Merge branch 'master' of https://github.com/Dipamc77/procgen-competition

experiments/custom-ppg.yaml

@@ -45,13 +45,13 @@ procgen-ppo:
         no_done_at_end: False
         
         # Custom switches
-        skips: 6
-        n_pi: 10
-        num_retunes: 16
-        retune_epochs: 6
+        skips: 2
+        n_pi: 16
+        num_retunes: 14
+        retune_epochs: 7
         standardize_rewards: True
-        aux_mbsize: 2
-        aux_num_accumulates: 4
+        aux_mbsize: 4
+        aux_num_accumulates: 2
         augment_buffer: True
         scale_reward: 1.0
         reset_returns: False
@@ -59,9 +59,9 @@ procgen-ppo:
         augment_randint_num: 3 ## Hacky name fix later
         aux_lr: 5.0e-4
         value_lr: 1.0e-3
-        same_lr_everywhere: False
+        same_lr_everywhere: True
         aux_phase_mixed_precision: True
-        single_optimizer: True
+        single_optimizer: False
         max_time: 7200
         pi_phase_mixed_precision: False
         
@@ -72,7 +72,7 @@ procgen-ppo:
         entropy_schedule: False
         
         # Memory management, if batch size overflow, batch splitting is done to handle it
-        max_minibatch_size: 500
+        max_minibatch_size: 1000
         updates_per_batch: 8
 
         normalize_actions: False
@@ -89,10 +89,8 @@ procgen-ppo:
         model:
             custom_model: impala_torch_ppg
             custom_model_config:
-#                 depths: [32, 64, 64]
-#                 nlatents: 512
-                depths: [64, 128, 128]
-                nlatents: 1024
+                depths: [32, 64, 64]
+                nlatents: 512
                 init_normed: True
                 use_layernorm: False
                 diff_framestack: True

experiments/custom-torch-ppo.yaml

@@ -46,15 +46,15 @@ procgen-ppo:
         no_done_at_end: False
         
         # Custom switches
-        retune_skips: 100000
-        retune_replay_size: 450000
+        retune_skips: 350000
+        retune_replay_size: 200000
         num_retunes: 13
-        retune_epochs: 3
+        retune_epochs: 6
         standardize_rewards: True
         scale_reward: 1.0
         return_reset: False
         aux_phase_mixed_precision: True
-        max_time: 1000000
+        max_time: 7200
         
         adaptive_gamma: False
         final_lr: 5.0e-5