Add ADNi to codebase

config/dataset/adni_tertiary.yaml

@@ -0,0 +1,4 @@
+_target_: sage.data.adni.ADNIClassification
+mode: train
+valid_ratio: 0.1
+seed: ${misc.seed}

config/sweep/adni_sweep.yaml

@@ -5,8 +5,6 @@ metric:
   goal: maximize
   name: test_acc
 parameters:
-  model:
-    values: [ resnet_binary , convnext_binary ]
   optim:
     values: [ adamw , lion ]
   scheduler:

config/train_cls.yaml

@@ -9,7 +9,7 @@ hydra:
 defaults:
   - _self_
   - model: convnext_cls
-  - dataset: ppmi
+  - dataset: adni
   - scheduler: exp_decay
   - optim: adamw
   - trainer: default

sage/constants.py

@@ -24,4 +24,4 @@
 
 EXT_BASE = DATA_BASE / "brain"
 PPMI_DIR = EXT_BASE / "PPMI"
-ADNI_DIR = EXT_BASE / "ADNI"
+ADNI_DIR = Path("adni") / "ADNI_3_reg"

sage/data/adni.py

@@ -15,20 +15,19 @@
 
 class ADNIBase(DatasetBase):
     NAME = "ADNI"
-    MAPPER2INT = {"ADNI 2": 0, "ADNI 3": 1}
     def __init__(self,
                  root: Path | str = C.ADNI_DIR,
-                 label_name: str = "adni_label.csv",
+                 label_name: str = "adni_labels_240509.csv",
                  mode: str = "train",
                  valid_ratio: float = .1,
-                 path_col: str = "abs_path",
-                 pk_col: str = "Subject ID",
-                 pid_col: str = "Subject ID",
-                 label_col: str = "Phase",
-                 strat_col: str = "Phase",
+                 path_col: str = "filepath",
+                 pk_col: str = "Subject",
+                 pid_col: str = "Subject",
+                 label_col: str = "Group",
+                 strat_col: str = "Group",
                  mod_col: str = None,
                  modality: List[str] = None,
-                 exclusion_fname: str = "donotuse-adni.txt",
+                 exclusion_fname: str = "",
                  augmentation: str = "monai",
                  seed: int = 42,):
         logger.warn("Please note that ADNI dataset label file should not have the exclusion file.")
@@ -58,20 +57,20 @@ def _exclude_data(self, labels: pd.DataFrame, pk_col: str, root: Path,
 
 class ADNIClassification(ADNIBase):
     NAME = "ADNI-CLS"
-    MAPPER2INT = {"ADNI 2": 0, "ADNI 3": 1}
+    MAPPER2INT = {"CN": 0, "MCI": 1, "AD": 2}
     def __init__(self,
                  root: Path | str = C.ADNI_DIR,
-                 label_name: str = "adni_label.csv",
+                 label_name: str = "adni_labels_240509.csv",
                  mode: str = "train",
                  valid_ratio: float = .1,
-                 path_col: str = "abs_path",
-                 pk_col: str = "Subject ID",
-                 pid_col: str = "Subject ID",
-                 label_col: str = "Phase",
-                 strat_col: str = "Phase",
-                 mod_col: str = None,
-                 modality: List[str] = None,
-                 exclusion_fname: str = "donotuse-adni.txt",
+                 path_col: str = "filepath",
+                 pk_col: str = "Subject",
+                 pid_col: str = "Subject",
+                 label_col: str = "Group",
+                 strat_col: str = "Group",
+                 mod_col: str = "Group",
+                 modality: List[str] = ["CN", "MCI", "AD"],
+                 exclusion_fname: str = "",
                  augmentation: str = "monai",
                  seed: int = 42,):
         super().__init__(root=root, label_name=label_name, mode=mode, valid_ratio=valid_ratio,
@@ -90,4 +89,28 @@ def _load_data(self, idx: int) -> Tuple[torch.Tensor]:
             logger.warn("Wrong label: %s\nData:%s", data[self.label_col], arr)
             raise
         label = torch.tensor(label, dtype=torch.long)
-        return arr, label
+        return arr, label
+
+
+class ADNIFullClassification(ADNIClassification):
+    NAME = "ADNI-ALL-CLS"
+    MAPPER2INT = {"CN": 0, "SMC": 1, "EMCI": 2, "MCI": 3, "LMCI": 4, "AD": 5}
+    def __init__(self,
+                 root: Path | str = C.ADNI_DIR,
+                 label_name: str = "adni_labels_240509.csv",
+                 mode: str = "train",
+                 valid_ratio: float = .1,
+                 path_col: str = "filepath",
+                 pk_col: str = "Subject",
+                 pid_col: str = "Subject",
+                 label_col: str = "Group",
+                 strat_col: str = "Group",
+                 mod_col: str = None,
+                 modality: List[str] = None,
+                 exclusion_fname: str = "",
+                 augmentation: str = "monai",
+                 seed: int = 42,):
+        super().__init__(root=root, label_name=label_name, mode=mode, valid_ratio=valid_ratio,
+                         path_col=path_col, pk_col=pk_col, pid_col=pid_col, label_col=label_col,
+                         strat_col=strat_col, mod_col=mod_col, modality=modality,
+                         exclusion_fname=exclusion_fname, augmentation=augmentation, seed=seed)

sage/data/dataloader.py

@@ -126,6 +126,9 @@ def __init__(self,
     def load_labels(self, root: Path, label_name: str = None, mode: str = None) -> pd.DataFrame:
         """ Load `.csv` """
         labels = pd.read_csv(root / label_name)
+        if mode is not None and "split" in labels.columns:
+            mode = "train" if mode != "test" else "test"
+            labels = labels[labels.split == mode]
         return labels
 
     def remove_duplicates(self, labels: pd.DataFrame) -> pd.DataFrame:
@@ -156,7 +159,7 @@ def split_data(self,
             trn_pid, val_pid = train_test_split(pid, test_size=valid_ratio, random_state=seed, stratify=y)
             trn = labels[labels[pid_col].isin(trn_pid)]
             val = labels[labels[pid_col].isin(val_pid)]
-        else:    
+        else:
             trn, val = train_test_split(labels, test_size=valid_ratio, random_state=seed)
         labels = {"train": trn, "valid": val}.get(mode, None)
         if labels is None:

sweep_command.sh

@@ -23,9 +23,9 @@ sweep_adni() {
     echo "Sweep on ADNI"
     python sweep.py --sweep_cfg_name=adni_sweep.yaml\
                     --wandb_project=adni\
-                    --config_name=train_binary.yaml\
+                    --config_name=train_cls.yaml\
                     --sweep_prefix='Scratch'\
-                    --overrides="['dataset=adni']"
+                    --overrides="['dataloader.batch_size=8']"
 }
 
 # Check the input argument and call the appropriate function