Save best model at the right place

fteicht · Mar 15, 2023 · 4c27aec · 4c27aec
1 parent 4e5f8fd
commit 4c27aec
Showing 1 changed file with 10 additions and 4 deletions.
diff --git a/gnn4rcpsp/learn_schedules.py b/gnn4rcpsp/learn_schedules.py
@@ -43,7 +43,6 @@ def check_solution(data):
         "makespan": [],
     }
     for data in data_batch.to_data_list():
-
         t2t, dur, r2t, rc, ref_makespan = (
             data.t2t.view(len(data.dur), -1).data.cpu().detach().numpy(),
             data.dur.data.cpu().detach().numpy(),
@@ -303,7 +302,6 @@ def hook(module, input, output, name=name):
                 writer.add_scalar("loss", loss.item(), step)
 
             if COMPUTE_METRICS:
-
                 for batch_idx, data in enumerate(test_loader):
                     data.to(device)
                     out = model(data)
@@ -356,7 +354,7 @@ def hook(module, input, output, name=name):
                     best_sgs_makespan_ref = epoch_violations["sgs_makespan_ref"]
                     shutil.copyfile(
                         f"saved_models/ResTransformer-256-50000/model_{kfold_idx}_{epoch}.tch",
-                        "best_model.tch",
+                        "saved_models/ResTransformer-256-50000/best_model.tch",
                     )
 
             writer.flush()
@@ -368,7 +366,6 @@ def hook(module, input, output, name=name):
 
 
 if __name__ == "__main__":
-
     file_path = os.path.realpath(__file__)
 
     if not os.path.exists(os.path.join(os.path.dirname(file_path), "data_list.tch")):
@@ -399,6 +396,14 @@ def hook(module, input, output, name=name):
         # train_list += random.sample(medium_list, int(0.5 * len(medium_list)))
         torch.save(train_list, "./train_list.tch")
 
+    # Only if we are in small-medium training mode
+    if os.path.exists(os.path.join(os.path.dirname(file_path), "validation_list.tch")):
+        validation_list = torch.load("validation_list.tch")
+    else:
+        big_inst_list = list(set(range(len(data_list))) - set(train_list))
+        validation_list = random.sample(big_inst_list, int(0.5 * len(big_inst_list)))
+        torch.save(validation_list, "./validation_list.tch")
+
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
     Net = ResTransformer
     # Net = ResGINE
@@ -419,4 +424,5 @@ def hook(module, input, output, name=name):
         optimizer=optimizer,
         device=device,
         writer=writer,
+        validation_list=validation_list,
     )