update data

.gitignore

@@ -11,12 +11,8 @@ __pycache__/
 
 .idea/
 
-# data & pretrain-model
-backup/
-data/
-pretrained/
-
 # private files
-utils_plot*
-imgs/
 1.md
+backup/
+all_data/
+pretrained/

EAS/run_search.py

@@ -47,7 +47,7 @@ def get_config():
     parser.add_argument('-p_runs', default=1, type=int)  # If batch_size is 1, set this to > 1 to do multiple runs for the instance in parallel
     parser.add_argument('-output_path', default="EAS_results", type=str)
     parser.add_argument('-norm', default="batch_no_track", choices=['instance', 'batch', 'batch_no_track', 'none'], type=str)
-    parser.add_argument('-gpu_id', default=2, type=int)
+    parser.add_argument('-gpu_id', default=0, type=int)
     parser.add_argument('-seed', default=2023, type=int, help="random seed")
 
     # EAS-Emb and EAS-Lay parameters
@@ -247,7 +247,9 @@ def search(run_id, config):
         logging.info(f"Runtime: {runtime}s")
         logging.info("MEM: " + str(cutorch.max_memory_reserved(config.gpu_id) / 1024 / 1024) + "MB")
         logging.info(f"Num. instances: {len(perf)}")
-        print(">> Solving {}, with sol {}".format(config.instances_path, perf))
+        print(">> Solved {}, with sol {}".format(config.instances_path, perf))
+
+    return np.mean(perf)
 
 
 def seed_everything(seed=2022):
@@ -263,4 +265,15 @@ def seed_everything(seed=2022):
     seed_everything(config.seed)
     if torch.cuda.is_available():
         torch.cuda.set_device(config.gpu_id)
-    search(run_id, config)
+
+    sol_list, path_list = [], []
+    if os.path.isdir(config.instances_path):
+        path_list = [os.path.join(config.instances_path, f) for f in sorted(os.listdir(config.instances_path))]
+    else:
+        path_list = config.instances_path
+
+    for path in path_list:
+        config.instances_path = path
+        sol = search(run_id, config)
+        sol_list.append(sol)
+    print(len(sol_list), sol_list)

POMO/CVRP/CVRPEnv.py

@@ -101,7 +101,8 @@ def load_problems(self, batch_size, problems=None, aug_factor=1):
             node_demand = self.saved_node_demand[self.saved_index:self.saved_index+batch_size]
             self.saved_index += batch_size
         else:
-            depot_xy, node_xy, node_demand = get_random_problems(batch_size, self.problem_size, distribution='uniform', problem="cvrp")
+            depot_xy, node_xy, node_demand, capacity = get_random_problems(batch_size, self.problem_size, distribution='uniform', problem="cvrp")
+            node_demand = node_demand / capacity.view(-1, 1)
         self.batch_size = depot_xy.size(0)
 
         if aug_factor > 1:

POMO/CVRP/CVRP_baseline.py

@@ -12,7 +12,7 @@
 from utils.functions import check_extension, load_dataset, save_dataset, run_all_in_pool, move_to
 
 
-def get_lkh_executable(url="http://www.akira.ruc.dk/~keld/research/LKH-3/LKH-3.0.7.tgz"):
+def get_lkh_executable(url="http://www.akira.ruc.dk/~keld/research/LKH-3/LKH-3.0.8.tgz"):
 
     cwd = os.path.abspath("lkh")
     os.makedirs(cwd, exist_ok=True)

POMO/TSP/TSPTrainer_Meta.py

@@ -229,7 +229,7 @@ def _train_one_epoch(self, epoch):
                 w[b] = self.task_w[idx][selected].item()
             selected_tasks.append(task_params)
         w = torch.softmax(torch.Tensor(w), dim=0)
-            
+
         for b in range(self.meta_params['B']):
             task_params, task_w = selected_tasks[b], w[b].item()
             # preparation

POMO/TSP/TSP_baseline.py

@@ -98,7 +98,7 @@ def solve_concorde_log(executable, directory, name, loc, disable_cache=False):
         return None
 
 
-def get_lkh_executable(url="http://www.akira.ruc.dk/~keld/research/LKH-3/LKH-3.0.7.tgz"):
+def get_lkh_executable(url="http://www.akira.ruc.dk/~keld/research/LKH-3/LKH-3.0.8.tgz"):
 
     cwd = os.path.abspath("lkh")
     os.makedirs(cwd, exist_ok=True)