Merge pull request #236 from hyeok9855/main

[Feat] add GFACS and replace PyVRP with HGS for CVRP local search

.gitignore

@@ -178,4 +178,7 @@ local_nbrun*
 
 # osx meta files
 .DS_Store
-outputs/
+outputs/
+
+# HGS-CVRP for cvrp local search
+HGS-CVRP/

configs/experiment/routing/deepaco.yaml

@@ -19,37 +19,62 @@ logger:
     name: deepaco-${env.name}${env.generator_params.num_loc}
 
 model:
-  batch_size: 512
-  val_batch_size: 1000
-  test_batch_size: 1000
-  train_data_size: 128_000
-  val_data_size: 10_000
-  test_data_size: 10_000
+  batch_size: 20
+  val_batch_size: 20
+  test_batch_size: 20
+  train_data_size: 400
+  val_data_size: 20
+  test_data_size: 100
+  optimizer: "AdamW"
   optimizer_kwargs:
-    lr: 1e-4
+    lr: 5e-4
     weight_decay: 0
   lr_scheduler:
-    "MultiStepLR"
+    "CosineAnnealingLR"
   lr_scheduler_kwargs:
-    milestones: [25, 35]
-    gamma: 0.1
+    T_max: 50
+    eta_min: 1e-5
+  metrics:
+    test:
+      - reward_000
+      - reward_002
+      - reward_009  # since n_iterations["text"] = 10
+  train_with_local_search: True
+  ls_reward_aug_W: 0.99
 
   policy_kwargs:
     n_ants:
-      train: 50
-      val: 20
-      test: 20
+      train: 30
+      val: 30
+      test: 100
     n_iterations:
       train: 1 # unused value
-      val: 20
-      test: 20
+      val: 5
+      test: 10
+    temperature: 1.0
+    top_p: 0.0
+    top_k: 0
+    start_node: null
+    multistart: False
+    k_sparse: 5  # this should be adjusted based on the `num_loc` value
+
     aco_kwargs:
       alpha: 1.0
       beta: 1.0
       decay: 0.95
-
+      use_local_search: True
+      use_nls: True
+      n_perturbations: 5
+      local_search_params:
+        max_iterations: 1000
+      perturbation_params:
+        max_iterations: 20
 
 trainer:
-  max_epochs: 40
+  max_epochs: 50
+  gradient_clip_val: 3.0
+  precision: "bf16-mixed"
+  device:
+    - 0
 
-seed: 1234
+seed: 1234

configs/experiment/routing/gfacs.yaml

@@ -0,0 +1,84 @@
+# @package _global_
+
+defaults:
+  - override /model: gfacs.yaml
+  - override /env: tsp.yaml
+  - override /callbacks: default.yaml
+  - override /trainer: default.yaml
+  - override /logger: wandb.yaml
+
+env:
+  generator_params:
+    num_loc: 50
+
+logger:
+  wandb:
+    project: "rl4co"
+    tags: ["gfacs", "${env.name}"]
+    group: ${env.name}${env.generator_params.num_loc}
+    name: gfacs-${env.name}${env.generator_params.num_loc}
+
+model:
+  batch_size: 20
+  val_batch_size: 20
+  test_batch_size: 20
+  train_data_size: 400
+  val_data_size: 20
+  test_data_size: 100
+  optimizer: "AdamW"
+  optimizer_kwargs:
+    lr: 5e-4
+    weight_decay: 0
+  lr_scheduler:
+    "CosineAnnealingLR"
+  lr_scheduler_kwargs:
+    T_max: 50
+    eta_min: 1e-5
+  metrics:
+    test:
+      - reward_000
+      - reward_002
+      - reward_009  # since n_iterations["text"] = 10
+  train_with_local_search: True
+  alpha_min: 0.5
+  alpha_max: 1.0
+  alpha_flat_epochs: 5
+  beta_min: 100
+  beta_max: 500
+  beta_flat_epochs: 5
+
+  policy_kwargs:
+    n_ants:
+      train: 30
+      val: 30
+      test: 100
+    n_iterations:
+      train: 1 # unused value
+      val: 5
+      test: 10
+    temperature: 1.0
+    top_p: 0.0
+    top_k: 0
+    multistart: False
+    k_sparse: 5  # this should be adjusted based on the `num_loc` value
+
+    aco_kwargs:
+      alpha: 1.0  # This alpha is different from the alpha in the model
+      beta: 1.0  # This beta is different from the beta in the model
+      decay: 0.95
+      use_local_search: True
+      use_nls: True
+      n_perturbations: 5
+      local_search_params:
+        max_iterations: 1000
+      perturbation_params:
+        max_iterations: 20
+
+trainer:
+  max_epochs: 50
+  gradient_clip_val: 3.0
+  precision: "bf16-mixed"
+  devices:
+    - 0
+
+seed: 1234

configs/model/deepaco.yaml

@@ -1,3 +1,3 @@
 _target_: rl4co.models.DeepACO
 
-baseline: "exponential"
+baseline: "no"

configs/model/gfacs.yaml

@@ -0,0 +1,3 @@
+_target_: rl4co.models.GFACS
+
+baseline: "no"

rl4co/envs/routing/cvrp/README.md

@@ -0,0 +1,20 @@
+As local search in CVRP, we adopted the SWAP* algorithm by Vidal et al. [1, 2] ([repo](https://github.com/vidalt/HGS-CVRP)). Specifically, we use the modified version of code provided by [DeepACO](https://github.com/henry-yeh/DeepACO/tree/main/cvrp_nls/HGS-CVRP-main), which we uploaded to [https://github.com/ai4co/HGS-CVRP](https://github.com/ai4co/HGS-CVRP) for convenience.
+
+
+### Installation
+
+```bash
+cd rl4co/envs/routing/cvrp
+git clone [email protected]:ai4co/HGS-CVRP.git
+cd HGS-CVRP
+bash build.sh
+```
+
+### References
+
+[1] Vidal, T., Crainic, T. G., Gendreau, M., Lahrichi, N., Rei, W. (2012). A hybrid genetic algorithm for multidepot and periodic vehicle routing problems. Operations Research, 60(3), 611-624.
+
+[2] Vidal, T. (2022). Hybrid genetic search for the CVRP: Open-source implementation and SWAP* neighborhood. Computers & Operations Research, 140, 105643.
+
+[3] Ye, H., Wang J., Cao, Z., Liang, H., Li, Y. (2023).
+DeepACO: Neural-enhanced ant systems for combinatorial optimization. Advances in neural information processing systems (NeurIPS) 36 (2023): 43706-43728.

rl4co/envs/routing/cvrp/env.py

@@ -14,8 +14,7 @@
 
 try:
     from .local_search import local_search
-except Exception:
-    # In case some dependencies are not installed (e.g., pyvrp)
+except:  # In case when we fail to build HGS
     local_search = None
 from .render import render
 
@@ -133,7 +132,7 @@ def _reset(
     @staticmethod
     def get_action_mask(td: TensorDict) -> torch.Tensor:
         # For demand steps_dim is inserted by indexing with id, for used_capacity insert node dim for broadcasting
-        exceeds_cap = td["demand"] + td["used_capacity"] > td["vehicle_capacity"]
+        exceeds_cap = td["demand"] + td["used_capacity"] > td["vehicle_capacity"] + 1e-5
 
         # Nodes that cannot be visited are already visited or too much demand to be served now
         mask_loc = td["visited"][..., 1:].to(exceeds_cap.dtype) | exceeds_cap
@@ -182,7 +181,7 @@ def check_solution_validity(td: TensorDict, actions: torch.Tensor):
             # Cannot use less than 0
             used_cap[used_cap < 0] = 0
             assert (
-                used_cap <= td["vehicle_capacity"] + 1e-5
+                used_cap <= td["vehicle_capacity"][:, 0] + 1e-5
             ).all(), "Used more than capacity"
 
     @staticmethod
@@ -257,7 +256,7 @@ def replace_selected_actions(
     def local_search(td: TensorDict, actions: torch.Tensor, **kwargs) -> torch.Tensor:
         assert (
             local_search is not None
-        ), "Cannot import local_search module. Check if `pyvrp` is installed."
+        ), "Cannot import local_search module. Check `rl4co/envs/routing/cvrp/README.md` for instructions to build HGS."
         return local_search(td, actions, **kwargs)
 
     @staticmethod