Some changes in hyperparameter config

1. Separate hyperparameter configs in model, optimizer and scheduler
2. Add more tunable hyperparameters in optimizer and scheduler

1 files changed, 42 insertions, 29 deletions

diff --git a/config.py b/config.py
index 47ded38..8a8d93a 100644
--- a/config.py
+++ b/config.py
@@ -41,35 +41,48 @@ config: Configuration = {
     },
     # Hyperparameter tuning
     'hyperparameter': {
-        # Auto-encoder feature channels coefficient
-        'ae_feature_channels': 64,
-        # Appearance, canonical and pose feature dimensions
-        'f_a_c_p_dims': (128, 128, 64),
-        # HPM pyramid scales, of which sum is number of parts
-        'hpm_scales': (1, 2, 4),
-        # Global pooling method
-        'hpm_use_avg_pool': True,
-        'hpm_use_max_pool': True,
-        # FConv feature channels coefficient
-        'fpfe_feature_channels': 32,
-        # FConv blocks kernel sizes
-        'fpfe_kernel_sizes': ((5, 3), (3, 3), (3, 3)),
-        # FConv blocks paddings
-        'fpfe_paddings': ((2, 1), (1, 1), (1, 1)),
-        # FConv blocks halving
-        'fpfe_halving': (0, 2, 3),
-        # Attention squeeze ratio
-        'tfa_squeeze_ratio': 4,
-        # Number of parts after Part Net
-        'tfa_num_parts': 16,
-        # Embedding dimension for each part
-        'embedding_dims': 256,
-        # Triplet loss margin
-        'triplet_margin': 0.2,
-        # Initial learning rate of Adam Optimizer
-        'lr': 1e-4,
-        # Betas of Adam Optimizer
-        'betas': (0.9, 0.999),
+        'model': {
+            # Auto-encoder feature channels coefficient
+            'ae_feature_channels': 64,
+            # Appearance, canonical and pose feature dimensions
+            'f_a_c_p_dims': (128, 128, 64),
+            # HPM pyramid scales, of which sum is number of parts
+            'hpm_scales': (1, 2, 4),
+            # Global pooling method
+            'hpm_use_avg_pool': True,
+            'hpm_use_max_pool': True,
+            # FConv feature channels coefficient
+            'fpfe_feature_channels': 32,
+            # FConv blocks kernel sizes
+            'fpfe_kernel_sizes': ((5, 3), (3, 3), (3, 3)),
+            # FConv blocks paddings
+            'fpfe_paddings': ((2, 1), (1, 1), (1, 1)),
+            # FConv blocks halving
+            'fpfe_halving': (0, 2, 3),
+            # Attention squeeze ratio
+            'tfa_squeeze_ratio': 4,
+            # Number of parts after Part Net
+            'tfa_num_parts': 16,
+            # Embedding dimension for each part
+            'embedding_dims': 256,
+            # Triplet loss margin
+            'triplet_margin': 0.2,
+        },
+        'optimizer': {
+            # Initial learning rate of Adam Optimizer
+            'lr': 1e-4,
+            # Coefficients used for computing running averages of
+            #   gradient and its square
+            'betas': (0.9, 0.999),
+            # Weight decay (L2 penalty)
+            'weight_decay': 0.001,
+        },
+        'scheduler': {
+            # Period of learning rate decay
+            'step_size': 500,
+            # Multiplicative factor of decay
+            'gamma': 0.9,
+        }
     },
     # Model metadata
     'model': {