@spaces.GPU
+@torch.inference_mode()
+def process_fn(
+    input_audio_path: str,
+    variant: str,
+    instruments_text: str,
+    use_sampling: bool,
+    temperature: float,
+) -> str:
+    model = _get_model(variant)
+    instruments = _parse_instruments(instruments_text)
+
+    midi_bytes = model.transcribe_to_midi(
+        input_audio_path,
+        use_sampling=use_sampling,
+        temperature=temperature,
+        instruments=instruments,
+    )
+
+    with tempfile.NamedTemporaryFile(suffix=".mid", delete=False) as f:
+        f.write(midi_bytes)
+        output_midi_path = f.name
+
+    return output_midi_path
+
+
+with gr.Blocks() as demo:
+    input_components = [
+        gr.Audio(type="filepath", label="Input Audio").harp_required(True),
+        gr.Dropdown(
+            choices=["small", "medium", "large"],
+            value="medium",
+            label="Model Size",
+            info="small = fastest, least accurate. medium = balanced (default, per repo). large = most accurate, slower.",
+        ),
+        gr.Textbox(
+            value="",
+            label="Instruments",
+            info="Comma-separated instrument names to restrict decoding to, e.g. acoustic_piano,violin,trumpet,drums,etc. Leave blank to let the model detect instruments on its own.",
+        ),
+        gr.Checkbox(
+            value=False,
+            label="Use Sampling",
+            info="Temperature sampling instead of greedy decoding (default: False, per repo).",
+        ),
+        gr.Slider(
+            minimum=0.1,
+            maximum=2.0,
+            step=0.1,
+            value=1.0,
+            label="Temperature",
+            info="Only used when Use Sampling is on (default: 1.0, per repo).",
+        ),
+    ]
+    output_components = [
+        gr.File(type="filepath", file_types=[".mid", ".midi"], label="Output MIDI").set_info(
+            "Transcribed MIDI notes."
+        ),
+    ]
+
+    build_endpoint(
+        model_card=model_card,
+        input_components=input_components,
+        output_components=output_components,
+        process_fn=process_fn,
+    )
+
+if __name__ == "__main__":
+    demo.queue().launch(pwa=True)

model.json

@@ -0,0 +1,12 @@
+{
+  "name": "MuScriptor",
+  "package_dir": null,
+  "package_source": "pip install muscriptor (published package, not vendored)",
+  "entry_point": "muscriptor.TranscriptionModel",
+  "checkpoints": {
+    "small": {"repo": "MuScriptor/muscriptor-small", "filename": "model.safetensors", "size_mb": 412},
+    "medium": {"repo": "MuScriptor/muscriptor-medium", "filename": "model.safetensors", "size_mb": 1228},
+    "large": {"repo": "MuScriptor/muscriptor-large", "filename": "model.safetensors", "size_mb": 5466}
+  },
+  "checkpoint_hosting": "gated on MuScriptor's own HF org, downloaded at runtime by muscriptor.TranscriptionModel.load_model() via hf_hub_download; requires HF_TOKEN Space secret from an account that has accepted the CC BY-NC 4.0 license on all three muscriptor-* repos"
+}

requirements.txt

@@ -0,0 +1,2 @@
+git+https://github.com/TEAMuP-dev/pyharp.git@develop
+muscriptor