Hyperparameter tuning

• update DLNs in DeepHit and PC-Hazard
• add w/o IPS to single events in SurvTRACE
• tune MTL in loss function
• begin outputting final results

experiments/configurations.py

@@ -16,47 +16,72 @@ batch_size_seer = 1024
 
 ### Metabric ###
 CPH_metabric = EasyDict({
-    'epochs': 200
+    'epochs': 100
 })
 
 ### SUPPORT ###
 CPH_support = EasyDict({
-    'epochs': 200
+    'epochs': 100
 })
 ### SEER ###
 CPH_seer = EasyDict({
-    'epochs': 200,
+    'epochs': 100,
 })
 
 
 ########################## DeepHit ##########################
+# alpha: governs the contribution of likelihoo and rank loss
+# sigma: scales loss function
 
 ### Metabric ###
 DeepHit_metabric = EasyDict({
     'batch_size': batch_size_metabric,
-    'learning_rate': 0.01,
     'epochs': 100,
-    'hidden_size': 32,
-    'dropout': 0.1
+
+    # network
+    'hidden_layers_size': [16],
+    'dropout': 0.1,
+
+    # loss
+    'alpha': 0.2,
+    'sigma': 0.1,
+
+    # optimizer
+    'learning_rate': 0.01
 })
 
 ### SUPPORT ###
 DeepHit_support = EasyDict({
     'batch_size': batch_size_support,
-    'learning_rate': 0.01,
     'epochs': 100,
-    'hidden_size': 32,
-    'dropout': 0.1
+
+    # network
+    'hidden_layers_size': [16, 16],
+    'dropout': 0.1,
+
+    # loss
+    'alpha': 0.2,
+    'sigma': 0.1,
+
+    # optimizer
+    'learning_rate': 0.01
 })
 
 ### SEER ###
 DeepHit_seer = EasyDict({
     'batch_size': batch_size_seer,
-    'learning_rate': 0.01,
     'epochs': 100,
-    'hidden_size_indiv': 32,
-    'hidden_size_shared': 64,
-    'dropout': 0.1
+
+    # network
+    'hidden_layers_size': [32, 32],
+    'dropout': 0.1,
+
+    # loss
+    'alpha': 0.2,
+    'sigma': 0.1,
+
+    # optimizer
+    'learning_rate': 0.001
 })
 
 
@@ -65,32 +90,49 @@ DeepHit_seer = EasyDict({
 ### Metabric ###
 DeepSurv_metabric = EasyDict({
     'batch_size': batch_size_metabric,
-    'learning_rate': 0.01,
     'epochs': 100,
-    'hidden_size': 32,
-    'dropout': 0.1
+
+    # Network
+    'hidden_layers_size': [32, 32],
+    'dropout': 0.1,
+
+    # Adam
+    'learning_rate': 1e-2,
+    'weight_decay': 0,
 })
 
 ### SUPPORT ###
 DeepSurv_support = EasyDict({
     'batch_size': batch_size_support,
-    'learning_rate': 0.01,
     'epochs': 100,
-    'hidden_size': 32,
-    'dropout': 0.1
+
+    # Network
+    'hidden_layers_size': [32, 32],
+    'dropout': 0.1,
+
+    # Adam
+    'learning_rate': 1e-2,
+    'weight_decay': 0,
 })
 
 ### SEER ###
 DeepSurv_seer = EasyDict({
-    'batch_size': batch_size_support,
-    'learning_rate': 0.01,
+    'batch_size': batch_size_seer,
     'epochs': 100,
-    'hidden_size': 32,
+
+    # Network
+    'hidden_layers_size': [64, 64],
     'dropout': 0.1,
+
+    # Adam
+    'learning_rate': 1e-2,
+    'weight_decay': 0.1,
 })
 
 
 ########################## Deep Survival Machines ##########################
+# hidden_size (list): number of layers and number of nodex
+# k: number of underlying parametric distributions
 
 ### Metabric ###
 DSM_metabric = EasyDict({
@@ -134,46 +176,70 @@ DSM_seer = EasyDict({
 ### Metabric ###
 PCHazard_metabric = EasyDict({
     'batch_size': batch_size_metabric,
-    'learning_rate': 0.01,
     'epochs': 100,
-    'hidden_size': 32,
-    'dropout': 0.1
+
+    # Network
+    'hidden_layers_size': [64, 64],
+    'dropout': 0.1,
+
+    # AdamWR
+    'learning_rate': 1e-2,
+    'decoupled_weight_decay': 0.8,
+    'cycle_multiplier': 2
 })
 
+
 ### SUPPORT ###
 PCHazard_support = EasyDict({
     'batch_size': batch_size_support,
-    'learning_rate': 0.01,
     'epochs': 100,
-    'hidden_size': 32,
-    'dropout': 0.1
+
+    # Network
+    'hidden_layers_size': [32, 32],
+    'dropout': 0.1,
+
+    # AdamWR
+    'learning_rate': 1e-2,
+    'decoupled_weight_decay': 0.8,
+    'cycle_multiplier': 2
 })
 
 ### SEER ###
 PCHazard_seer = EasyDict({
     'batch_size': batch_size_seer,
-    'learning_rate': 0.01,
     'epochs': 100,
-    'hidden_size': 32,
+
+    # Network
+    'hidden_layers_size': [32, 32, 32, 32],
     'dropout': 0.1,
+
+    # AdamWR
+    'learning_rate': 1e-3,
+    'decoupled_weight_decay': 0.8,
+    'cycle_multiplier': 2
 })
 
 
 ########################## Random Survival Forests ##########################
+# epoch: in this context, refers to number of trees to generate in the forest.
+# max_depth: maximum depth of the tree.
 
 ### Metabric ###
 RSF_metabric = EasyDict({
-    'epochs': 100
+    'epochs': 200,
+    'max_depth': 4
 })
 
 ### SUPPORT ###
 RSF_support = EasyDict({
-    'epochs': 100
+    'epochs': 200,
+    'max_depth': 4
 })
 
 ### SEER ###
 RSF_seer = EasyDict({
-    'epochs': 100,
+    'epochs': 200,
+    'max_depth': 4
 })
 
 ########################## SurvTRACE #########################
@@ -181,7 +247,7 @@ RSF_seer = EasyDict({
 #   and not on different variants.
 
 ### METABRIC ###
-SurvTRACE_metabric = EasyDict(
+survtrace_metabric = EasyDict(
     {
         'num_durations': 5, # num of discrete intervals for prediction, e.g., num_dur = 5 means the whole period is discretized to be 5 intervals
         'seed': 1234,
@@ -210,15 +276,17 @@ SurvTRACE_metabric = EasyDict(
         'pruned_heads': {}, # no use
 
         # hyperparameters
-        'batch_size': 64,
+        'batch_size': batch_size_metabric,
         'weight_decay': 1e-4,
         'learning_rate': 1e-3,
-        'epochs': 100
+        'epochs': 100,
+        'gamma1': 1,
+        'gamma2': 1
     }
 )
 
 ### SUPPORT ###
-SurvTRACE_support = EasyDict(
+survtrace_support = EasyDict(
     {
         'num_durations': 5, # num of discrete intervals for prediction, e.g., num_dur = 5 means the whole period is discretized to be 5 intervals
         'seed': 1234,
@@ -247,16 +315,18 @@ SurvTRACE_support = EasyDict(
         'pruned_heads': {}, # no use
 
         # hyperparameters
-        'batch_size': 128,
+        'batch_size': batch_size_support,
         'weight_decay': 0,
         'learning_rate': 1e-3,
-        'epochs': 100
+        'epochs': 100,
+        'gamma1': 1,
+        'gamma2': 1
     }
 )
 
 
-### SUPPORT ###
-SurvTRACE_seer = EasyDict(
+### SEER ###
+survtrace_seer = EasyDict(
     {
         'num_durations': 5, # num of discrete intervals for prediction, e.g., num_dur = 5 means the whole period is discretized to be 5 intervals
         'seed': 1234,
@@ -274,7 +344,7 @@ SurvTRACE_seer = EasyDict(
         'num_event': 1, # only set when using SurvTraceMulti for competing risks
         'hidden_act': 'gelu',
         'attention_probs_dropout_prob': 0.1,
-        'early_stop_patience': 5,
+        'early_stop_patience': 10,
         'initializer_range': 0.02,
         'layer_norm_eps': 1e-12,
         'max_position_embeddings': 512, # # no use
@@ -286,9 +356,11 @@ SurvTRACE_seer = EasyDict(
         'val_batch_size': 10000,
 
         # hyperparameters
-        'batch_size': 1024,
+        'batch_size': batch_size_seer,
         'weight_decay': 0,
         'learning_rate': 1e-4,
-        'epochs': 100
+        'epochs': 100,
+        'gamma1': 1,
+        'gamma2': 1
     }
 )
\ No newline at end of file