Show page

Differences

This shows you the differences between two versions of the page.

@@ Line 1: / Line 1: @@
+<code python train_model.py>
+import pandas as pd
+import numpy as np
+from sklearn.linear_model import LogisticRegression
+from sklearn.metrics import classification_report
+IDLE_CSV  = "raw_data_label0.csv"
+SHAKE_CSV = "raw_data_label1.csv"
+WINDOW_SIZE = 50   # 0.5s window @ ~100 Hz
+STEP_SIZE   = 25   # 50% overlap
+MODEL_HEADER_OUT = "include/model.h"
+def load_and_merge():
+    df0 = pd.read_csv(IDLE_CSV)
+    df1 = pd.read_csv(SHAKE_CSV)
+    # Just sanity: enforce int label for safety
+    df0["label"] = 0
+    df1["label"] = 1
+    df = pd.concat([df0, df1], ignore_index=True)
+    # Sort by timestamp is not strictly required, but helps consistent windows
+    df = df.sort_values(by="timestamp_ms").reset_index(drop=True)
+    return df
+def extract_features(df):
+    """
+    df columns: timestamp_ms, ax_g, ay_g, az_g, label
+    We'll slide windows, compute features, and assign the majority label.
+    """
+    feats = []
+    ax = df["ax_g"].to_numpy()
+    ay = df["ay_g"].to_numpy()
+    az = df["az_g"].to_numpy()
+    lbl = df["label"].to_numpy()
+    mag = np.sqrt(ax**2 + ay**2 + az**2)
+    n = len(df)
+    idx = 0
+    while idx + WINDOW_SIZE <= n:
+        sl = slice(idx, idx+WINDOW_SIZE)
+        win_mag = mag[sl]
+        win_lbl = lbl[sl]
+        mean_mag = np.mean(win_mag)
+        std_mag  = np.std(win_mag)
+        p2p_mag  = np.max(win_mag) - np.min(win_mag)
+        # majority/avg label in this window
+        label_window = int(round(np.mean(win_lbl)))
+        feats.append({
+            "mean_mag": mean_mag,
+            "std_mag":  std_mag,
+            "p2p_mag":  p2p_mag,
+            "label":    label_window
+        })
+        idx += STEP_SIZE
+    return pd.DataFrame(feats)
+def main():
+    df = load_and_merge()
+    feats = extract_features(df)
+    X = feats[["mean_mag", "std_mag", "p2p_mag"]].to_numpy()
+    y = feats["label"].to_numpy()
+    clf = LogisticRegression()
+    clf.fit(X, y)
+    y_pred = clf.predict(X)
+    print(classification_report(y, y_pred))
+    w = clf.coef_[0]      # [w1, w2, w3]
+    b = clf.intercept_[0] # bias
+    print("Weights:", w)
+    print("Bias:", b)
+    with open(MODEL_HEADER_OUT, "w") as f:
+        f.write("// Auto-generated by train_model.py\n")
+        f.write("#pragma once\n\n")
+        f.write("struct ShakeModel {\n")
+        f.write("    static constexpr float W0 = %.8ff;\n" % b)
+        f.write("    static constexpr float W1 = %.8ff;\n" % w[0])
+        f.write("    static constexpr float W2 = %.8ff;\n" % w[1])
+        f.write("    static constexpr float W3 = %.8ff;\n" % w[2])
+        f.write("};\n\n")
+        f.write("// Feature extraction params\n")
+        f.write("#define WINDOW_SIZE %d\n" % WINDOW_SIZE)
+        f.write("#define STEP_SIZE   %d\n" % STEP_SIZE)
+if __name__ == "__main__":
+    main()
+</code>