Initial MoveTSA dataset builder (gated Gradio app)

README.md

@@ -1,13 +1,50 @@
 ---
-title: Movetsa Builder
-emoji: 🐢
-colorFrom: yellow
-colorTo: gray
+title: MoveTSA Dataset Builder
+emoji: 🚗
+colorFrom: blue
+colorTo: indigo
 sdk: gradio
-sdk_version: 6.17.3
-python_version: '3.13'
+sdk_version: 5.9.1
 app_file: app.py
 pinned: false
+hf_oauth: true
+short_description: Generate parametrised MoveTSA windows, no raw access
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# MoveTSA dataset builder
+
+A gated Gradio Space that generates a **parametrised** MoveTSA windows parquet
+(HRV / BSI / ECG-derived respiration / simulator aggregates / DATEX / subjective
+labels) on demand — **without** giving users access to the raw recordings.
+
+## How it works
+
+- The raw recordings live in the **private** dataset `thbndi/movetsa-raw` and are
+  downloaded into the Space at startup. They never leave the Space.
+- The pipeline code is pulled from `thbndi/MoveTSA` and imported as the
+  `MoveTSA` package.
+- Authorised users sign in with Hugging Face, pick `window_size` / `overlap` /
+  `normalize` / baselines / familiarization, and download **only the generated
+  parquet**.
+
+## Access control (gated)
+
+The Space is public, but generation requires:
+1. signing in with Hugging Face (`hf_oauth`), and
+2. being listed in the `ALLOWLIST` Space variable (comma-separated usernames).
+
+Non-listed users are told to request access. Add a username by editing the
+`ALLOWLIST` variable in **Settings → Variables and secrets** — no redeploy of
+code needed.
+
+## Configuration (Settings → Variables and secrets)
+
+| Name | Kind | Purpose |
+|------|------|---------|
+| `HF_TOKEN` | **secret** | Fine-grained **read** token with access to `thbndi/movetsa-raw` and `thbndi/MoveTSA`. |
+| `ALLOWLIST` | variable | Comma-separated HF usernames allowed to generate (default: `thbndi`). |
+| `RAW_REPO` | variable (optional) | Raw dataset id (default `thbndi/movetsa-raw`). |
+| `CODE_REPO` | variable (optional) | Pipeline-code dataset id (default `thbndi/MoveTSA`). |
+
+> First load downloads ~13 GB of raw SIMU logs; expect a slow cold start. Enable
+> **persistent storage** on the Space to avoid re-downloading on each restart.

app.py

@@ -0,0 +1,114 @@
+"""MoveTSA dataset builder — Hugging Face Space (gated via HF login + allowlist).
+
+Generates a **parametrised** windows parquet (HRV / BSI / RSP / simulator
+aggregates / DATEX / subjective labels) from the PRIVATE raw recordings,
+entirely server-side. Authorised users only ever download the generated
+parquet — they never get access to the raw ECG/SIMU files.
+
+Backing repos (private, read with the ``HF_TOKEN`` Space secret):
+  - ``thbndi/movetsa-raw`` : raw recordings (stay inside the Space)
+  - ``thbndi/MoveTSA``     : pipeline code (downloaded at startup, importable
+                             as the ``MoveTSA`` package)
+
+Access control: the Space is public, but generation requires signing in with
+Hugging Face and being on the allowlist (the ``ALLOWLIST`` Space variable, a
+comma-separated list of usernames).
+"""
+
+import os
+import sys
+import tempfile
+
+import gradio as gr
+from huggingface_hub import snapshot_download
+
+TOKEN = os.environ.get("HF_TOKEN")
+CODE_REPO = os.environ.get("CODE_REPO", "thbndi/MoveTSA")
+RAW_REPO = os.environ.get("RAW_REPO", "thbndi/movetsa-raw")
+OWNER = os.environ.get("OWNER_HANDLE", "thbndi")
+
+# Comma-separated HF usernames allowed to generate. Edit via the ALLOWLIST
+# Space *variable* (Settings → Variables and secrets) — no code change needed.
+ALLOWLIST = {
+    u.strip().lower()
+    for u in os.environ.get("ALLOWLIST", OWNER).split(",")
+    if u.strip()
+}
+
+# ----------------------------------------------------------------- startup
+# 1) Pipeline code → importable as the `MoveTSA` package (cwd on sys.path).
+snapshot_download(CODE_REPO, repo_type="dataset", token=TOKEN,
+                  local_dir="MoveTSA", allow_patterns=["*.py", "*.yaml"])
+sys.path.insert(0, os.getcwd())
+
+# 2) Raw recordings — private, kept inside the Space, never served to users.
+RAW_DIR = snapshot_download(RAW_REPO, repo_type="dataset", token=TOKEN,
+                            allow_patterns=["S*/**", "subjective_scores.csv"])
+
+from MoveTSA.export_hf_dataset import build_windows  # noqa: E402  (needs sys.path)
+
+
+def generate(window_size, overlap, normalize, include_baselines,
+             include_familiarization, profile: gr.OAuthProfile | None):
+    """Run the pipeline with the chosen parameters and return a parquet file."""
+    if profile is None:
+        return None, "🔒 Connecte-toi avec ton compte Hugging Face pour générer."
+    if profile.username.lower() not in ALLOWLIST:
+        return None, (
+            f"⛔ Accès non accordé pour **@{profile.username}**.\n\n"
+            f"Demande l'accès à **@{OWNER}** (ajout à l'allowlist)."
+        )
+
+    df = build_windows(
+        RAW_DIR,
+        window_size=int(window_size),
+        window_overlap=float(overlap),
+        normalize=(None if normalize == "none" else normalize),
+        include_baselines=bool(include_baselines),
+        include_familiarization=bool(include_familiarization),
+        verbose=False,
+    )
+
+    fname = (f"movetsa_w{int(window_size)}_ov{int(float(overlap) * 100)}"
+             f"_{normalize}.parquet")
+    out = os.path.join(tempfile.mkdtemp(), fname)
+    df.to_parquet(out, index=False)
+    msg = (f"✅ **{len(df)} fenêtres × {df.shape[1]} colonnes** "
+           f"({df['subject'].nunique()} sujets) — `{fname}`")
+    return out, msg
+
+
+with gr.Blocks(title="MoveTSA dataset builder") as demo:
+    gr.Markdown(
+        "# 🚗 MoveTSA dataset builder\n"
+        "Génère un parquet **HRV / BSI / RSP / simulateur** paramétré à partir "
+        "des enregistrements bruts (privés). Tu télécharges **uniquement** le "
+        "parquet généré — jamais les données brutes.\n\n"
+        "1. Connecte-toi avec Hugging Face. 2. Règle les paramètres. 3. Génère."
+    )
+    gr.LoginButton()
+    with gr.Row():
+        with gr.Column():
+            window_size = gr.Slider(15, 180, value=60, step=5,
+                                    label="window_size (s)")
+            overlap = gr.Slider(0.0, 0.9, value=0.5, step=0.05, label="overlap")
+            normalize = gr.Dropdown(["zscore", "center", "none"],
+                                    value="zscore", label="normalize")
+            include_baselines = gr.Checkbox(
+                value=True, label="inclure les baselines (B1–B4)")
+            include_familiarization = gr.Checkbox(
+                value=True, label="inclure la familiarisation (F)")
+            btn = gr.Button("Générer le parquet", variant="primary")
+        with gr.Column():
+            status = gr.Markdown()
+            out_file = gr.File(label="parquet généré")
+
+    btn.click(
+        generate,
+        inputs=[window_size, overlap, normalize, include_baselines,
+                include_familiarization],
+        outputs=[out_file, status],
+    )
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,9 @@
+huggingface_hub>=0.25
+pandas
+numpy
+scipy
+mne
+neurokit2
+pyyaml
+tqdm
+pyarrow