kurzti88066/Fahrsimulator_MSY2526_AI
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Compare commits

base: kurzti88066:13bd76631f8341f76439943bfe7cc151a3ec1a16
Branches Tags
kurzti88066:main
..
compare: kurzti88066:de0084dc09acbed3c3a2ecfe8e284bcb3026d844
Branches Tags
kurzti88066:main

No commits in common. "13bd76631f8341f76439943bfe7cc151a3ec1a16" and "de0084dc09acbed3c3a2ecfe8e284bcb3026d844" have entirely different histories.
13bd76631f ... de0084dc09

7 changed files with 2619 additions and 119 deletions
Show Stats Expand all files Collapse all files

331
model_training/DeepSVDD/deepSVDD.ipynb
View File

@ -41,6 +41,7 @@
"import time\n", "import time\n",
"base_dir = os.path.abspath(os.path.join(os.getcwd(), \"..\"))\n", "base_dir = os.path.abspath(os.path.join(os.getcwd(), \"..\"))\n",
"sys.path.append(base_dir)\n", "sys.path.append(base_dir)\n",
"print(base_dir)\n",
"\n", "\n",
"from Fahrsimulator_MSY2526_AI.model_training.tools import evaluation_tools, scaler, mad_outlier_removal, performance_split\n", "from Fahrsimulator_MSY2526_AI.model_training.tools import evaluation_tools, scaler, mad_outlier_removal, performance_split\n",
"from sklearn.preprocessing import StandardScaler, MinMaxScaler\n", "from sklearn.preprocessing import StandardScaler, MinMaxScaler\n",
@ -50,7 +51,7 @@
"import tensorflow as tf\n", "import tensorflow as tf\n",
"from tensorflow.keras import layers, models, regularizers\n", "from tensorflow.keras import layers, models, regularizers\n",
"import pickle\n", "import pickle\n",
"from sklearn.metrics import (accuracy_score, auc, roc_curve, f1_score) " "from sklearn.metrics import (roc_auc_score, accuracy_score, precision_score, recall_score, f1_score, confusion_matrix, classification_report, balanced_accuracy_score, ConfusionMatrixDisplay, auc, roc_curve) "
] ]
}, },
{ {
@ -88,40 +89,15 @@
"id": "f00a477c", "id": "f00a477c",
"metadata": {}, "metadata": {},
"source": [ "source": [
"### Configuration of paths and data preprocessing" "### Data Preprocessing"
] ]
}, },
{ {
"cell_type": "code", "cell_type": "markdown",
"execution_count": null, "id": "504c1df7",
"id": "5136fcec",
"metadata": {}, "metadata": {},
"outputs": [],
"source": [ "source": [
"# TODO: set path where to save normalizer\n", "Laden der Daten"
"normalizer_path=Path('.pkl') # TODO: set manually"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "c2115f65",
"metadata": {},
"outputs": [],
"source": [
"performance_path = Path(r\".csv\") # TODO: set manually\n",
"performance_df = pd.read_csv(performance_path)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "559eb8d2",
"metadata": {},
"outputs": [],
"source": [
"encoder_save_path = Path('.keras') # TODO: set manually\n",
"deep_svdd_save_path = Path('.keras') # TODO: set manually"
] ]
}, },
{ {
@ -131,7 +107,8 @@
"metadata": {}, "metadata": {},
"outputs": [], "outputs": [],
"source": [ "source": [
"dataset_path = Path(r\".parquet\") # TODO: set manually" "dataset_path = Path(r\"data-paulusjafahrsimulator-gpu/new_datasets/combined_dataset_25hz.parquet\")\n",
"# dataset_path = Path(r\"/home/jovyan/data-paulusjafahrsimulator-gpu/new_datasets/120s_combined_dataset_25hz.parquet\")"
] ]
}, },
{ {
@ -144,6 +121,17 @@
"df = pd.read_parquet(path=dataset_path)" "df = pd.read_parquet(path=dataset_path)"
] ]
}, },
{
"cell_type": "code",
"execution_count": null,
"id": "c2115f65",
"metadata": {},
"outputs": [],
"source": [
"performance_path = Path(r\"/home/jovyan/data-paulusjafahrsimulator-gpu/subject_performance/3new_au_performance.csv\")\n",
"performance_df = pd.read_csv(performance_path)"
]
},
{ {
"cell_type": "markdown", "cell_type": "markdown",
"id": "c045c46d", "id": "c045c46d",
@ -227,7 +215,7 @@
"\n", "\n",
"data = pd.concat([low, high], ignore_index=True)\n", "data = pd.concat([low, high], ignore_index=True)\n",
"df = data.drop_duplicates()\n", "df = data.drop_duplicates()\n",
"df = df.dropna()\n", "\n",
"print(\"Label distribution:\")\n", "print(\"Label distribution:\")\n",
"print(df[\"label\"].value_counts())" "print(df[\"label\"].value_counts())"
] ]
@ -287,6 +275,210 @@
"Normalization" "Normalization"
] ]
}, },
{
"cell_type": "code",
"execution_count": null,
"id": "acec4a03",
"metadata": {},
"outputs": [],
"source": [
"def fit_normalizer(train_data, au_columns, method='standard', scope='global'):\n",
" \"\"\"\n",
" Fit normalization scalers on training data.\n",
" \n",
" Parameters:\n",
" -----------\n",
" train_data : pd.DataFrame\n",
" Training dataframe with AU columns and subjectID\n",
" au_columns : list\n",
" List of AU column names to normalize\n",
" method : str, default='standard'\n",
" Normalization method: 'standard' for StandardScaler or 'minmax' for MinMaxScaler\n",
" scope : str, default='global'\n",
" Normalization scope: 'subject' for per-subject or 'global' for across all subjects\n",
" \n",
" Returns:\n",
" --------\n",
" dict\n",
" Dictionary containing fitted scalers and statistics for new subjects\n",
" \"\"\"\n",
" if method == 'standard':\n",
" Scaler = StandardScaler\n",
" elif method == 'minmax':\n",
" Scaler = MinMaxScaler\n",
" else:\n",
" raise ValueError(\"method must be 'standard' or 'minmax'\")\n",
" \n",
" scalers = {}\n",
" if scope == 'subject':\n",
" # Fit one scaler per subject\n",
" subject_stats = []\n",
" \n",
" for subject in train_data['subjectID'].unique():\n",
" subject_mask = train_data['subjectID'] == subject\n",
" scaler = Scaler()\n",
" scaler.fit(train_data.loc[subject_mask, au_columns].values)\n",
" scalers[subject] = scaler\n",
" \n",
" # Store statistics for averaging\n",
" if method == 'standard':\n",
" subject_stats.append({\n",
" 'mean': scaler.mean_,\n",
" 'std': scaler.scale_\n",
" })\n",
" elif method == 'minmax':\n",
" subject_stats.append({\n",
" 'min': scaler.data_min_,\n",
" 'max': scaler.data_max_\n",
" })\n",
" \n",
" # Calculate average statistics for new subjects\n",
" if method == 'standard':\n",
" avg_mean = np.mean([s['mean'] for s in subject_stats], axis=0)\n",
" avg_std = np.mean([s['std'] for s in subject_stats], axis=0)\n",
" fallback_scaler = StandardScaler()\n",
" fallback_scaler.mean_ = avg_mean\n",
" fallback_scaler.scale_ = avg_std\n",
" fallback_scaler.var_ = avg_std ** 2\n",
" fallback_scaler.n_features_in_ = len(au_columns)\n",
" elif method == 'minmax':\n",
" avg_min = np.mean([s['min'] for s in subject_stats], axis=0)\n",
" avg_max = np.mean([s['max'] for s in subject_stats], axis=0)\n",
" fallback_scaler = MinMaxScaler()\n",
" fallback_scaler.data_min_ = avg_min\n",
" fallback_scaler.data_max_ = avg_max\n",
" fallback_scaler.data_range_ = avg_max - avg_min\n",
" fallback_scaler.scale_ = 1.0 / fallback_scaler.data_range_\n",
" fallback_scaler.min_ = -avg_min * fallback_scaler.scale_\n",
" fallback_scaler.n_features_in_ = len(au_columns)\n",
" \n",
" scalers['_fallback'] = fallback_scaler\n",
" \n",
" elif scope == 'global':\n",
" # Fit one scaler for all subjects\n",
" scaler = Scaler()\n",
" scaler.fit(train_data[au_columns].values)\n",
" scalers['global'] = scaler\n",
" \n",
" else:\n",
" raise ValueError(\"scope must be 'subject' or 'global'\")\n",
" \n",
" return {'scalers': scalers, 'method': method, 'scope': scope}\n",
"\n",
"def apply_normalizer(data, columns, normalizer_dict):\n",
" \"\"\"\n",
" Apply fitted normalization scalers to data.\n",
" \n",
" Parameters:\n",
" -----------\n",
" data : pd.DataFrame\n",
" Dataframe with AU columns and subjectID\n",
" au_columns : list\n",
" List of AU column names to normalize\n",
" normalizer_dict : dict\n",
" Dictionary containing fitted scalers from fit_normalizer()\n",
" \n",
" Returns:\n",
" --------\n",
" pd.DataFrame\n",
" DataFrame with normalized AU columns\n",
" \"\"\"\n",
" normalized_data = data.copy()\n",
" scalers = normalizer_dict['scalers']\n",
" scope = normalizer_dict['scope']\n",
" normalized_data[columns] = normalized_data[columns].astype(np.float64)\n",
"\n",
" if scope == 'subject':\n",
" # Apply per-subject normalization\n",
" for subject in data['subjectID'].unique():\n",
" subject_mask = data['subjectID'] == subject\n",
" \n",
" # Use the subject's scaler if available, otherwise use fallback\n",
" if subject in scalers:\n",
" scaler = scalers[subject]\n",
" else:\n",
" # Use averaged scaler for new subjects\n",
" scaler = scalers['_fallback']\n",
" print(f\"Info: Subject {subject} not in training data. Using averaged scaler from training subjects.\")\n",
" \n",
" normalized_data.loc[subject_mask, columns] = scaler.transform(\n",
" data.loc[subject_mask, columns].values\n",
" )\n",
" \n",
" elif scope == 'global':\n",
" # Apply global normalization\n",
" scaler = scalers['global']\n",
" normalized_data[columns] = scaler.transform(data[columns].values)\n",
" \n",
" return normalized_data\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "53c6ee6f",
"metadata": {},
"outputs": [],
"source": [
"def save_normalizer(normalizer_dict, filepath):\n",
" \"\"\"\n",
" Save fitted normalizer to disk.\n",
"\n",
" Parameters:\n",
" -----------\n",
" normalizer_dict : dict\n",
" Dictionary containing fitted scalers from fit_normalizer()\n",
" filepath : str\n",
" Path to save the normalizer (e.g., 'normalizer.pkl')\n",
" \"\"\"\n",
" # Create directory if it does not exist\n",
" dirpath = os.path.dirname(filepath)\n",
" if dirpath:\n",
" os.makedirs(dirpath, exist_ok=True)\n",
"\n",
" with open(filepath, 'wb') as f:\n",
" pickle.dump(normalizer_dict, f)\n",
"\n",
" print(f\"Normalizer saved to {filepath}\")\n",
"\n",
"def load_normalizer(filepath):\n",
" \"\"\"\n",
" Load fitted normalizer from disk.\n",
" \n",
" Parameters:\n",
" -----------\n",
" filepath : str\n",
" Path to the saved normalizer file\n",
" \n",
" Returns:\n",
" --------\n",
" dict\n",
" Dictionary containing fitted scalers\n",
" \"\"\"\n",
" with open(filepath, 'rb') as f:\n",
" normalizer_dict = pickle.load(f)\n",
" print(f\"Normalizer loaded from {filepath}\")\n",
" return normalizer_dict"
]
},
{
"cell_type": "markdown",
"id": "7280f64f",
"metadata": {},
"source": [
"save Normalizer"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "8420afc2",
"metadata": {},
"outputs": [],
"source": [
"normalizer_path=Path('data-paulusjafahrsimulator-gpu/saved_models/deepsvdd_save/normalizer_min_max_global.pkl')"
]
},
{ {
"cell_type": "code", "cell_type": "code",
"execution_count": null, "execution_count": null,
@ -303,10 +495,8 @@
"print(len(eye_cols))\n", "print(len(eye_cols))\n",
"all_signal_columns = face_au_cols+eye_cols\n", "all_signal_columns = face_au_cols+eye_cols\n",
"print(len(all_signal_columns))\n", "print(len(all_signal_columns))\n",
"\n", "normalizer = fit_normalizer(train_df, all_signal_columns, method='minmax', scope='global')\n",
"# fit and save normalizer\n", "save_normalizer(normalizer, normalizer_path )"
"normalizer = scaler.fit_normalizer(train_df, all_signal_columns, method='minmax', scope='global')\n",
"scaler.save_normalizer(normalizer, normalizer_path )"
] ]
}, },
{ {
@ -316,11 +506,11 @@
"metadata": {}, "metadata": {},
"outputs": [], "outputs": [],
"source": [ "source": [
"normalizer = scaler.load_normalizer(normalizer_path)\n", "normalizer = load_normalizer(normalizer_path)\n",
"# Apply normalization to all sets\n", "# 3. Apply normalization to all sets\n",
"train_df_norm = scaler.apply_normalizer(train_df, all_signal_columns, normalizer)\n", "train_df_norm = apply_normalizer(train_df, all_signal_columns, normalizer)\n",
"val_df_norm = scaler.apply_normalizer(val_df, all_signal_columns, normalizer)\n", "val_df_norm = apply_normalizer(val_df, all_signal_columns, normalizer)\n",
"test_df_norm = scaler.apply_normalizer(test_df, all_signal_columns, normalizer)" "test_df_norm = apply_normalizer(test_df, all_signal_columns, normalizer)"
] ]
}, },
{ {
@ -376,13 +566,13 @@
"def build_intermediate_fusion_autoencoder(\n", "def build_intermediate_fusion_autoencoder(\n",
" input_dim_mod1=15,\n", " input_dim_mod1=15,\n",
" input_dim_mod2=20,\n", " input_dim_mod2=20,\n",
" encoder_hidden_dim_mod1=12, # TODO: set manually\n", " encoder_hidden_dim_mod1=12, # individuell\n",
" encoder_hidden_dim_mod2=20, # TODO: set manually\n", " encoder_hidden_dim_mod2=20, # individuell\n",
" latent_dim=6, # TODO: set manually\n", " latent_dim=6, # Änderung: Bottleneck vergrößert für stabilere Repräsentation\n",
" dropout_rate=0.4, # TODO: set manually\n", " dropout_rate=0.4, # Dropout in Hidden Layers\n",
" neg_slope=0.1, # TODO: set manually\n", " neg_slope=0.1,\n",
" weight_decay=1e-4, # TODO: set manually\n", " weight_decay=1e-4,\n",
" decoder_hidden_dims=[16, 32] # TODO: set manually\n", " decoder_hidden_dims=[16, 32] # Änderung: Decoder größer für bessere Rekonstruktion\n",
"):\n", "):\n",
" \"\"\"\n", " \"\"\"\n",
" Verbesserter Intermediate-Fusion Autoencoder für Deep SVDD.\n", " Verbesserter Intermediate-Fusion Autoencoder für Deep SVDD.\n",
@ -407,10 +597,10 @@
" kernel_regularizer=l2\n", " kernel_regularizer=l2\n",
" )(x1_in)\n", " )(x1_in)\n",
" e1 = act(e1)\n", " e1 = act(e1)\n",
" e1 = layers.Dropout(dropout_rate)(e1) \n", " e1 = layers.Dropout(dropout_rate)(e1) # Dropout nur hier\n",
"\n", "\n",
" e1 = layers.Dense(\n", " e1 = layers.Dense(\n",
" 16, \n", " 16, # Änderung: Hidden Layer größer für stabilere Fusion\n",
" use_bias=False,\n", " use_bias=False,\n",
" kernel_regularizer=l2\n", " kernel_regularizer=l2\n",
" )(e1)\n", " )(e1)\n",
@ -423,20 +613,20 @@
" kernel_regularizer=l2\n", " kernel_regularizer=l2\n",
" )(x2_in)\n", " )(x2_in)\n",
" e2 = act(e2)\n", " e2 = act(e2)\n",
" e2 = layers.Dropout(dropout_rate)(e2) \n", " e2 = layers.Dropout(dropout_rate)(e2) # Dropout nur hier\n",
"\n", "\n",
" e2 = layers.Dense(\n", " e2 = layers.Dense(\n",
" 16, \n", " 16, # Änderung: Hidden Layer größer\n",
" use_bias=False,\n", " use_bias=False,\n",
" kernel_regularizer=l2\n", " kernel_regularizer=l2\n",
" )(e2)\n", " )(e2)\n",
" e2 = act(e2)\n", " e2 = act(e2)\n",
"\n", "\n",
" # -------- Intermediate Fusion --------\n", " # -------- Intermediate Fusion --------\n",
" fused = layers.Concatenate(name=\"fusion\")([e1, e2]) # 16+16=32 dimensions\n", " fused = layers.Concatenate(name=\"fusion\")([e1, e2]) # 16+16=32 Dimensionen\n",
"\n", "\n",
" # -------- Joint Encoder / Bottleneck --------\n", " # -------- Joint Encoder / Bottleneck --------\n",
"\n", " # sinnvoll kleiner als Fusion\n",
" h = layers.Dense(\n", " h = layers.Dense(\n",
" latent_dim,\n", " latent_dim,\n",
" use_bias=False,\n", " use_bias=False,\n",
@ -447,16 +637,16 @@
"\n", "\n",
" z = layers.Dense(\n", " z = layers.Dense(\n",
" latent_dim,\n", " latent_dim,\n",
" activation=None, # linear for Deep SVDD\n", " activation=None, # linear, für Deep SVDD\n",
" use_bias=False,\n", " use_bias=False,\n",
" kernel_regularizer=l2,\n", " kernel_regularizer=l2,\n",
" name=\"latent\"\n", " name=\"latent\"\n",
" )(h)\n", " )(h)\n",
"\n", " # Dropout entfernt direkt vor Bottleneck\n",
"\n", "\n",
" # -------- Decoder --------\n", " # -------- Decoder --------\n",
" d = layers.Dense(\n", " d = layers.Dense(\n",
" decoder_hidden_dims[0], \n", " decoder_hidden_dims[0], # größerer Decoder\n",
" use_bias=False,\n", " use_bias=False,\n",
" kernel_regularizer=l2\n", " kernel_regularizer=l2\n",
" )(z)\n", " )(z)\n",
@ -502,10 +692,10 @@
"model = build_intermediate_fusion_autoencoder(\n", "model = build_intermediate_fusion_autoencoder(\n",
" input_dim_mod1=len(face_au_cols),\n", " input_dim_mod1=len(face_au_cols),\n",
" input_dim_mod2=len(eye_cols),\n", " input_dim_mod2=len(eye_cols),\n",
" encoder_hidden_dim_mod1=12, # TODO: set manually\n", " encoder_hidden_dim_mod1=12, # individuell\n",
" encoder_hidden_dim_mod2=8, # TODO: set manually\n", " encoder_hidden_dim_mod2=8, # individuell\n",
" latent_dim=4,\n", " latent_dim=4,\n",
" dropout_rate=0.7, # TODO: set manually\n", " dropout_rate=0.7, # einstellbar\n",
" neg_slope=0.1,\n", " neg_slope=0.1,\n",
" weight_decay=1e-3\n", " weight_decay=1e-3\n",
")\n", ")\n",
@ -590,6 +780,7 @@
"metadata": {}, "metadata": {},
"outputs": [], "outputs": [],
"source": [ "source": [
"encoder_save_path =Path('data-paulusjafahrsimulator-gpu/saved_models/deepsvdd_save/encoder_8_deep.keras')\n",
"encoder.save(encoder_save_path)" "encoder.save(encoder_save_path)"
] ]
}, },
@ -685,6 +876,22 @@
"center = get_center(deep_svdd_net, [X_face, X_eye])" "center = get_center(deep_svdd_net, [X_face, X_eye])"
] ]
}, },
{
"cell_type": "code",
"execution_count": null,
"id": "da140072",
"metadata": {},
"outputs": [],
"source": [
"# def get_radius(nu, dataset):\n",
"# x_face, x_eye = dataset # <-- zwingend entpacken\n",
"\n",
"# dataset_tuple=[x_face, x_eye]\n",
"\n",
"# dists = dist_per_sample(deep_svdd_net.predict(dataset_tuple), center)\n",
"# return np.quantile(np.sqrt(dists), 1-nu).astype(np.float32)"
]
},
{ {
"cell_type": "code", "cell_type": "code",
"execution_count": null, "execution_count": null,
@ -737,9 +944,10 @@
" return get_radius_from_arrays(nu, X_face, X_eye)\n", " return get_radius_from_arrays(nu, X_face, X_eye)\n",
"\n", "\n",
"\n", "\n",
"nu = 0.05 # Set nu respectively\n", "nu = 0.25\n",
"\n", "\n",
"train_dataset = tf.data.Dataset.from_tensor_slices((X_face, X_eye)).shuffle(64).batch(64)\n", "train_dataset = tf.data.Dataset.from_tensor_slices((X_face, X_eye)).shuffle(64).batch(64)\n",
"# train_dataset = tf.data.Dataset.from_tensor_slices((X_face, X_eye))\n",
"\n", "\n",
"optimizer = tf.keras.optimizers.Adam(1e-3)\n", "optimizer = tf.keras.optimizers.Adam(1e-3)\n",
"train(train_dataset, epochs=150, nu=nu)\n", "train(train_dataset, epochs=150, nu=nu)\n",
@ -811,6 +1019,7 @@
"metadata": {}, "metadata": {},
"outputs": [], "outputs": [],
"source": [ "source": [
"deep_svdd_save_path =Path('data-paulusjafahrsimulator-gpu/saved_models/deepsvdd_save/deep_svdd_06.keras')\n",
"deep_svdd_net.save(deep_svdd_save_path)" "deep_svdd_net.save(deep_svdd_save_path)"
] ]
}, },

140
model_training/IsolationForest/iforest_training.ipynb
View File

@ -28,7 +28,7 @@
"sys.path.append(base_dir)\n", "sys.path.append(base_dir)\n",
"print(base_dir)\n", "print(base_dir)\n",
"\n", "\n",
"from Fahrsimulator_MSY2526_AI.model_training.tools import evaluation_tools, scaler\n", "from Fahrsimulator_MSY2526_AI.model_training.tools import evaluation_tools\n",
"from sklearn.preprocessing import StandardScaler, MinMaxScaler\n", "from sklearn.preprocessing import StandardScaler, MinMaxScaler\n",
"from sklearn.ensemble import IsolationForest\n", "from sklearn.ensemble import IsolationForest\n",
"from sklearn.model_selection import GridSearchCV, KFold\n", "from sklearn.model_selection import GridSearchCV, KFold\n",
@ -52,7 +52,7 @@
"metadata": {}, "metadata": {},
"outputs": [], "outputs": [],
"source": [ "source": [
"data_path = Path(r\".parquet\") # TODO: set manually" "data_path = Path(r\"/home/jovyan/data-paulusjafahrsimulator-gpu/new_datasets/50s_25Hz_dataset.parquet\")"
] ]
}, },
{ {
@ -115,6 +115,118 @@
"print(f\"high all: {high_all.shape}\")" "print(f\"high all: {high_all.shape}\")"
] ]
}, },
{
"cell_type": "code",
"execution_count": null,
"id": "47a0f44d",
"metadata": {},
"outputs": [],
"source": [
"def fit_normalizer(train_data, au_columns, method='standard', scope='global'):\n",
" \"\"\"\n",
" Fit normalization scalers on training data.\n",
" \n",
" Parameters:\n",
" -----------\n",
" train_data : pd.DataFrame\n",
" Training dataframe with AU columns and subjectID\n",
" au_columns : list\n",
" List of AU column names to normalize\n",
" method : str, default='standard'\n",
" Normalization method: 'standard' for StandardScaler or 'minmax' for MinMaxScaler\n",
" scope : str, default='global'\n",
" Normalization scope: 'subject' for per-subject or 'global' for across all subjects\n",
" \n",
" Returns:\n",
" --------\n",
" dict\n",
" Dictionary containing fitted scalers\n",
" \"\"\"\n",
" # Select scaler based on method\n",
" if method == 'standard':\n",
" Scaler = StandardScaler\n",
" elif method == 'minmax':\n",
" Scaler = MinMaxScaler\n",
" else:\n",
" raise ValueError(\"method must be 'standard' or 'minmax'\")\n",
" \n",
" scalers = {}\n",
" \n",
" if scope == 'subject':\n",
" # Fit one scaler per subject\n",
" for subject in train_data['subjectID'].unique():\n",
" subject_mask = train_data['subjectID'] == subject\n",
" scaler = Scaler()\n",
" scaler.fit(train_data.loc[subject_mask, au_columns])\n",
" scalers[subject] = scaler\n",
" \n",
" elif scope == 'global':\n",
" # Fit one scaler for all subjects\n",
" scaler = Scaler()\n",
" scaler.fit(train_data[au_columns])\n",
" scalers['global'] = scaler\n",
" \n",
" else:\n",
" raise ValueError(\"scope must be 'subject' or 'global'\")\n",
" \n",
" return {'scalers': scalers, 'method': method, 'scope': scope}"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "642d0017",
"metadata": {},
"outputs": [],
"source": [
"def apply_normalizer(data, au_columns, normalizer_dict):\n",
" \"\"\"\n",
" Apply fitted normalization scalers to data.\n",
" \n",
" Parameters:\n",
" -----------\n",
" data : pd.DataFrame\n",
" Dataframe with AU columns and subjectID\n",
" au_columns : list\n",
" List of AU column names to normalize\n",
" normalizer_dict : dict\n",
" Dictionary containing fitted scalers from fit_normalizer()\n",
" \n",
" Returns:\n",
" --------\n",
" pd.DataFrame\n",
" DataFrame with normalized AU columns\n",
" \"\"\"\n",
" normalized_data = data.copy()\n",
" scalers = normalizer_dict['scalers']\n",
" scope = normalizer_dict['scope']\n",
" \n",
" if scope == 'subject':\n",
" # Apply per-subject normalization\n",
" for subject in data['subjectID'].unique():\n",
" subject_mask = data['subjectID'] == subject\n",
" \n",
" # Use the subject's scaler if available, otherwise use a fitted scaler from training\n",
" if subject in scalers:\n",
" scaler = scalers[subject]\n",
" else:\n",
" # For new subjects not seen in training, use the first available scaler\n",
" # (This is a fallback - ideally all test subjects should be in training for subject-level normalization)\n",
" print(f\"Warning: Subject {subject} not found in training data. Using fallback scaler.\")\n",
" scaler = list(scalers.values())[0]\n",
" \n",
" normalized_data.loc[subject_mask, au_columns] = scaler.transform(\n",
" data.loc[subject_mask, au_columns]\n",
" )\n",
" \n",
" elif scope == 'global':\n",
" # Apply global normalization\n",
" scaler = scalers['global']\n",
" normalized_data[au_columns] = scaler.transform(data[au_columns])\n",
" \n",
" return normalized_data"
]
},
{ {
"cell_type": "markdown", "cell_type": "markdown",
"id": "697b3cf7", "id": "697b3cf7",
@ -223,7 +335,7 @@
"outputs": [], "outputs": [],
"source": [ "source": [
"# Fit normalizer on training data\n", "# Fit normalizer on training data\n",
"normalizer = scaler.fit_normalizer(train_data, cols, method='minmax', scope='global')\n", "normalizer = fit_normalizer(train_data, cols, method='minmax', scope='global')\n",
"print(\"Normalizer fitted on training data\")" "print(\"Normalizer fitted on training data\")"
] ]
}, },
@ -235,11 +347,11 @@
"outputs": [], "outputs": [],
"source": [ "source": [
"# Apply normalization to all datasets\n", "# Apply normalization to all datasets\n",
"train_normalized = scaler.apply_normalizer(train_data, cols, normalizer)\n", "train_normalized = apply_normalizer(train_data, cols, normalizer)\n",
"val_normal_normalized = scaler.apply_normalizer(val_normal_data, cols, normalizer)\n", "val_normal_normalized = apply_normalizer(val_normal_data, cols, normalizer)\n",
"val_high_normalized = scaler.apply_normalizer(val_high_data, cols, normalizer)\n", "val_high_normalized = apply_normalizer(val_high_data, cols, normalizer)\n",
"test_normal_normalized = scaler.apply_normalizer(test_normal_data, cols, normalizer)\n", "test_normal_normalized = apply_normalizer(test_normal_data, cols, normalizer)\n",
"test_high_normalized = scaler.apply_normalizer(test_high_data, cols, normalizer)\n", "test_high_normalized = apply_normalizer(test_high_data, cols, normalizer)\n",
"\n", "\n",
"print(\"Normalization applied to all datasets\")" "print(\"Normalization applied to all datasets\")"
] ]
@ -378,6 +490,18 @@
"display_name": "Python 3 (ipykernel)", "display_name": "Python 3 (ipykernel)",
"language": "python", "language": "python",
"name": "python3" "name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.10"
} }
}, },
"nbformat": 4, "nbformat": 4,

1312
model_training/VAE_SVM/AEdannSVM.ipynb Normal file
View File

File diff suppressed because it is too large Load Diff

877
model_training/VAE_SVM/vaesvm.ipynb Normal file
View File

@ -0,0 +1,877 @@
{
"cells": [
{
"cell_type": "markdown",
"id": "708c9745",
"metadata": {},
"source": [
"### Imports"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "53b10294",
"metadata": {},
"outputs": [],
"source": [
"import pandas as pd\n",
"import numpy as np\n",
"from pathlib import Path\n",
"import sys\n",
"import os\n",
"\n",
"base_dir = os.path.abspath(os.path.join(os.getcwd(), \"..\"))\n",
"sys.path.append(base_dir)\n",
"print(base_dir)\n",
"\n",
"from Fahrsimulator_MSY2526_AI.model_training.tools import evaluation_tools, scaler, mad_outlier_removal\n",
"from sklearn.preprocessing import StandardScaler, MinMaxScaler\n",
"from sklearn.svm import OneClassSVM\n",
"from sklearn.model_selection import GridSearchCV, KFold, ParameterGrid, train_test_split\n",
"import matplotlib.pyplot as plt\n",
"import tensorflow as tf\n",
"import pickle\n",
"from sklearn.metrics import (roc_auc_score, accuracy_score, precision_score, \n",
" recall_score, f1_score, confusion_matrix, classification_report) "
]
},
{
"cell_type": "markdown",
"id": "68101229",
"metadata": {},
"source": [
"### load Dataset"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "24a765e8",
"metadata": {},
"outputs": [],
"source": [
"dataset_path = Path(r\"/home/jovyan/data-paulusjafahrsimulator-gpu/first_AU_dataset/output_windowed.parquet\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "471001b0",
"metadata": {},
"outputs": [],
"source": [
"df = pd.read_parquet(path=dataset_path)"
]
},
{
"cell_type": "markdown",
"id": "0fdecdaa",
"metadata": {},
"source": [
"### Load Performance data and Subject Split"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "692d1b47",
"metadata": {},
"outputs": [],
"source": [
"performance_path = Path(r\"/home/jovyan/data-paulusjafahrsimulator-gpu/subject_performance/3new_au_performance.csv\")\n",
"performance_df = pd.read_csv(performance_path)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "ea617e3f",
"metadata": {},
"outputs": [],
"source": [
"# Subject IDs aus dem Haupt-Dataset nehmen\n",
"subjects_from_df = df[\"subjectID\"].unique()\n",
"\n",
"# Performance-Subset nur für vorhandene Subjects\n",
"perf_filtered = performance_df[\n",
" performance_df[\"subjectID\"].isin(subjects_from_df)\n",
"][[\"subjectID\", \"overall_score\"]]\n",
"\n",
"# Merge: nur Subjects, die sowohl im df als auch im Performance-CSV vorkommen\n",
"merged = (\n",
" pd.DataFrame({\"subjectID\": subjects_from_df})\n",
" .merge(perf_filtered, on=\"subjectID\", how=\"inner\")\n",
")\n",
"\n",
"# Sicherstellen, dass keine Scores fehlen\n",
"if merged[\"overall_score\"].isna().any():\n",
" raise ValueError(\"Es fehlen Score-Werte für manche Subjects.\")\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "ae43df8d",
"metadata": {},
"outputs": [],
"source": [
"merged_sorted = merged.sort_values(\"overall_score\", ascending=False).reset_index(drop=True)\n",
"\n",
"scores = merged_sorted[\"overall_score\"].values\n",
"n_total = len(merged_sorted)\n",
"n_small = n_total // 3\n",
"n_large = n_total - n_small\n",
"\n",
"# Schritt 1: zufällige Start-Aufteilung\n",
"idx = np.arange(n_total)\n",
"np.random.shuffle(idx)\n",
"\n",
"small_idx = idx[:n_small]\n",
"large_idx = idx[n_small:]\n",
"\n",
"def score_diff(small_idx, large_idx):\n",
" return abs(scores[small_idx].mean() - scores[large_idx].mean())\n",
"\n",
"diff = score_diff(small_idx, large_idx)\n",
"threshold = 0.01\n",
"max_iter = 100\n",
"count = 0\n",
"\n",
"# Schritt 2: random swaps bis Differenz klein genug\n",
"while diff > threshold and count < max_iter:\n",
" # Zwei zufällige Elemente auswählen\n",
" si = np.random.choice(small_idx)\n",
" li = np.random.choice(large_idx)\n",
" \n",
" # Tausch durchführen\n",
" new_small_idx = small_idx.copy()\n",
" new_large_idx = large_idx.copy()\n",
" \n",
" new_small_idx[new_small_idx == si] = li\n",
" new_large_idx[new_large_idx == li] = si\n",
"\n",
" # neue Differenz berechnen\n",
" new_diff = score_diff(new_small_idx, new_large_idx)\n",
"\n",
" # Swap akzeptieren, wenn es besser wird\n",
" if new_diff < diff:\n",
" small_idx = new_small_idx\n",
" large_idx = new_large_idx\n",
" diff = new_diff\n",
"\n",
" count += 1\n",
"\n",
"# Finalgruppen\n",
"group_small = merged_sorted.loc[small_idx].reset_index(drop=True)\n",
"group_large = merged_sorted.loc[large_idx].reset_index(drop=True)\n",
"\n",
"print(\"Finale Score-Differenz:\", diff)\n",
"print(\"Größe Gruppe 1:\", len(group_small))\n",
"print(\"Größe Gruppe 2:\", len(group_large))\n"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "9d1b414e",
"metadata": {},
"outputs": [],
"source": [
"group_large['overall_score'].mean()"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "fa71f9a5",
"metadata": {},
"outputs": [],
"source": [
"group_small['overall_score'].mean()"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "79ecb4a2",
"metadata": {},
"outputs": [],
"source": [
"training_subjects = group_large['subjectID'].values\n",
"test_subjects = group_small['subjectID'].values\n",
"print(training_subjects)\n",
"print(test_subjects)"
]
},
{
"cell_type": "markdown",
"id": "4353f87c",
"metadata": {},
"source": [
"### Data cleaning with mad"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "76610052",
"metadata": {},
"outputs": [],
"source": [
"# SET\n",
"threshold_mad = 5\n",
"column_praefix ='AU'\n",
"\n",
"au_columns = [col for col in df.columns if col.startswith(column_praefix)]\n",
"cleaned_df = mad_outlier_removal.mad_outlier_removal(df,columns=au_columns, threshold=threshold_mad)\n",
"print(cleaned_df.shape)\n",
"print(df.shape)"
]
},
{
"cell_type": "markdown",
"id": "9a6c1732",
"metadata": {},
"source": [
"#### TO DO\n",
" * pipeline aus Autoencoder und SVM\n",
" * group k fold\n",
" * AE überpüfen, loss dokumentieren"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "877309d9",
"metadata": {},
"outputs": [],
"source": [
"### Variational Autoencoder with Classifier Head\n",
"import pandas as pd\n",
"import numpy as np\n",
"import tensorflow as tf\n",
"from tensorflow import keras\n",
"from tensorflow.keras import layers, Model\n",
"from sklearn.model_selection import GroupKFold\n",
"from sklearn.preprocessing import StandardScaler\n",
"from sklearn.metrics import (\n",
" accuracy_score, precision_score, recall_score, f1_score, \n",
" roc_auc_score, confusion_matrix, classification_report\n",
")\n",
"import matplotlib.pyplot as plt\n",
"from collections import defaultdict\n",
"\n",
"# ============================================================================\n",
"# 1. CREATE LABELS\n",
"# ============================================================================\n",
"\n",
"# Low workload: baseline + n-back level 1,4\n",
"low_all = cleaned_df[\n",
" ((cleaned_df[\"PHASE\"] == \"baseline\") |\n",
" ((cleaned_df[\"STUDY\"] == \"n-back\") & (cleaned_df[\"PHASE\"] != \"baseline\") & (cleaned_df[\"LEVEL\"].isin([1,4]))))\n",
"].copy()\n",
"low_all['label'] = 0\n",
"print(f\"Low workload samples: {low_all.shape[0]}\")\n",
"\n",
"# High workload n-back: level 2,3,5,6\n",
"high_nback = cleaned_df[\n",
" (cleaned_df[\"STUDY\"]==\"n-back\") &\n",
" (cleaned_df[\"LEVEL\"].isin([2, 3, 5, 6])) &\n",
" (cleaned_df[\"PHASE\"].isin([\"train\", \"test\"]))\n",
"].copy()\n",
"high_nback['label'] = 1\n",
"print(f\"High n-back samples: {high_nback.shape[0]}\")\n",
"\n",
"# High workload k-drive\n",
"high_kdrive = cleaned_df[\n",
" (cleaned_df[\"STUDY\"] == \"k-drive\") & (cleaned_df[\"PHASE\"] != \"baseline\")\n",
"].copy()\n",
"high_kdrive['label'] = 1\n",
"print(f\"High k-drive samples: {high_kdrive.shape[0]}\")\n",
"\n",
"# Combine all high workload\n",
"high_all = pd.concat([high_nback, high_kdrive])\n",
"print(f\"Total high workload samples: {high_all.shape[0]}\")\n",
"\n",
"# Complete labeled dataset\n",
"labeled_df = pd.concat([low_all, high_all]).reset_index(drop=True)\n",
"print(f\"\\nTotal labeled samples: {labeled_df.shape[0]}\")\n",
"print(f\"Class distribution:\\n{labeled_df['label'].value_counts()}\")\n",
"\n",
"# ============================================================================\n",
"# 2. TRAIN/TEST SPLIT BY SUBJECTS\n",
"# ============================================================================\n",
"\n",
"train_df = labeled_df[labeled_df['subjectID'].isin(training_subjects)].copy()\n",
"test_df = labeled_df[labeled_df['subjectID'].isin(test_subjects)].copy()\n",
"\n",
"print(f\"\\nTraining subjects: {training_subjects}\")\n",
"print(f\"Test subjects: {test_subjects}\")\n",
"print(f\"Train samples: {train_df.shape[0]}, Test samples: {test_df.shape[0]}\")\n",
"\n",
"# Extract features and labels\n",
"au_columns = [col for col in labeled_df.columns if col.startswith('AU')]\n",
"print(f\"\\nUsing {len(au_columns)} AU features: {au_columns}\")\n",
"\n",
"X_train = train_df[au_columns].values\n",
"y_train = train_df['label'].values\n",
"groups_train = train_df['subjectID'].values\n",
"\n",
"X_test = test_df[au_columns].values\n",
"y_test = test_df['label'].values\n",
"\n",
"# Normalize features\n",
"scaler = StandardScaler()\n",
"X_train_scaled = scaler.fit_transform(X_train)\n",
"X_test_scaled = scaler.transform(X_test)\n",
"\n",
"print(f\"\\nTrain class distribution: {np.bincount(y_train)}\")\n",
"print(f\"Test class distribution: {np.bincount(y_test)}\")\n",
"\n",
"# ============================================================================\n",
"# 3. VAE WITH CLASSIFIER HEAD MODEL\n",
"# ============================================================================\n",
"\n",
"class Sampling(layers.Layer):\n",
" \"\"\"Reparameterization trick for VAE\"\"\"\n",
" def call(self, inputs):\n",
" z_mean, z_log_var = inputs\n",
" batch = tf.shape(z_mean)[0]\n",
" dim = tf.shape(z_mean)[1]\n",
" epsilon = tf.random.normal(shape=(batch, dim))\n",
" return z_mean + tf.exp(0.5 * z_log_var) * epsilon\n",
"\n",
"def build_vae_classifier(input_dim, latent_dim, encoder_dims=[32, 16], \n",
" decoder_dims=[16, 32], classifier_dims=[16]):\n",
" \"\"\"\n",
" Build VAE with classifier head\n",
" \n",
" Args:\n",
" input_dim: Number of input features (20 AUs)\n",
" latent_dim: Dimension of latent space (2-5)\n",
" encoder_dims: Hidden layer sizes for encoder\n",
" decoder_dims: Hidden layer sizes for decoder\n",
" classifier_dims: Hidden layer sizes for classifier\n",
" \"\"\"\n",
" \n",
" # ---- ENCODER ----\n",
" encoder_inputs = keras.Input(shape=(input_dim,), name='encoder_input')\n",
" x = encoder_inputs\n",
" \n",
" for i, dim in enumerate(encoder_dims):\n",
" x = layers.Dense(dim, activation='relu', name=f'encoder_dense_{i}')(x)\n",
" x = layers.BatchNormalization(name=f'encoder_bn_{i}')(x)\n",
" x = layers.Dropout(0.2, name=f'encoder_dropout_{i}')(x)\n",
" \n",
" z_mean = layers.Dense(latent_dim, name='z_mean')(x)\n",
" z_log_var = layers.Dense(latent_dim, name='z_log_var')(x)\n",
" z = Sampling()([z_mean, z_log_var])\n",
" \n",
" encoder = Model(encoder_inputs, [z_mean, z_log_var, z], name='encoder')\n",
" \n",
" # ---- DECODER ----\n",
" latent_inputs = keras.Input(shape=(latent_dim,), name='latent_input')\n",
" x = latent_inputs\n",
" \n",
" for i, dim in enumerate(decoder_dims):\n",
" x = layers.Dense(dim, activation='relu', name=f'decoder_dense_{i}')(x)\n",
" x = layers.BatchNormalization(name=f'decoder_bn_{i}')(x)\n",
" \n",
" decoder_outputs = layers.Dense(input_dim, activation='linear', name='decoder_output')(x)\n",
" decoder = Model(latent_inputs, decoder_outputs, name='decoder')\n",
" \n",
" # ---- CLASSIFIER HEAD ----\n",
" x = latent_inputs\n",
" for i, dim in enumerate(classifier_dims):\n",
" x = layers.Dense(dim, activation='relu', name=f'classifier_dense_{i}')(x)\n",
" x = layers.Dropout(0.3, name=f'classifier_dropout_{i}')(x)\n",
" \n",
" classifier_output = layers.Dense(1, activation='sigmoid', name='classifier_output')(x)\n",
" classifier = Model(latent_inputs, classifier_output, name='classifier')\n",
" \n",
" # ---- FULL MODEL ----\n",
" inputs = keras.Input(shape=(input_dim,), name='vae_input')\n",
" z_mean, z_log_var, z = encoder(inputs)\n",
" reconstructed = decoder(z)\n",
" classification = classifier(z)\n",
" \n",
" model = Model(inputs, [reconstructed, classification], name='vae_classifier')\n",
" \n",
" return model, encoder, decoder, classifier\n",
"\n",
"# ============================================================================\n",
"# 4. CUSTOM TRAINING LOOP WITH COMBINED LOSS\n",
"# ============================================================================\n",
"\n",
"class VAEClassifier(keras.Model):\n",
" def __init__(self, encoder, decoder, classifier, **kwargs):\n",
" super().__init__(**kwargs)\n",
" self.encoder = encoder\n",
" self.decoder = decoder\n",
" self.classifier = classifier\n",
" self.total_loss_tracker = keras.metrics.Mean(name=\"total_loss\")\n",
" self.reconstruction_loss_tracker = keras.metrics.Mean(name=\"reconstruction_loss\")\n",
" self.kl_loss_tracker = keras.metrics.Mean(name=\"kl_loss\")\n",
" self.classification_loss_tracker = keras.metrics.Mean(name=\"classification_loss\")\n",
" self.accuracy_tracker = keras.metrics.BinaryAccuracy(name=\"accuracy\")\n",
" \n",
" @property\n",
" def metrics(self):\n",
" return [\n",
" self.total_loss_tracker,\n",
" self.reconstruction_loss_tracker,\n",
" self.kl_loss_tracker,\n",
" self.classification_loss_tracker,\n",
" self.accuracy_tracker,\n",
" ]\n",
" \n",
" def train_step(self, data):\n",
" x, y = data\n",
" \n",
" with tf.GradientTape() as tape:\n",
" # Forward pass\n",
" z_mean, z_log_var, z = self.encoder(x, training=True)\n",
" reconstruction = self.decoder(z, training=True)\n",
" classification = self.classifier(z, training=True)\n",
" \n",
" # Reconstruction loss (MSE)\n",
" reconstruction_loss = tf.reduce_mean(\n",
" keras.losses.mse(x, reconstruction))\n",
" \n",
" # KL divergence loss\n",
" kl_loss = -0.5 * tf.reduce_mean(\n",
" tf.reduce_sum(\n",
" 1 + z_log_var - tf.square(z_mean) - tf.exp(z_log_var),\n",
" axis=1\n",
" )\n",
" )\n",
" \n",
" # Classification loss (binary crossentropy)\n",
" # Classification loss (binary crossentropy)\n",
" classification_loss = tf.reduce_mean(\n",
" keras.losses.binary_crossentropy(tf.expand_dims(y, -1), classification)\n",
" )\n",
" \n",
" # Combined loss with weights\n",
" total_loss = reconstruction_loss + kl_loss + classification_loss\n",
" \n",
" # Backpropagation\n",
" grads = tape.gradient(total_loss, self.trainable_weights)\n",
" self.optimizer.apply_gradients(zip(grads, self.trainable_weights))\n",
" \n",
" # Update metrics\n",
" self.total_loss_tracker.update_state(total_loss)\n",
" self.reconstruction_loss_tracker.update_state(reconstruction_loss)\n",
" self.kl_loss_tracker.update_state(kl_loss)\n",
" self.classification_loss_tracker.update_state(classification_loss)\n",
" self.accuracy_tracker.update_state(y, classification)\n",
" \n",
" return {\n",
" \"total_loss\": self.total_loss_tracker.result(),\n",
" \"reconstruction_loss\": self.reconstruction_loss_tracker.result(),\n",
" \"kl_loss\": self.kl_loss_tracker.result(),\n",
" \"classification_loss\": self.classification_loss_tracker.result(),\n",
" \"accuracy\": self.accuracy_tracker.result(),\n",
" }\n",
" \n",
" def test_step(self, data):\n",
" x, y = data\n",
" \n",
" z_mean, z_log_var, z = self.encoder(x, training=False)\n",
" reconstruction = self.decoder(z, training=False)\n",
" classification = self.classifier(z, training=False)\n",
" \n",
" # Reconstruction loss (MSE)\n",
" reconstruction_loss = tf.reduce_mean(\n",
" keras.losses.mse(x, reconstruction))\n",
" kl_loss = -0.5 * tf.reduce_mean(\n",
" tf.reduce_sum(1 + z_log_var - tf.square(z_mean) - tf.exp(z_log_var), axis=1)\n",
" )\n",
" # Classification loss (binary crossentropy)\n",
" classification_loss = tf.reduce_mean(\n",
" keras.losses.binary_crossentropy(tf.expand_dims(y, -1), classification)\n",
" )\n",
" total_loss = reconstruction_loss + kl_loss + classification_loss\n",
" \n",
" self.total_loss_tracker.update_state(total_loss)\n",
" self.reconstruction_loss_tracker.update_state(reconstruction_loss)\n",
" self.kl_loss_tracker.update_state(kl_loss)\n",
" self.classification_loss_tracker.update_state(classification_loss)\n",
" self.accuracy_tracker.update_state(y, classification)\n",
" \n",
" return {\n",
" \"total_loss\": self.total_loss_tracker.result(),\n",
" \"reconstruction_loss\": self.reconstruction_loss_tracker.result(),\n",
" \"kl_loss\": self.kl_loss_tracker.result(),\n",
" \"classification_loss\": self.classification_loss_tracker.result(),\n",
" \"accuracy\": self.accuracy_tracker.result(),\n",
" }\n",
"\n",
"# ============================================================================\n",
"# 5. GROUP K-FOLD CROSS-VALIDATION WITH GRID SEARCH\n",
"# ============================================================================\n",
"\n",
"# Hyperparameter grid\n",
"param_grid = {\n",
" 'latent_dim': [2, 5],\n",
" 'encoder_dims': [[32, 16], [64, 32]],\n",
" 'learning_rate': [0.001, 0.005],\n",
" 'batch_size': [32, 64],\n",
"}\n",
"\n",
"# Generate all combinations\n",
"from itertools import product\n",
"keys = param_grid.keys()\n",
"values = param_grid.values()\n",
"param_combinations = [dict(zip(keys, v)) for v in product(*values)]\n",
"\n",
"print(f\"\\nTotal hyperparameter combinations: {len(param_combinations)}\")\n",
"\n",
"# Group K-Fold setup\n",
"n_splits = 5\n",
"gkf = GroupKFold(n_splits=n_splits)\n",
"\n",
"# Store results\n",
"cv_results = []\n",
"\n",
"# Grid search with cross-validation\n",
"for idx, params in enumerate(param_combinations):\n",
" print(f\"\\n{'='*80}\")\n",
" print(f\"Testing combination {idx+1}/{len(param_combinations)}: {params}\")\n",
" print(f\"{'='*80}\")\n",
" \n",
" fold_results = []\n",
" \n",
" for fold, (train_idx, val_idx) in enumerate(gkf.split(X_train_scaled, y_train, groups_train)):\n",
" print(f\"\\nFold {fold+1}/{n_splits}\")\n",
" \n",
" X_fold_train, X_fold_val = X_train_scaled[train_idx], X_train_scaled[val_idx]\n",
" y_fold_train, y_fold_val = y_train[train_idx], y_train[val_idx]\n",
" \n",
" # Build model\n",
" model, encoder, decoder, classifier = build_vae_classifier(\n",
" input_dim=len(au_columns),\n",
" latent_dim=params['latent_dim'],\n",
" encoder_dims=params['encoder_dims'],\n",
" decoder_dims=list(reversed(params['encoder_dims'])),\n",
" classifier_dims=[16]\n",
" )\n",
" \n",
" vae_classifier = VAEClassifier(encoder, decoder, classifier)\n",
" vae_classifier.compile(optimizer=keras.optimizers.Adam(params['learning_rate']))\n",
" \n",
" # Early stopping\n",
" early_stop = keras.callbacks.EarlyStopping(\n",
" monitor='val_total_loss',\n",
" patience=10,\n",
" restore_best_weights=True,\n",
" mode='min'\n",
" )\n",
" \n",
" # Train\n",
" history = vae_classifier.fit(\n",
" X_fold_train, y_fold_train,\n",
" validation_data=(X_fold_val, y_fold_val),\n",
" epochs=60,\n",
" batch_size=params['batch_size'],\n",
" callbacks=[early_stop],\n",
" verbose=0\n",
" )\n",
" \n",
" # Evaluate on validation fold\n",
" z_mean_val, _, _ = encoder.predict(X_fold_val, verbose=0)\n",
" y_pred_proba = classifier.predict(z_mean_val, verbose=0).flatten()\n",
" y_pred = (y_pred_proba > 0.5).astype(int)\n",
" \n",
" fold_metrics = {\n",
" 'accuracy': accuracy_score(y_fold_val, y_pred),\n",
" 'precision': precision_score(y_fold_val, y_pred, zero_division=0),\n",
" 'recall': recall_score(y_fold_val, y_pred, zero_division=0),\n",
" 'f1': f1_score(y_fold_val, y_pred, zero_division=0),\n",
" 'roc_auc': roc_auc_score(y_fold_val, y_pred_proba),\n",
" 'final_recon_loss': history.history['val_reconstruction_loss'][-1],\n",
" 'final_kl_loss': history.history['val_kl_loss'][-1],\n",
" 'final_class_loss': history.history['val_classification_loss'][-1],\n",
" }\n",
" \n",
" fold_results.append(fold_metrics)\n",
" print(f\" Accuracy: {fold_metrics['accuracy']:.4f}, F1: {fold_metrics['f1']:.4f}, AUC: {fold_metrics['roc_auc']:.4f}\")\n",
" \n",
" # Clear session to free memory\n",
" keras.backend.clear_session()\n",
" \n",
" # Average across folds\n",
" avg_results = {\n",
" 'params': params,\n",
" 'mean_accuracy': np.mean([r['accuracy'] for r in fold_results]),\n",
" 'std_accuracy': np.std([r['accuracy'] for r in fold_results]),\n",
" 'mean_f1': np.mean([r['f1'] for r in fold_results]),\n",
" 'std_f1': np.std([r['f1'] for r in fold_results]),\n",
" 'mean_roc_auc': np.mean([r['roc_auc'] for r in fold_results]),\n",
" 'std_roc_auc': np.std([r['roc_auc'] for r in fold_results]),\n",
" 'mean_recon_loss': np.mean([r['final_recon_loss'] for r in fold_results]),\n",
" 'mean_kl_loss': np.mean([r['final_kl_loss'] for r in fold_results]),\n",
" 'mean_class_loss': np.mean([r['final_class_loss'] for r in fold_results]),\n",
" 'fold_results': fold_results\n",
" }\n",
" \n",
" cv_results.append(avg_results)\n",
" \n",
" print(f\"\\nMean CV Accuracy: {avg_results['mean_accuracy']:.4f} ± {avg_results['std_accuracy']:.4f}\")\n",
" print(f\"Mean CV F1: {avg_results['mean_f1']:.4f} ± {avg_results['std_f1']:.4f}\")\n",
" print(f\"Mean CV AUC: {avg_results['mean_roc_auc']:.4f} ± {avg_results['std_roc_auc']:.4f}\")\n",
"\n",
"# ============================================================================\n",
"# 6. SELECT BEST MODEL AND EVALUATE ON TEST SET\n",
"# ============================================================================\n",
"\n",
"# Find best hyperparameters based on mean F1 score\n",
"best_idx = np.argmax([r['mean_f1'] for r in cv_results])\n",
"best_params = cv_results[best_idx]['params']\n",
"\n",
"print(f\"\\n{'='*80}\")\n",
"print(\"BEST HYPERPARAMETERS (based on CV F1 score):\")\n",
"print(f\"{'='*80}\")\n",
"for key, value in best_params.items():\n",
" print(f\"{key}: {value}\")\n",
"print(f\"\\nCV Performance:\")\n",
"print(f\" Accuracy: {cv_results[best_idx]['mean_accuracy']:.4f} ± {cv_results[best_idx]['std_accuracy']:.4f}\")\n",
"print(f\" F1 Score: {cv_results[best_idx]['mean_f1']:.4f} ± {cv_results[best_idx]['std_f1']:.4f}\")\n",
"print(f\" ROC-AUC: {cv_results[best_idx]['mean_roc_auc']:.4f} ± {cv_results[best_idx]['std_roc_auc']:.4f}\")\n",
"\n",
"# Train final model on all training data\n",
"print(f\"\\n{'='*80}\")\n",
"print(\"TRAINING FINAL MODEL ON ALL TRAINING DATA\")\n",
"print(f\"{'='*80}\")\n",
"\n",
"final_model, final_encoder, final_decoder, final_classifier = build_vae_classifier(\n",
" input_dim=len(au_columns),\n",
" latent_dim=best_params['latent_dim'],\n",
" encoder_dims=best_params['encoder_dims'],\n",
" decoder_dims=list(reversed(best_params['encoder_dims'])),\n",
" classifier_dims=[16]\n",
")\n",
"\n",
"final_vae_classifier = VAEClassifier(final_encoder, final_decoder, final_classifier)\n",
"final_vae_classifier.compile(optimizer=keras.optimizers.Adam(best_params['learning_rate']))\n",
"\n",
"final_history = final_vae_classifier.fit(\n",
" X_train_scaled, y_train,\n",
" validation_split=0.2,\n",
" epochs=100,\n",
" batch_size=best_params['batch_size'],\n",
" callbacks=[keras.callbacks.EarlyStopping(monitor='val_total_loss', patience=15, restore_best_weights=True, mode='min')],\n",
" verbose=1\n",
")\n",
"\n",
"# Evaluate on held-out test set\n",
"print(f\"\\n{'='*80}\")\n",
"print(\"EVALUATION ON HELD-OUT TEST SET\")\n",
"print(f\"{'='*80}\")\n",
"\n",
"z_mean_test, _, _ = final_encoder.predict(X_test_scaled, verbose=0)\n",
"y_test_pred_proba = final_classifier.predict(z_mean_test, verbose=0).flatten()\n",
"y_test_pred = (y_test_pred_proba > 0.5).astype(int)\n",
"\n",
"test_metrics = {\n",
" 'accuracy': accuracy_score(y_test, y_test_pred),\n",
" 'precision': precision_score(y_test, y_test_pred),\n",
" 'recall': recall_score(y_test, y_test_pred),\n",
" 'f1': f1_score(y_test, y_test_pred),\n",
" 'roc_auc': roc_auc_score(y_test, y_test_pred_proba),\n",
"}\n",
"\n",
"print(\"\\nTest Set Performance:\")\n",
"for metric, value in test_metrics.items():\n",
" print(f\" {metric.capitalize()}: {value:.4f}\")\n",
"\n",
"print(\"\\nConfusion Matrix:\")\n",
"print(confusion_matrix(y_test, y_test_pred))\n",
"\n",
"print(\"\\nClassification Report:\")\n",
"print(classification_report(y_test, y_test_pred, target_names=['Low Workload', 'High Workload']))\n",
"\n",
"# ============================================================================\n",
"# 7. VISUALIZATION\n",
"# ============================================================================\n",
"\n",
"# Plot training history\n",
"fig, axes = plt.subplots(2, 2, figsize=(15, 10))\n",
"\n",
"axes[0, 0].plot(final_history.history['reconstruction_loss'], label='Train')\n",
"axes[0, 0].plot(final_history.history['val_reconstruction_loss'], label='Val')\n",
"axes[0, 0].set_title('Reconstruction Loss')\n",
"axes[0, 0].set_xlabel('Epoch')\n",
"axes[0, 0].set_ylabel('Loss')\n",
"axes[0, 0].legend()\n",
"axes[0, 0].grid(True)\n",
"\n",
"axes[0, 1].plot(final_history.history['kl_loss'], label='Train')\n",
"axes[0, 1].plot(final_history.history['val_kl_loss'], label='Val')\n",
"axes[0, 1].set_title('KL Divergence Loss')\n",
"axes[0, 1].set_xlabel('Epoch')\n",
"axes[0, 1].set_ylabel('Loss')\n",
"axes[0, 1].legend()\n",
"axes[0, 1].grid(True)\n",
"\n",
"axes[1, 0].plot(final_history.history['classification_loss'], label='Train')\n",
"axes[1, 0].plot(final_history.history['val_classification_loss'], label='Val')\n",
"axes[1, 0].set_title('Classification Loss')\n",
"axes[1, 0].set_xlabel('Epoch')\n",
"axes[1, 0].set_ylabel('Loss')\n",
"axes[1, 0].legend()\n",
"axes[1, 0].grid(True)\n",
"\n",
"axes[1, 1].plot(final_history.history['accuracy'], label='Train')\n",
"axes[1, 1].plot(final_history.history['val_accuracy'], label='Val')\n",
"axes[1, 1].set_title('Classification Accuracy')\n",
"axes[1, 1].set_xlabel('Epoch')\n",
"axes[1, 1].set_ylabel('Accuracy')\n",
"axes[1, 1].legend()\n",
"axes[1, 1].grid(True)\n",
"\n",
"plt.tight_layout()\n",
"plt.show()\n",
"\n",
"# Visualize latent space (if 2D or 3D)\n",
"if best_params['latent_dim'] == 2:\n",
" z_mean_train, _, _ = final_encoder.predict(X_train_scaled, verbose=0)\n",
" \n",
" plt.figure(figsize=(10, 8))\n",
" scatter = plt.scatter(z_mean_train[:, 0], z_mean_train[:, 1], \n",
" c=y_train, cmap='RdYlBu', alpha=0.6, edgecolors='k')\n",
" plt.colorbar(scatter, label='Workload (0=Low, 1=High)')\n",
" plt.xlabel('Latent Dimension 1')\n",
" plt.ylabel('Latent Dimension 2')\n",
" plt.title('2D Latent Space Representation (Training Data)')\n",
" plt.grid(True, alpha=0.3)\n",
" plt.show()\n",
" \n",
" # Test set latent space\n",
" plt.figure(figsize=(10, 8))\n",
" scatter = plt.scatter(z_mean_test[:, 0], z_mean_test[:, 1], \n",
" c=y_test, cmap='RdYlBu', alpha=0.6, edgecolors='k')\n",
" plt.colorbar(scatter, label='Workload (0=Low, 1=High)')\n",
" plt.xlabel('Latent Dimension 1')\n",
" plt.ylabel('Latent Dimension 2')\n",
" plt.title('2D Latent Space Representation (Test Data)')\n",
" plt.grid(True, alpha=0.3)\n",
" plt.show()\n",
"\n",
"print(\"\\n\" + \"=\"*80)\n",
"print(\"TRAINING COMPLETE!\")\n",
"print(\"=\"*80)"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "79bcfc58",
"metadata": {},
"outputs": [],
"source": [
"### Save Trained VAE Classifier Model\n",
"from pathlib import Path\n",
"from datetime import datetime\n",
"\n",
"# Define save path\n",
"model_dir = Path(\"/home/jovyan/data-paulusjafahrsimulator-gpu/trained_models\")\n",
"model_dir.mkdir(parents=True, exist_ok=True)\n",
"\n",
"timestamp = datetime.now().strftime(\"%Y%m%d_%H%M%S\")\n",
"model_path = model_dir / f\"vae_classifier_{timestamp}.keras\"\n",
"\n",
"# Save the complete model\n",
"final_vae_classifier.save(model_path)\n",
"\n",
"print(f\"Model saved to: {model_path}\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"id": "d700e517",
"metadata": {},
"outputs": [],
"source": []
},
{
"cell_type": "code",
"execution_count": null,
"id": "30d8d100",
"metadata": {},
"outputs": [],
"source": [
"### Plot Confusion Matrix for Final Model\n",
"from sklearn.metrics import ConfusionMatrixDisplay\n",
"x = Path(\"/home/jovyan/data-paulusjafahrsimulator-gpu/trained_models/vae_classifier_20251210_230121.keras\")\n",
"# Load the saved model\n",
"print(f\"Loading model from: {x}\")\n",
"# loaded_vae_classifier = tf.keras.models.load_model(x)\n",
"loaded_vae_classifier = final_vae_classifier\n",
"print(\"✓ Model loaded successfully!\")\n",
"\n",
"# Extract encoder and classifier from loaded model\n",
"loaded_encoder = loaded_vae_classifier.encoder\n",
"loaded_classifier = loaded_vae_classifier.classifier\n",
"\n",
"# Get predictions on test set\n",
"z_mean_test, _, _ = loaded_encoder.predict(X_test_scaled, verbose=0)\n",
"y_test_pred_proba = loaded_classifier.predict(z_mean_test, verbose=0).flatten()\n",
"y_test_pred = (y_test_pred_proba > 0.5).astype(int)\n",
"\n",
"# Create and plot confusion matrix\n",
"cm = confusion_matrix(y_test, y_test_pred)\n",
"disp = ConfusionMatrixDisplay(confusion_matrix=cm, \n",
" display_labels=['Low Workload', 'High Workload'])\n",
"\n",
"fig, ax = plt.subplots(figsize=(8, 6))\n",
"disp.plot(ax=ax, cmap='Blues', values_format='d')\n",
"plt.title('Confusion Matrix - Test Set (Loaded Model)')\n",
"plt.tight_layout()\n",
"plt.show()\n",
"\n",
"# Print metrics\n",
"print(f\"\\nTest Set Performance (Loaded Model):\")\n",
"print(f\" Accuracy: {accuracy_score(y_test, y_test_pred):.4f}\")\n",
"print(f\" Precision: {precision_score(y_test, y_test_pred):.4f}\")\n",
"print(f\" Recall: {recall_score(y_test, y_test_pred):.4f}\")\n",
"print(f\" F1 Score: {f1_score(y_test, y_test_pred):.4f}\")\n",
"print(f\" ROC-AUC: {roc_auc_score(y_test, y_test_pred_proba):.4f}\")"
]
},
{
"cell_type": "markdown",
"id": "e826a998",
"metadata": {},
"source": [
"TO DO\n",
" * autoencoder langsam anfangen mit 19 schichten\n",
" * dann AE und SVM mit hybridem training wie bei claude?!\n",
" * dataset aus eyetracking verwenden?"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.10"
}
},
"nbformat": 4,
"nbformat_minor": 5
}

11
predict_pipeline/check_python_version.py Normal file
View File

@ -0,0 +1,11 @@
# from tools import db_helpers
import sys
def main():
print(sys.version)
# db_helpers.add_columns_to_table()
if __name__ == "__main__":
main()

9
predict_pipeline/feature_extraction.py Normal file
View File

@ -0,0 +1,9 @@
import sqlite3
def main():
return 0
if __name__ == "__main__":
main()

58
predict_pipeline/fill_db.ipynb
View File

@ -1,13 +1,5 @@
{ {
"cells": [ "cells": [
{
"cell_type": "markdown",
"id": "fb68b447",
"metadata": {},
"source": [
"## Database creation and filling (for live system) "
]
},
{ {
"cell_type": "code", "cell_type": "code",
"execution_count": null, "execution_count": null,
@ -29,10 +21,8 @@
"metadata": {}, "metadata": {},
"outputs": [], "outputs": [],
"source": [ "source": [
"# TODO: set paths and table name\n", "database_path = Path(r\"/home/edgekit/MSY_FS/databases/rawdata.sqlite\")\n",
"database_path = Path(r\"database.sqlite\") # this path references an empty, but already created sqlite file\n", "parquet_path = Path(r\"/home/edgekit/MSY_FS/fahrsimulator_msy2526_ai/files_for_testing/both_mod_0000.parquet\")"
"parquet_path = Path(r\"...parquet\") # this path leads to the data that should be used to fill the databse\n",
"table_name = \"XXX\" # name of the new table"
] ]
}, },
{ {
@ -66,14 +56,6 @@
"con, cursor = db_helpers.connect_db(database_path)" "con, cursor = db_helpers.connect_db(database_path)"
] ]
}, },
{
"cell_type": "markdown",
"id": "7007c68f",
"metadata": {},
"source": [
"Select a subset to insert into database "
]
},
{ {
"cell_type": "code", "cell_type": "code",
"execution_count": null, "execution_count": null,
@ -87,14 +69,6 @@
"df_first_100.insert(0, '_Id', df_first_100.index + 1)" "df_first_100.insert(0, '_Id', df_first_100.index + 1)"
] ]
}, },
{
"cell_type": "markdown",
"id": "92171186",
"metadata": {},
"source": [
"Type conversion"
]
},
{ {
"cell_type": "code", "cell_type": "code",
"execution_count": null, "execution_count": null,
@ -114,14 +88,6 @@
" return \"TEXT\"\n" " return \"TEXT\"\n"
] ]
}, },
{
"cell_type": "markdown",
"id": "45af9956",
"metadata": {},
"source": [
"Define constraints and primary key"
]
},
{ {
"cell_type": "code", "cell_type": "code",
"execution_count": null, "execution_count": null,
@ -143,14 +109,6 @@
"}\n" "}\n"
] ]
}, },
{
"cell_type": "markdown",
"id": "133e92ee",
"metadata": {},
"source": [
"Create the table"
]
},
{ {
"cell_type": "code", "cell_type": "code",
"execution_count": null, "execution_count": null,
@ -161,7 +119,7 @@
"sql = db_helpers.create_table(\n", "sql = db_helpers.create_table(\n",
" conn=con,\n", " conn=con,\n",
" cursor=cursor,\n", " cursor=cursor,\n",
" table_name=table_name,\n", " table_name=\"rawdata\",\n",
" columns=columns,\n", " columns=columns,\n",
" constraints=constraints,\n", " constraints=constraints,\n",
" primary_key=primary_key,\n", " primary_key=primary_key,\n",
@ -192,7 +150,7 @@
"db_helpers.insert_rows_into_table(\n", "db_helpers.insert_rows_into_table(\n",
" conn=con,\n", " conn=con,\n",
" cursor=cursor,\n", " cursor=cursor,\n",
" table_name=table_name,\n", " table_name=\"rawdata\",\n",
" columns=columns_to_insert,\n", " columns=columns_to_insert,\n",
" commit=True\n", " commit=True\n",
")\n" ")\n"
@ -205,7 +163,7 @@
"metadata": {}, "metadata": {},
"outputs": [], "outputs": [],
"source": [ "source": [
"request = db_helpers.get_data_from_table(conn=con, table_name='rawdata',columns_list=['*'])" "a = db_helpers.get_data_from_table(conn=con, table_name='rawdata',columns_list=['*'])"
] ]
}, },
{ {
@ -215,7 +173,7 @@
"metadata": {}, "metadata": {},
"outputs": [], "outputs": [],
"source": [ "source": [
"request.head()" "a.head()"
] ]
}, },
{ {
@ -231,7 +189,7 @@
], ],
"metadata": { "metadata": {
"kernelspec": { "kernelspec": {
"display_name": "310", "display_name": "MSY_FS_env",
"language": "python", "language": "python",
"name": "python3" "name": "python3"
}, },
@ -245,7 +203,7 @@
"name": "python", "name": "python",
"nbconvert_exporter": "python", "nbconvert_exporter": "python",
"pygments_lexer": "ipython3", "pygments_lexer": "ipython3",
"version": "3.10.19" "version": "3.12.12"
} }
}, },
"nbformat": 4, "nbformat": 4,