SwiMate — Aplikasi Pembantu Latihan Renang

SwiMate membantu perenang memantau performa latihan secara real-time menggunakan GPS, sensor gerak (accelerometer + gyroscope), dan model machine learning untuk mendeteksi gaya renang. README ini diperbarui untuk memberikan penjelasan mendetail mengenai arsitektur frontend (Flutter), backend (Supabase / PostgreSQL) dan pipeline machine learning (training → TFLite → inference), lengkap dengan contoh, checklist dan langkah debugging yang umum dijumpai saat ujian.

Demo

Video demo: https://drive.google.com/file/d/1MjcQMe2_oCaYc2Xd8yo435FslvocRyuB/view?usp=sharing

Ringkasan Fitur

• Menghitung pace (real-time) menggunakan GPS dan stopwatch
• Deteksi stroke otomatis dari sensor gerak
• Estimasi kalori berbasis jarak, durasi, dan stroke
• Analisis efisiensi renang (Stroke Rate, Stroke Length, SWOLF)
• Deteksi gaya renang (Freestyle, Backstroke, Breaststroke, Butterfly) menggunakan model ML
• Penyimpanan sesi & breakdown gaya per sesi ke Supabase (PostgreSQL)

Arsitektur tingkat tinggi

Frontend (Flutter) ⇄ Supabase SDK (supabase_flutter) ⇄ Supabase API (HTTP) ⇄ PostgreSQL (tabel: users, activity_table, activity_segments)

• Autentikasi: Supabase Auth (JWT) — client menerima session (access + refresh token) melalui SDK.
• Database: PostgreSQL yang dikelola Supabase. Gunakan Row Level Security (RLS) untuk membatasi akses berdasarkan auth.uid().
• ML: Model dilatih di folder machine_learning/ (Keras). Hasil training diekspor ke .h5 dan dikonversi ke .tflite untuk digunakan di aplikasi.

Struktur penting di repo

• applications/ — kode aplikasi Flutter (source utama: applications/lib/).

  • applications/lib/main.dart — entry point app (pastikan Supabase di-initialize di sini).
  • applications/lib/pages/ — screens (signin, signup, record, history, age/height/weight, dll).
  • applications/lib/services/ — services seperti activity_service.dart (query Supabase) dan swimming_style_detector.dart (tflite inference wrapper).
  • applications/assets/models/ — model tflite dan (opsional) labels.json/labels.npy yang dipaket ke app.

• machine_learning/ — skrip training, dataset, dan model hasil pelatihan (model.h5, labels.npy, runs/).

  • swimate_train.py, swimate_train_with_unknown.py, swimate_train_default.py — skrip pelatihan dengan variasi augmentasi/setting.
  • dataset/ — CSV sumber per gaya renang.
  • runs/ & runs_no_aug/ — artifacts pelatihan (model.h5, tflite, labels.npy).

Frontend (Flutter) — penjelasan lengkap

1. Inisialisasi Supabase

Sebelum kode lain yang mengakses Supabase.instance.client, inisialisasi Supabase harus dipanggil di main.dart:

import 'package:flutter/material.dart';
import 'package:supabase_flutter/supabase_flutter.dart';

Future<void> main() async {
	WidgetsFlutterBinding.ensureInitialized();

	await Supabase.initialize(
		url: 'https://<PROJECT_REF>.supabase.co',
		anonKey: const String.fromEnvironment('SUPABASE_ANON_KEY'),
	);

	runApp(const SwimScienceApp());
}

• Gunakan environment variables atau CI secrets (mis. flutter build --dart-define=SUPABASE_ANON_KEY=xxx) untuk menghindari menyimpan kunci di repo.

2. Auth flow (ringkas)

• Login email/password: supabase.auth.signInWithPassword(email:..., password:...).
• OAuth (Google): dapat menggunakan signInWithIdToken setelah mendapatkan idToken dari Google SDK.
• SDK menyimpan session; SDK otomatis menyertakan Authorization header pada request ke API Supabase.

3. Contoh CRUD (Dart)

• Ambil semua aktivitas user (history):

final userId = Supabase.instance.client.auth.currentUser?.id;
final res = await Supabase.instance.client
	.from('activity_table')
	.select('*')
	.eq('user_id', userId)
	.order('started_at', ascending: false);

// response biasanya List<Map<String,dynamic>>

• Simpan activity dan segments (non-atomic example):

final insertRes = await Supabase.instance.client
	.from('activity_table')
	.insert({
		'user_id': userId,
		'started_at': DateTime.now().toIso8601String(),
		'distance_m': 0,
	})
	.select()
	.single();

final activityId = insertRes['id'];

await Supabase.instance.client.from('activity_segments').insert({
	'activity_id': activityId,
	'start_offset_seconds': 0,
	'end_offset_seconds': 30,
	'stroke_type': 'freestyle',
	'confidence': 0.92,
});

• Catatan: operasi multi-table di client tidak bersifat atomik — untuk atomicity buat RPC (Postgres function) di Supabase.

4. Best-practices frontend

• Debounce/throttle penulisan ke DB (mis. saat user scroll untuk mengubah age) — hindari terlalu banyak request.
• Tangani tipe data dengan hati-hati: API bisa mengembalikan jenis yang berbeda tergantung versi SDK (Map atau PostgrestResponse). Gunakan safe-casts.
• Pastikan nama kolom di query sesuai dengan skema DB (user_id vs id_user).

Backend (Supabase / PostgreSQL) — penjelasan lengkap

1. Skema tabel (contoh)

• users / profiles (terdiri dari data user tambahan):

  • id uuid PRIMARY KEY (supabase auth UID)
  • email text
  • username text
  • height numeric
  • weight numeric
  • created_at timestamptz

• activity_table:

  • id uuid PRIMARY KEY
  • user_id uuid REFERENCES profiles(id)
  • started_at timestamptz
  • ended_at timestamptz
  • distance_m numeric
  • calories numeric
  • strokes_count int
  • timestamp timestamptz -- (opsional) record created timestamp

• activity_segments:

  • id uuid PRIMARY KEY
  • activity_id uuid REFERENCES activity_table(id)
  • start_offset_seconds numeric
  • end_offset_seconds numeric
  • stroke_type text
  • confidence numeric
  • created_at timestamptz

2. Row Level Security (RLS) dan policy

• Supabase memberikan kemampuan RLS — sangat penting untuk aplikasi yang memanggil DB langsung dari client.
• Contoh kebijakan minimal (ganti user_id sesuai kolom Anda):

ALTER TABLE public.activity_table ENABLE ROW LEVEL SECURITY;

CREATE POLICY "Users can select own activities"
	ON public.activity_table
	FOR SELECT
	USING (auth.uid() = user_id);

CREATE POLICY "Users can insert own activities"
	ON public.activity_table
	FOR INSERT
	WITH CHECK (auth.uid() = user_id);

CREATE POLICY "Users can update own activities"
	ON public.activity_table
	FOR UPDATE
	USING (auth.uid() = user_id)
	WITH CHECK (auth.uid() = user_id);

CREATE POLICY "Users can delete own activities"
	ON public.activity_table
	FOR DELETE
	USING (auth.uid() = user_id);

• Aktifkan kebijakan serupa untuk activity_segments dan profiles (jika perlu).

3. Atomic insert (activity + segments)

• Buat fungsi PL/pgSQL (RPC) yang melakukan insert activity dan segments dalam satu transaksi, lalu panggil via supabase.rpc('create_activity_with_segments', params) dari client.

Contoh ringkas:

CREATE FUNCTION public.create_activity_with_segments(p_activity json, p_segments json)
RETURNS json AS $$
DECLARE new_activity_id uuid;
BEGIN
	INSERT INTO activity_table (user_id, started_at, distance_m)
	VALUES ((p_activity->>'user_id')::uuid, (p_activity->>'started_at')::timestamptz, (p_activity->>'distance_m')::numeric)
	RETURNING id INTO new_activity_id;

	INSERT INTO activity_segments (activity_id, start_offset_seconds, end_offset_seconds, stroke_type, confidence)
	SELECT new_activity_id, (s->>'start_offset_seconds')::numeric, (s->>'end_offset_seconds')::numeric, s->>'stroke_type', (s->>'confidence')::numeric
	FROM json_array_elements(p_segments) AS s;

	RETURN json_build_object('activity_id', new_activity_id);
END;
$$ LANGUAGE plpgsql SECURITY DEFINER;

4. Security notes

• Jangan simpan service_role key di client. Gunakan hanya anonKey untuk client.
• Validasi input di DB (constraints) untuk pertahanan lapis kedua.

Machine Learning — penjelasan lengkap

1. Data & preprocessing

• Dataset CSV ada di machine_learning/dataset/ (file per gaya). Setiap CSV berisi kolom sensor seperti accel_x, accel_y, accel_z, gyro_x, gyro_y, gyro_z dan label.
• Preprocessing utama di skrip training:
  • Windowing (mis. SAMPLING_HZ=10, WINDOW_SEC=4 → window_size=40 sampel)
  • Stride (overlap)
  • Normalisasi/standardisasi (pakai mean/std dari training set — simpan stat jika perlu)
  • Augmentasi (opsional): gaussian noise, magnitude scaling, axis scaling, random dropout

2. Training

• Skrip: swimate_train.py, swimate_train_with_unknown.py, swimate_train_default.py.
• Model: Transformer-like architecture (MultiHeadAttention, LayerNorm, FeedForward, GlobalAvgPool). Output: softmax K-class.
• Training menyimpan labels.npy (urutan kelas) dan model .h5 di folder runs/.

3. Evaluasi

• Confusion matrix disimpan/ditampilkan: baris = ground truth, kolom = prediksi.
• Perhatikan metrik per kelas (accuracy, precision, recall, F1) dan latency inference untuk deployment.

4. Export ke TFLite

• Konversi .h5 → .tflite (contoh):

import tensorflow as tf
model = tf.keras.models.load_model('swimate_model.h5')
converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()
open('swimate_model.tflite', 'wb').write(tflite_model)

• Pertimbangkan quantization (post-training) untuk mengurangi ukuran dan latency.

5. Integrasi ke Flutter

• Salin swimate_model.tflite dan labels.npy/labels.json ke applications/assets/models/.
• Tambahkan assets di applications/pubspec.yaml (pastikan path benar).
• Di runtime, swimming_style_detector.dart memuat model via tflite_flutter, menyiapkan input tensor sesuai shape, menjalankan interpreter, lalu memetakan indeks ke label.

6. Hal-hal penting untuk inference

• Sampling rate dan window size harus sama dengan setting saat training.
• Preprocessing (order channel, normalization) harus identik.
• Labels ordering harus sinkron (simpan labels.npy yang sama dengan model) — jika tidak, hasil label akan keliru.

Cara cepat menjalankan & contoh perintah (fish shell)

Frontend (Flutter):

cd applications
flutter pub get
# Jalankan di device/emulator
flutter run

# Analisis kode statis
flutter analyze

# Build release (Android)
flutter build apk --release

Machine learning (Python):

cd machine_learning
python3 -m venv .venv
source .venv/bin/activate.fish
pip install -r requirements.txt  # jika tersedia
python swimate_train_default.py

Jika hasil training menghasilkan runs/<timestamp>/swimate_model.h5 dan .tflite, salin .tflite ke applications/assets/models/ dan tambahkan path di applications/pubspec.yaml.