Blame for src/phyphox_app_block.py

"""Streamlit UI block for Tab 2: Phyphox live sensor upload.

Call from streamlit_app.py:

    from phyphox_app_block import render_phyphox_tab
    with tab2:
        render_phyphox_tab(ffn_model, ffn_status, cnn_model, cnn_status)
"""

import json
import os

import numpy as np
import pandas as pd
import streamlit as st

from phyphox_pipeline import process_phyphox_files, FS, WINDOW, STEP

LABEL_MAP = {
    0: "WALKING",
    1: "WALKING_UPSTAIRS",
    2: "WALKING_DOWNSTAIRS",
    3: "SITTING",
    4: "STANDING",
    5: "LAYING",
}

EXPLANATIONS = {
    "LAYING": "Minimal movement detected across all axes: consistent with a stationary horizontal posture.",
    "SITTING": "Low dynamic acceleration with stable gravity: stationary upright posture.",
    "STANDING": "Similar to sitting with slight postural micro-movements.",
    "WALKING": "Rhythmic periodic acceleration on the vertical axis: level walking at normal cadence.",
    "WALKING_DOWNSTAIRS": "Downward gravitational shift with higher impact peaks: descending stairs.",
    "WALKING_UPSTAIRS": "Elevated vertical acceleration effort: climbing stairs.",
}

@st.cache_resource
def _load_norm_params(norm_path: str):
    """Load per-feature min/max from norm_params.json (keys are string indices 0-560)."""
    with open(norm_path) as f:
        d = json.load(f)
    min_vals = np.array([d[str(i)]["min"] for i in range(561)], dtype=np.float32)
    max_vals = np.array([d[str(i)]["max"] for i in range(561)], dtype=np.float32)
    return min_vals, max_vals


def _normalize(features: np.ndarray, min_vals: np.ndarray, max_vals: np.ndarray) -> np.ndarray:
    """Best-effort feature-level min-max scaling to [-1, 1].

    Uses per-feature min/max observed in the UCI HAR training set. This is
    an approximation: the UCI pipeline normalises raw signals before feature
    extraction, so physical-unit features may fall outside the training range.
    Values are clipped before scaling to keep outputs bounded.
    """
    rng = max_vals - min_vals
    rng = np.where(rng < 1e-8, 1.0, rng)          # avoid div-by-zero
    clipped = np.clip(features, min_vals, max_vals)
    return 2.0 * (clipped - min_vals) / rng - 1.0


def render_phyphox_tab(
    ffn_model, ffn_status: str,
    cnn_model, cnn_status: str,
    norm_params_path: str,
) -> None:
    st.subheader("Upload Phyphox sensor recording")
    with st.expander("How to record and export your data - step by step", expanded=True):
        st.markdown("""
**Step 1 - Install Phyphox**
Download the free [Phyphox](https://phyphox.org/) app from the Play Store (Android) or App Store (iOS).

**Step 2 - Create a new experiment**
- Open the app and tap the **+** icon in the top-right corner
- Select **Add simple experiment**
- In the active sensors list, add **Accelerometer** and **Gyroscope**
- Set the **sensor rate to 50** (Hz) for both
- Give the experiment a title (e.g. "Walking") then tap **Proceed**

**Step 3 - Configure a timed run**
- Tap the **three-dot menu** in the top-right corner and select **Timed run**
- Set a **recording duration** (15-20 seconds recommended)
- Optionally set a **start delay** (e.g. 5 seconds) so you have time to get into position before recording begins

**Step 4 - Position the device**
Place your phone in your **trouser pocket** or attach it at your **waist**. This mirrors how the original UCI dataset was collected and gives the most accurate results.

**Step 5 - Record one activity**
- Press the **play button** to start
- Perform a **single activity continuously** - do not stop and restart mid-recording
- Stay in steady motion for the full duration before stopping

**Step 6 - Export the data**
- After the recording ends, tap the **three-dot menu** and select **Export data**
- Choose **CSV (comma separated values)** and confirm
- You will receive a **.zip file** - unzip it to find separate CSV files for the **Accelerometer** and **Gyroscope**

**Step 7 - Upload below**
Upload each CSV file into its corresponding field below, then the pipeline will extract features and classify your activity.

> ⚠️ **Important:** Each export is a fresh experiment. Do not reuse an old experiment - it will contain data from previous recordings stitched together, which will confuse the classifier.
        """)

    if "phyphox_run" not in st.session_state:
        st.session_state.phyphox_run = 0

    col1, col2 = st.columns(2)
    with col1:
        acc_file = st.file_uploader(
            "Accelerometer CSV",
            type=["csv"],
            key=f"acc_upload_{st.session_state.phyphox_run}",
            help="Columns: Time (s), X (m/s²), Y (m/s²), Z (m/s²)",
        )
    with col2:
        gyro_file = st.file_uploader(
            "Gyroscope CSV",
            type=["csv"],
            key=f"gyro_upload_{st.session_state.phyphox_run}",
            help="Columns: Time (s), X (rad/s), Y (rad/s), Z (rad/s)",
        )

    if acc_file is not None or gyro_file is not None:
        if st.button("Clear / new recording"):
            st.session_state.phyphox_run += 1
            st.rerun()

    if acc_file is None or gyro_file is None:
        st.info("Upload both files to continue.")
        return

    try:
        import io as _io
        _raw = acc_file.read()
        acc_file.seek(0)
        _df = pd.read_csv(_io.StringIO(_raw.decode("utf-8") if isinstance(_raw, bytes) else _raw))
        _df.columns = [c.strip('"').strip() for c in _df.columns]
        _num_cols = [c for c in _df.columns if c != "Time (s)"]
        _mag = float((_df[_num_cols].apply(pd.to_numeric, errors="coerce") ** 2).sum(axis=1).mean() ** 0.5)
        if _mag < 3.0:
            st.error(
                f"Mean accelerometer magnitude is only {_mag:.2f} m/s² — this looks like "
                "'Acceleration **without** g'. Please re-record using **'Acceleration (with g)'** "
                "so gravity is included in the signal."
            )
            return
        else:
            st.caption(f"Signal check: mean magnitude = {_mag:.2f} m/s² (gravity present)")

        # Detect Phyphox pause/resume sessions: large gaps in the time column
        # mean different activities were stitched together in one CSV.
        _t = pd.to_numeric(_df["Time (s)"], errors="coerce").dropna().values
        _gaps = np.diff(_t)
        _expected_dt = np.median(_gaps)
        _session_breaks = int(np.sum(_gaps > _expected_dt * 20))
        if _session_breaks > 0:
            st.warning(
                f"**{_session_breaks} session break(s) detected** in this recording. "
                "Phyphox accumulates data across pause/resume cycles — your CSV contains "
                f"{_session_breaks + 1} separate recordings stitched together. "
                "Only windows from a single activity will predict correctly. "
                "To fix: tap the **trash icon** in Phyphox to clear data before each new recording."
            )
    except Exception:
        pass

    try:
        with st.spinner("Extracting 561 features from sensor data…"):
            features, pipeline_warnings = process_phyphox_files(acc_file, gyro_file)
    except ValueError as err:
        st.error(str(err))
        return
    except Exception as err:
        st.error(f"Unexpected error during feature extraction: {err}")
        return

    for w in pipeline_warnings:
        st.warning(w)

    n_windows = len(features)
    duration_s = (n_windows - 1) * (STEP / FS) + (WINDOW / FS)

    c1, c2, c3 = st.columns(3)
    c1.metric("Windows extracted", n_windows)
    c2.metric("Approx. duration", f"{duration_s:.1f} s")
    c3.metric("Features per window", 561)
    st.caption(
        f"Each window = {WINDOW / FS:.2f} s at {FS} Hz · "
        f"50% overlap ({STEP / FS:.2f} s hop)"
    )

    if os.path.exists(norm_params_path):
        min_vals, max_vals = _load_norm_params(norm_params_path)
        features = _normalize(features, min_vals, max_vals)
        st.caption(
            "Accelerometer converted from m/s² to g units to match UCI training data. "
            "Features scaled to [−1, 1] using physical-unit min/max computed from the "
            "UCI HAR training set raw inertial signals."
        )
    else:
        st.warning(
            "norm_params.json not found: features are in physical units. "
            "Predictions will be unreliable until normalisation is applied."
        )

    if ffn_status != "ready" and cnn_status != "ready":
        st.warning("Models not loaded: cannot predict yet.")
        return

    st.markdown("---")
    st.subheader("Model comparison")

    left, right = st.columns(2)

    def _render_model_col(col, model, status, name):
        with col:
            st.markdown(f"#### {name}")
            if status != "ready":
                st.error(f"Model not loaded: {status}")
                return

            probs_all = model.predict(features, verbose=0)       # (n_windows, 6)
            pred_labels = [LABEL_MAP[int(np.argmax(p))] for p in probs_all]

            from collections import Counter
            # Skip first and last window for the final vote: these are typically
            # contaminated by recording start/stop transients (person not yet
            # in full motion, or the gravity filter still warming up).
            core = probs_all[1:-1] if n_windows > 3 else probs_all
            core_labels = [LABEL_MAP[int(np.argmax(p))] for p in core]
            vote = Counter(core_labels).most_common(1)[0][0]
            avg_conf = float(np.mean(np.max(core, axis=1))) * 100

            st.success(f"**{vote}** · {avg_conf:.1f}% avg confidence")
            st.markdown(f"_{EXPLANATIONS[vote]}_")

            if n_windows > 1:
                with st.expander(f"Per-window breakdown ({n_windows} windows)"):
                    rows = []
                    for i, (p, label) in enumerate(zip(probs_all, pred_labels)):
                        t_start = i * STEP / FS
                        is_edge = (i == 0 or i == n_windows - 1) and n_windows > 3
                        rows.append({
                            "Window": i + 1,
                            "Time (s)": f"{t_start:.1f}–{t_start + WINDOW/FS:.1f}",
                            "Prediction": label + (" *" if is_edge else ""),
                            "Confidence": f"{float(np.max(p))*100:.1f}%",
                        })
                    st.dataframe(pd.DataFrame(rows), use_container_width=True)
                    if n_windows > 3:
                        st.caption("* Edge windows excluded from overall vote (recording start/stop transient).")

            mean_probs = core.mean(axis=0)
            st.markdown("**Average confidence across all classes**")
            st.bar_chart(pd.DataFrame(
                {"Confidence (%)": [float(mean_probs[i]) * 100 for i in range(6)]},
                index=[LABEL_MAP[i] for i in range(6)],
            ))

    _render_model_col(left,  ffn_model, ffn_status, "Feedforward Network")
    _render_model_col(right, cnn_model, cnn_status, "1D Convolutional Network")