Minor fix to ensure samples from all classes are present in spider_thorax validation set

Author: PierreMarza

src/thunder/config/dataset/spider_thorax.yaml

@@ -13,10 +13,10 @@ compatible_tasks:
     "adversarial_attack",
   ]
 
-nb_train_samples: 50413
-nb_val_samples: 12906
+nb_train_samples: 49562
+nb_val_samples: 13757
 nb_test_samples: 14988
-md5sum: "5d91551c7b0a4d82639411fcc2af847e"
+md5sum: "5d1aab1441a442b1cc4a6ed6c14d4718"
 image_sizes: [[224, 224]]
 mpp: 0.5
 cancer_type: skin

src/thunder/datasets/dataset/spider_thorax.py

@@ -97,6 +97,17 @@ def create_splits_spider_thorax(base_folder: str, dataset_cfg: dict) -> None:
     val_slide_names = random.sample(
         unique_slide_names, int(0.2 * len(unique_slide_names))
     )
+
+    # Class 12 is only present on 4 slides in the spider_thorax dataset
+    # (significantly less represented than other classes)
+    # so we need to make sure to have one in val set and the 3 others in train set
+    for slide_name in ["slide_0021", "slide_0048", "slide_0123", "slide_0178"]:
+        # Looping over the only 4 slides containing class 12 samples
+        if slide_name != "slide_0048" and slide_name in val_slide_names:
+            val_slide_names.remove(slide_name)
+        elif slide_name == "slide_0048" and slide_name not in val_slide_names:
+            val_slide_names.append(slide_name)
+
     val_mask = np.array([slide_name in val_slide_names for slide_name in slide_names])
 
     (