Upload folder using huggingface_hub

ZoomLDM-demo-dataset.py

@@ -24,9 +24,11 @@ _MAG_DICT = {
 
 _FIXED_SSL_FEATURE_DIM_1 = 1024
 
-def get_ssl_feat_shape(mag_level):
+def get_ssl_feat_shape(mag_level, pool=False):
     first_dim = _N_EMBED[mag_level]
     h = int(np.sqrt(first_dim))
+    if pool:
+        h = min(h, 8)
     return (_FIXED_SSL_FEATURE_DIM_1, h, h)
 
 def preprocess_features(feat_array):
@@ -38,35 +40,29 @@ def preprocess_features(feat_array):
     std = feat_array.std(axis=0, keepdims=True)
     feat_array = (feat_array - mean) / (std + 1e-8)
 
-    h = np.sqrt(feat_array.shape[1]).astype(int)
-    feat_array = torch.tensor(feat_array.reshape((-1, h, h)))
-
-    if h > 8:
-        shape = (8, 8)
-        feat_array = F.adaptive_avg_pool2d(feat_array, shape)
-
     return feat_array
 
 class MagnificationConfig(datasets.BuilderConfig):
-    def __init__(self, mag_level=None, ssl_feat_shape=None, data_dir=None, **kwargs):
+    def __init__(self, mag_level=None, ssl_feat_shape_pooled=None, ssl_feat_shape_unpooled=None, data_dir=None, **kwargs):
         super(MagnificationConfig, self).__init__(**kwargs)
         self.mag_level = mag_level
-        self.ssl_feat_shape = ssl_feat_shape
+        self.ssl_feat_shape_pooled = ssl_feat_shape_pooled
+        self.ssl_feat_shape_unpooled = ssl_feat_shape_unpooled
         self.data_dir = data_dir
 
 class TCGADataset(datasets.GeneratorBasedBuilder):
     VERSION = _DATASET_VERSION
-    MAGNIFICATIONS = ["20x", "10x", "5x", "2.5x", "1.25x"]
     BUILDER_CONFIGS = []
     for mag_level_str in _MAG_DICT.keys():
-        feature_shape = get_ssl_feat_shape(mag_level_str)
+
         builder_config_instance = MagnificationConfig(
             name=mag_level_str,
             version=_DATASET_VERSION,
-            description=f"Dataset at {mag_level_str} mag. SSL feature shape: {feature_shape}",
+            description=f"Dataset at {mag_level_str} mag",
             data_dir=mag_level_str,
             mag_level=mag_level_str,
-            ssl_feat_shape=feature_shape
+            ssl_feat_shape_pooled=get_ssl_feat_shape(mag_level_str, pool=True),
+            ssl_feat_shape_unpooled=get_ssl_feat_shape(mag_level_str, pool=False)
         )
         BUILDER_CONFIGS.append(builder_config_instance)
 
@@ -78,10 +74,9 @@ class TCGADataset(datasets.GeneratorBasedBuilder):
             features=datasets.Features(
                 {
                     "image": datasets.Image(),
-                    "ssl_feat": datasets.Array3D(shape=self.config.ssl_feat_shape, dtype="float32"),
-                    "filename_img": datasets.Value("string"),
-                    "filename_feat": datasets.Value("string"),
-                    "mag": datasets.Value("string"),
+                    "ssl_feat": datasets.Array3D(shape=self.config.ssl_feat_shape_pooled, dtype="float32"),
+                    "ssl_feat_unpooled": datasets.Array3D(shape=self.config.ssl_feat_shape_unpooled, dtype="float32"),
+                    "mag": datasets.Value("int32"),
                 }
             ),
             homepage="https://github.com/cvlab-stonybrook/ZoomLDM",
@@ -118,9 +113,21 @@ class TCGADataset(datasets.GeneratorBasedBuilder):
             ssl_feat_data = np.load(feat_path)
             processed_feature = preprocess_features(ssl_feat_data)
 
+            h = np.sqrt(processed_feature.shape[1]).astype(int)
+            feat_array_unpooled = torch.tensor(processed_feature.reshape((-1, h, h)))
+
+            if h > 8:
+                shape = (8, 8)
+                feat_array_pooled = F.adaptive_avg_pool2d(feat_array_unpooled, shape)
+
+            else:
+                feat_array_pooled = feat_array_unpooled
+
+
             yield idx, {
                 "image": str(img_path), 
-                "ssl_feat": processed_feature,
+                "ssl_feat": feat_array_pooled,
+                "ssl_feat_unpooled": feat_array_unpooled,
                 "mag": _MAG_DICT[mag_level],
             }
             idx += 1