Added single experiment and other scripts; mostly debugging large mar…

…gin loss + logit suppression outlier exposure
tnoe1 · Jul 28, 2022 · baf7c1e · baf7c1e
1 parent 4f3953e
commit baf7c1e
Show file tree

Hide file tree

Showing 12 changed files with 5,729 additions and 36 deletions.
diff --git a/experiment.py b/experiment.py
@@ -194,9 +194,9 @@ def main():
     # Setup Weights and Biases and specify hyperparameters
     test_method = "oe_test" if args.oe_test else "new_class_test"
     if args.baseline:
-        project_name = "TN_Masters_Proj_Baseline_{}".format(test_method)
+        project_name = "TN_Masters_Proj_Baseline_Val_{}".format(test_method)
     else:
-        project_name = "TN_Masters_Proj_{}_{}_{}".format(args.detection_type, args.loss, test_method)
+        project_name = "TN_Masters_Proj_Val_{}_{}_{}".format(args.detection_type, args.loss, test_method)
 
     wandb.init(project=project_name)
 
@@ -323,34 +323,34 @@ def main():
         wandb.log({"loss": loss, "ID_Accuracy": acc, "AUROC": auc, "metric_combined": metric_combined, "epoch": i})
 
         # Save Model
-        if i % 5 == 0 or metric_combined > metric_combined_running_max:
-            if metric_combined > metric_combined_running_max:
-                metric_combined_running_max = metric_combined
-
-            if args.baseline:
-                directory = "baseline_{}".format(test_method)
-                torch.save({
-                    'epoch': i,
-                    'model_state_dict': net.state_dict(),
-                    'optimizer_state_dict': optim.state_dict(),
-                    'auc': auc,
-                    'id_accuracy': acc,
-                    'split': args.split,
-                    'test_method': test_method,
-                }, 'models/{}/day_{}_{}_time_{}_{}_split_{}_epoch_{}.pth'.format(directory, now.month, now.day, now.hour, now.minute, args.split, i))
-            else:
-                directory = "{}_{}_{}".format(args.loss, args.detection_type, test_method)
-                torch.save({
-                    'epoch': i,
-                    'model_state_dict': net.state_dict(),
-                    'optimizer_state_dict': optim.state_dict(),
-                    'auc': auc,
-                    'id_accuracy': acc,
-                    'split': args.split,
-                    'test_method': test_method,
-                    'loss': args.loss,
-                    'detection_type': args.detection_type,
-                }, 'models/{}/day_{}_{}_time_{}_{}_split_{}_epoch_{}.pth'.format(directory, now.month, now.day, now.hour, now.minute, args.split, i))
+        # if i % 5 == 0 or metric_combined > metric_combined_running_max:
+        #     if metric_combined > metric_combined_running_max:
+        #         metric_combined_running_max = metric_combined
+
+        if args.baseline:
+            directory = "val_baseline_{}".format(test_method)
+            torch.save({
+                'epoch': i,
+                'model_state_dict': net.state_dict(),
+                'optimizer_state_dict': optim.state_dict(),
+                'auc': auc,
+                'id_accuracy': acc,
+                'split': args.split,
+                'test_method': test_method,
+            }, 'models/{}/day_{}_{}_time_{}_{}_split_{}_epoch_{}.pth'.format(directory, now.month, now.day, now.hour, now.minute, args.split, i))
+        else:
+            directory = "val_{}_{}_{}".format(args.loss, args.detection_type, test_method)
+            torch.save({
+                'epoch': i,
+                'model_state_dict': net.state_dict(),
+                'optimizer_state_dict': optim.state_dict(),
+                'auc': auc,
+                'id_accuracy': acc,
+                'split': args.split,
+                'test_method': test_method,
+                'loss': args.loss,
+                'detection_type': args.detection_type,
+            }, 'models/{}/day_{}_{}_time_{}_{}_split_{}_epoch_{}.pth'.format(directory, now.month, now.day, now.hour, now.minute, args.split, i))
 
 if __name__ == '__main__':
     main()