[Configs] updated configs

configs/experiment/routing/am-critic.yaml

@@ -0,0 +1,14 @@
+# @package _global_
+
+# Use the following to take the default values from am.yaml
+# Replace below only the values that you want to change compared to the default values
+defaults:
+  - routing/am.yaml
+
+logger:
+  wandb:
+    tags: ["am-critic", "${env.name}"]
+    name: am-critic-${env.name}${env.num_loc}
+
+model:
+  baseline: "critic"

configs/experiment/routing/am-ppo.yaml

@@ -0,0 +1,50 @@
+# @package _global_
+
+defaults:
+  - override /model: am-ppo.yaml
+  - override /env: tsp.yaml
+  - override /callbacks: default.yaml
+  - override /trainer: default.yaml
+  - override /logger: wandb.yaml
+
+env:
+  num_loc: 50
+
+logger:
+  wandb:
+    project: "rl4co"
+    tags: ["am-ppo", "${env.name}"]
+    group: ${env.name}${env.num_loc}
+    name: am-${env.name}${env.num_loc}
+
+model:
+  batch_size: 512
+  val_batch_size: 1024
+  test_batch_size: 1024
+  train_data_size: 1_280_000
+  val_data_size: 10_000
+  test_data_size: 10_000
+  clip_range: 0.2
+  ppo_epochs: 2
+  mini_batch_size: 512
+  vf_lambda: 0.5
+  entropy_lambda: 0.01
+  normalize_adv: False
+  max_grad_norm: 0.5
+  optimizer_kwargs:
+    lr: 1e-4
+    weight_decay: 0
+  lr_scheduler:
+    "MultiStepLR"
+  lr_scheduler_kwargs:
+    milestones: [80, 95]
+    gamma: 0.1
+
+trainer:
+  max_epochs: 100
+  gradient_clip_val: Null # not supported in manual optimization
+
+seed: 1234
+
+metrics:
+  train: ["loss", "reward", "surrogate_loss", "value_loss", "entropy_bonus"]

configs/experiment/tsp/am.yaml → configs/experiment/routing/am-xl.yaml

@@ -15,26 +15,29 @@ logger:
     project: "rl4co"
     tags: ["am", "${env.name}"]
     group: ${env.name}${env.num_loc}"
-    name: "am-${env.name}${env.num_loc}"
-
+    name: "am-xl-${env.name}${env.num_loc}"
 
 model:
-  batch_size: 512
+  policy_kwargs:
+    num_encoder_layers: 6
+    normalization: 'instance'
+  batch_size: 2048
   val_batch_size: 1024
   test_batch_size: 1024
   train_data_size: 1_280_000
   val_data_size: 10_000
   test_data_size: 10_000
   optimizer_kwargs:
     lr: 1e-4
+    weight_decay: 0
   lr_scheduler:
     "MultiStepLR"
   lr_scheduler_kwargs:
-    milestones: [80, 95]
+    milestones: [480, 495]
     gamma: 0.1
 
 trainer:
-  max_epochs: 100
+  max_epochs: 500
 
 seed: 1234
 

configs/experiment/routing/am.yaml

@@ -14,8 +14,8 @@ logger:
   wandb:
     project: "rl4co"
     tags: ["am", "${env.name}"]
-    group: ${env.name}${env.num_loc}"
-    name: "am-${env.name}${env.num_loc}"
+    group: ${env.name}${env.num_loc}
+    name: am-${env.name}${env.num_loc}
 
 model:
   batch_size: 512

configs/experiment/routing/pomo.yaml

@@ -0,0 +1,38 @@
+# @package _global_
+
+defaults:
+  - override /model: pomo.yaml
+  - override /env: tsp.yaml
+  - override /callbacks: default.yaml
+  - override /trainer: default.yaml
+  - override /logger: wandb.yaml
+
+env:
+  num_loc: 50
+
+logger:
+  wandb:
+    project: "rl4co"
+    tags: ["pomo", "${env.name}"]
+    group: ${env.name}${env.num_loc}"
+    name: "pomo-${env.name}${env.num_loc}"
+
+model:
+  batch_size: 64
+  train_data_size: 160_000
+  val_data_size: 10_000
+  test_data_size: 10_000
+  optimizer_kwargs:
+    lr: 1e-4
+    weight_decay: 1e-6
+  lr_scheduler:
+    "MultiStepLR"
+  lr_scheduler_kwargs:
+    milestones: [95]
+    gamma: 0.1
+
+trainer:
+  max_epochs: 100
+
+seed: 1234
+

configs/experiment/routing/symnco.yaml

@@ -0,0 +1,42 @@
+# @package _global_
+
+defaults:
+  - override /model: symnco.yaml
+  - override /env: tsp.yaml
+  - override /callbacks: default.yaml
+  - override /trainer: default.yaml
+  - override /logger: wandb.yaml
+
+env:
+  num_loc: 50
+
+logger:
+  wandb:
+    project: "rl4co"
+    tags: ["symnco", "${env.name}"]
+    group: ${env.name}${env.num_loc}"
+    name: "symnco-${env.name}${env.num_loc}"
+
+model:
+  batch_size: 512
+  val_batch_size: 1024
+  test_batch_size: 1024
+  train_data_size: 1_280_000
+  val_data_size: 10_000
+  test_data_size: 10_000
+  num_starts: 0 # 0 for no augmentation for multi-starts
+  num_augment: 10
+  optimizer_kwargs:
+    lr: 1e-4
+    weight_decay: 0
+  lr_scheduler:
+    "MultiStepLR"
+  lr_scheduler_kwargs:
+    milestones: [80, 95]
+    gamma: 0.1
+
+trainer:
+  max_epochs: 100
+
+seed: 1234
+

configs/model/pomo.yaml

@@ -1 +1,10 @@
-_target_: rl4co.models.POMO
+_target_: rl4co.models.POMO
+
+# During val/test only we apply dihedral transformations
+num_augment: 8
+
+metrics:
+  train: ["loss", "reward"]
+  val: ["reward", "max_reward", "max_aug_reward"]
+  test: ${metrics.val}
+  log_on_step: True

configs/model/symnco.yaml

@@ -1,4 +1,10 @@
 _target_: rl4co.models.SymNCO
 
-num_augment: 8
-num_starts: ${env.num_loc}
+num_augment: 10
+num_starts: 0 # by default we use only symmetric augmentations
+
+metrics:
+  train: ["loss", "loss_ss", "loss_ps", "loss_inv", "reward"]
+  val: ["reward", "max_reward", "max_aug_reward"]
+  test: ${metrics.val}
+  log_on_step: True