From 3e194dc9a40559efc0d85d55e9b7bf6e81c3bfbc Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Borb=C3=A1la=20Farkas?= <36622811+lujzi05@users.noreply.github.com> Date: Mon, 19 Jan 2026 16:09:47 +0100 Subject: [PATCH] changing hyperparameters for dancing queen --- .../configs/config_hyperparameters.py | 12 ++++++------ models/dancing_queen/configs/config_sweep.py | 16 ++++++++++------ models/dancing_queen/run.sh | 0 3 files changed, 16 insertions(+), 12 deletions(-) mode change 100644 => 100755 models/dancing_queen/run.sh diff --git a/models/dancing_queen/configs/config_hyperparameters.py b/models/dancing_queen/configs/config_hyperparameters.py index f31c5f0b..5cba5ac4 100644 --- a/models/dancing_queen/configs/config_hyperparameters.py +++ b/models/dancing_queen/configs/config_hyperparameters.py @@ -13,15 +13,15 @@ def get_hp_config(): "mc_dropout": True, "activation": "ReLU", "batch_size": 128, - "delta": 0.1031278505804136, + "delta": 0.3, #0.1031278505804136, "dropout": 0.3, - "early_stopping_min_delta": 0.01, - "early_stopping_patience": 10, - "false_negative_weight": 2.228018327985685, - "false_positive_weight": 1.5395763736325097, + "early_stopping_min_delta": 0.05, #0.01, + "early_stopping_patience": 12, #10, + "false_negative_weight": 3.5, #2.228018327985685, + "false_positive_weight": 2.2, #1.5395763736325097, "feature_scaler": "MinMaxScaler", "force_reset": True, - "gradient_clip_val": 0.5415811286559888, + "gradient_clip_val": 0.7, #0.5415811286559888, "hidden_dim": 512, "input_chunk_length": 36, "log_features": [ diff --git a/models/dancing_queen/configs/config_sweep.py b/models/dancing_queen/configs/config_sweep.py index e70306f3..9754a38e 100644 --- a/models/dancing_queen/configs/config_sweep.py +++ b/models/dancing_queen/configs/config_sweep.py @@ -32,11 +32,11 @@ def get_sweep_config(): # ============== TEMPORAL CONFIGURATION ============== 'steps': {'values': [[*range(1, 36 + 1)]]}, 'input_chunk_length': {'values': [36, 48, 60]}, # RNNs handle long sequences well - 'output_chunk_shift': {'values': [0]}, + 'ut_chunk_shift': {'values': [0]}, # ============== TRAINING BASICS ============== - 'batch_size': {'values': [64, 128, 256]}, # Larger batches stabilize RNN training - 'n_epochs': {'values': [300]}, + 'batch_size': {'values': [64, 128, 256, 512]}, # Larger batches stabilize RNN training + 'n_epochs': {'values': [200,300,400]}, 'early_stopping_patience': {'values': [10, 12]}, 'early_stopping_min_delta': {'values': [0.001, 0.005]}, 'force_reset': {'values': [True]}, @@ -64,6 +64,10 @@ def get_sweep_config(): 'min': 0.5, 'max': 1.0, }, + 'lr_scheduler_type': { + 'values': ['none', 'ReduceLROnPlateau', 'cosine', 'onecycle'] + }, + # ============== SCALING ============== # RobustScaler as default fallback for unmapped features @@ -134,7 +138,7 @@ def get_sweep_config(): 'hidden_dim': {'values': [256, 512, 768]}, # Larger for rare event patterns 'n_rnn_layers': {'values': [2, 3, 4]}, # Deeper for complex temporal patterns 'activation': {'values': ['ReLU', 'GELU', 'Tanh']}, - 'dropout': {'values': [0.2, 0.3, 0.4]}, # Moderate dropout prevents overfitting on zeros + 'dropout': {'values': [0.2, 0.3, 0.4, 0.5]}, # Moderate dropout prevents overfitting on zeros 'use_reversible_instance_norm': {'values': [True]}, # Helps with non-stationary conflict # ============== LOSS FUNCTION ============== @@ -148,8 +152,8 @@ def get_sweep_config(): }, 'delta': { 'distribution': 'log_uniform_values', - 'min': 0.1, - 'max': 0.8, + 'min': 0.05, #0.1, + 'max': 2.0, #0.8, }, 'non_zero_weight': { 'distribution': 'uniform', diff --git a/models/dancing_queen/run.sh b/models/dancing_queen/run.sh old mode 100644 new mode 100755