Fix path to save last_model.pkl

classify_utils/cal_acc_MCP.py

@@ -116,8 +116,11 @@ def cal_top1_topk(feats, cent_MCP, concept_vecs, concept_means, selected_node, a
     parser.add_argument('--l2', default = [], type = int, nargs = "+", help = "Select the node to drop")
     parser.add_argument('--l3', default = [], type = int, nargs = "+", help = "Select the node to drop")
     parser.add_argument('--l4', default = [], type = int, nargs = "+", help = "Select the node to drop")
+    parser.add_argument("--saved_dir", default = ".", type = str)
     args = parser.parse_args()
 
+    args.dst = f"{args.saved_dir}/pkl/{args.case_name}/{args.model.lower()}_{args.basic_model.lower()}"
+
     print("Calculate accuracy !!")
     os.environ["CUDA_VISIBLE_DEVICES"] = args.device
 
@@ -138,7 +141,7 @@ def cal_top1_topk(feats, cent_MCP, concept_vecs, concept_means, selected_node, a
         num_class = int(num_class * 0.8 + 0.5)
 
     model = load_model(args.model, args.basic_model.lower(), num_class)
-    load_weight(model, f"./pkl/{case_name}/{args.model.lower()}_{args.basic_model}/best_model.pkl")
+    load_weight(model, f"{args.dst}/best_model.pkl")
     model.eval()
 
     post_name = ""

classify_utils/cal_class_MCP.py

@@ -78,10 +78,13 @@ def cal_class_MCP(model, concept_vecs, concept_means, data_transforms, data_path
     parser.add_argument('--w_mode', default = "best", choices=['best', 'last'], type = str)
     parser.add_argument('--all_class', default = False, action = "store_true", help = "if true, generate the whole dataset's classes MCP distribution.")
     parser.add_argument('--few_shot', default = False, action = "store_true")
+    parser.add_argument("--saved_dir", default = ".", type = str)
 
     args = parser.parse_args()
     print("Calculate the class MCP distribution !!")
     os.environ["CUDA_VISIBLE_DEVICES"] = args.device
+    args.dst = f"{args.saved_dir}/pkl/{args.case_name}/{args.model.lower()}_{args.basic_model.lower()}"
+
 
     case_name = args.case_name
     args, image_size = get_model_set(args)
@@ -100,7 +103,7 @@ def cal_class_MCP(model, concept_vecs, concept_means, data_transforms, data_path
         num_class = int(num_class * 0.8 + 0.5)
 
     model = load_model(args.model, args.basic_model.lower(), num_class)
-    trained_param_path = f"./pkl/{args.case_name}/{args.model.lower()}_{args.basic_model}/{args.w_mode}_model.pkl"
+    trained_param_path = f"{args.dst}/{args.w_mode}_model.pkl"
     load_weight(model, trained_param_path)
     model.eval()
 

extract_prototypes.py

@@ -119,10 +119,13 @@ def load_dataset(data_path: str, image_size: int, args, data_transform = None) -
     parser.add_argument('--basic_model', default = "resnet50", type = str)
     parser.add_argument('--model', default = "aix_model", type = str)
     parser.add_argument('--w_mode', default = "best", choices=['best', 'last'], type = str)
+    parser.add_argument("--saved_dir", default = ".", type = str)
     args = parser.parse_args()
 
     os.environ["CUDA_VISIBLE_DEVICES"] = args.device
 
+    args.dst = f"{args.saved_dir}/pkl/{args.case_name}/{args.model.lower()}_{args.basic_model.lower()}"
+
     os.makedirs(f"./PCA_concept_specific_tmp/{args.case_name}/{args.basic_model}/", exist_ok = True)
     data_path, train_path, val_path, num_class = get_dataset(args.case_name)
 
@@ -143,7 +146,7 @@ def load_dataset(data_path: str, image_size: int, args, data_transform = None) -
     print(args)
 
     train_loader = load_dataset(data_path, image_size, args)
-    trained_param_path = f"./pkl/{args.case_name}/{args.model.lower()}_{args.basic_model}/best_model.pkl"
+    trained_param_path = f"{args.dst}/best_model.pkl"
     load_weight(model, trained_param_path)
 
     post_name = ""

train.py

@@ -404,7 +404,7 @@ def runs(args):
                                 "Lr_scheduler" : lr_state_dict,
                                 "Best ACC" : max_acc[phase].avg,
                                 "concept_cha" : args.concept_cha}
-                torch.save(save_data, './pkl/{}/{}_{}/last_model.pkl'.format(args.index, args.model.lower(), args.basic_model.lower()))
+                torch.save(save_data, f'{args.dst}/last_model.pkl')
                 # -------------------------------------------------------------
                 info_log('Index : {}'.format(args.index), args.global_rank, args.log_type, args.log)
                 info_log("dataset : {}".format(args.dataset_name), args.global_rank, args.log_type, args.log)

vis_utils/find_topk_area.py

@@ -102,10 +102,13 @@ def save_masked(ori_imgs, masks, args):
     parser.add_argument('--eigen_topk', default = 1, type = int)
     parser.add_argument('--heatmap', action = "store_true", default = False)
     parser.add_argument('--masked', action = "store_true", default = False)
+    parser.add_argument("--saved_dir", default = ".", type = str)
     args = parser.parse_args()
     os.environ["CUDA_VISIBLE_DEVICES"] = args.device
     print(args)
 
+    args.dst = f"{args.saved_dir}/pkl/{args.case_name}/{args.model.lower()}_{args.basic_model.lower()}"
+
     nrow = args.topk
 
     layer_sizes = []
@@ -147,7 +150,7 @@ def save_masked(ori_imgs, masks, args):
         if args.model.lower() != "aix_model":
             post_name = "_resp"
         model = load_model(args.model, args.basic_model, num_class)
-        trained_param_path = f"./pkl/{args.case_name}/{args.model.lower()}_{args.basic_model}/best_model.pkl"
+        trained_param_path = f"{args.dst}/best_model.pkl"
         print(trained_param_path)
         load_weight(model, trained_param_path)
 

vis_utils/find_topk_response.py

@@ -44,7 +44,10 @@ def __getitem__(self, idx):
     parser.add_argument('--concept_per_layer', default = [32, 32, 32, 32], type = int, nargs = "+")
     parser.add_argument('--eigen_topk', default = 1, type = int)
     parser.add_argument('--use_CLS_token', action = "store_true", default = False)
+    parser.add_argument("--saved_dir", default = ".", type = str)
     args = parser.parse_args()
+
+
 
     print(args)
     os.environ["CUDA_VISIBLE_DEVICES"] = args.device
@@ -63,6 +66,8 @@ def __getitem__(self, idx):
                                      transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
     TOP_RATE = 0.1
     for step, case_name in enumerate(case_names):
+        args.dst = f"{args.saved_dir}/pkl/{case_name}/{args.model.lower()}_{args.basic_model.lower()}"
+
         data_path, train_path, val_path, num_class = get_dataset(case_name)
         data_path = data_path + train_path
 
@@ -76,7 +81,7 @@ def __getitem__(self, idx):
         concept_vecs, concept_means = load_concept(concept_covs, concept_means, args.eigen_topk)
 
         model = load_model(args.model, args.basic_model, num_class).cuda()
-        trained_param_path = f"./pkl/{case_name}/{args.model.lower()}_{args.basic_model}/best_model.pkl"
+        trained_param_path = f"{args.dst}/best_model.pkl"
         load_weight(model, trained_param_path)
 
         max_resp_path = [np.array([]),