Add csv reader

timm/data/readers/reader_factory.py

@@ -3,6 +3,7 @@
 
 from .reader_image_folder import ReaderImageFolder
 from .reader_image_in_tar import ReaderImageInTar
+from .reader_paths_csv import ReaderPathsCsv
 
 
 def create_reader(
@@ -34,6 +35,13 @@ def create_reader(
         from .reader_wds import ReaderWds
         kwargs.pop('download', False)
         reader = ReaderWds(root=root, name=name, split=split, **kwargs)
+    elif "samples_csv_path" in kwargs:
+        assert "class_map" in kwargs
+        reader = ReaderPathsCsv(
+            images_dir=root,
+            samples_csv_path=kwargs["samples_csv_path"], 
+            class_map=kwargs["class_map"],
+        )
     else:
         assert os.path.exists(root)
         # default fallback path (backwards compat), use image tar if root is a .tar file, otherwise image folder

timm/data/readers/reader_paths_csv.py

@@ -0,0 +1,47 @@
+"""
+A dataset reader that extracts images from a single folder
+based on a csv with labels and filenames relative to that folder.
+"""
+import os
+import pandas as pd
+
+from .reader import Reader
+
+
+class ReaderPathsCsv(Reader):
+    def __init__(
+            self,
+            images_dir,
+            samples_csv_path,
+            class_map: dict[str, int],
+    ):
+        super().__init__()
+        assert isinstance(class_map, dict)
+
+        self.images_dir = images_dir
+        samples_df = pd.read_csv(samples_csv_path).astype(str)
+
+        if not samples_df["label"].isin(class_map).all():
+            unrecognized_ids = ~samples_df["label"].isin(class_map)
+            unrecognized_labels = set(samples_df.loc[unrecognized_ids, "label"])
+            raise ValueError(f"Unrecognized labels found in samples_df: {unrecognized_labels}")
+
+        samples_df["label"] = samples_df["label"].map(class_map)
+
+        self.samples_df = samples_df
+
+    def __getitem__(self, index):
+        filename, target = self.samples_df.iloc[index]
+        path = os.path.join(self.images_dir, filename)
+        return open(path, 'rb'), target
+
+    def __len__(self):
+        return len(self.samples_df.index)
+
+    def _filename(self, index, basename=False, absolute=False):
+        filename = self.samples_df.iloc[index, "filename"]
+        if basename:
+            filename = os.path.basename(filename)
+        elif not absolute:
+            filename = os.path.relpath(filename, self.images_dir)
+        return filename

timm/train.py

@@ -93,6 +93,10 @@
                     help='path to dataset (root dir)')
 parser.add_argument('--dataset', metavar='NAME', default='',
                     help='dataset type + name ("<type>/<name>") (default: ImageFolder or ImageTar if empty)')
+parser.add_argument('--train-samples-csv-path', metavar='PATH',
+                    help='path to csv with train filenames and labels')
+parser.add_argument('--val-samples-csv-path', metavar='PATH',
+                    help='path to csv with train filenames and labels')
 group.add_argument('--train-split', metavar='NAME', default='train',
                    help='dataset train split (default: train)')
 group.add_argument('--val-split', metavar='NAME', default='validation',
@@ -669,6 +673,7 @@ def train(config: dict[str, t.Any]):
         input_key=args.input_key,
         target_key=args.target_key,
         num_samples=args.train_num_samples,
+        samples_csv_path=args.train_samples_csv_path,
     )
 
     if args.val_split:
@@ -684,6 +689,7 @@ def train(config: dict[str, t.Any]):
             input_key=args.input_key,
             target_key=args.target_key,
             num_samples=args.val_num_samples,
+            samples_csv_path=args.val_samples_csv_path,
         )
 
     # setup mixup / cutmix