0.01

data/dataset.py

@@ -1,5 +1,9 @@
 from scipy.io import loadmat
 import torch
+import lmdb
+import os
+import pickle
+from io import BytesIO
 from torch.utils.data import Dataset
 from torchvision.datasets.folder import default_loader
 from torchvision.datasets import ImageFolder
@@ -22,7 +26,7 @@ class CARS(Dataset):
         return len(self.annotations)
 
     def __getitem__(self, item):
-        file_name = "{:05d}.jpg".format(item+1)
+        file_name = "{:05d}.jpg".format(item + 1)
         target = self.annotations[file_name]
         sample = self.loader(self.root / "cars_train" / file_name)
         if self.transform is not None:
@@ -41,6 +45,22 @@ class ImprovedImageFolder(ImageFolder):
         return super().__getitem__(item)[0]
 
 
+class LMDBDataset(Dataset):
+    def __init__(self, lmdb_path):
+        self.db = lmdb.open(lmdb_path, subdir=os.path.isdir(lmdb_path), readonly=True, lock=False,
+                            readahead=False, meminit=False)
+        with self.db.begin(write=False) as txn:
+            self.classes_list = pickle.loads(txn.get(b"classes_list"))
+            self._len = pickle.loads(txn.get(b"__len__"))
+
+    def __len__(self):
+        return self._len
+
+    def __getitem__(self, i):
+        with self.db.begin(write=False) as txn:
+            return torch.load(BytesIO(txn.get("{}".format(i).encode())))
+
+
 class EpisodicDataset(Dataset):
     def __init__(self, origin_dataset, num_class, num_set, num_episodes):
         self.origin = origin_dataset
@@ -60,12 +80,12 @@ class EpisodicDataset(Dataset):
         for i, c in enumerate(random_classes):
             image_list = self.origin.classes_list[c]
             if len(image_list) > self.num_set * 2:
-                idx_list = torch.randperm(len(image_list))[:self.num_set*2].tolist()
+                idx_list = torch.randperm(len(image_list))[:self.num_set * 2].tolist()
             else:
-                idx_list = torch.randint(high=len(image_list), size=(self.num_set*2,)).tolist()
+                idx_list = torch.randint(high=len(image_list), size=(self.num_set * 2,)).tolist()
             support_set_list.extend([self.origin[idx] for idx in idx_list[:self.num_set]])
             query_set_list.extend([self.origin[idx] for idx in idx_list[self.num_set:]])
-            target_list.extend([i]*self.num_set)
+            target_list.extend([i] * self.num_set)
         return {
             "support": torch.stack(support_set_list),
             "query": torch.stack(query_set_list),

data/lmdbify.py

@@ -0,0 +1,35 @@
+import torch
+import lmdb
+import os
+import pickle
+from io import BytesIO
+from data.dataset import CARS, ImprovedImageFolder
+import torchvision
+from tqdm import tqdm
+
+
+def dataset_to_lmdb(dataset, lmdb_path):
+    env = lmdb.open(lmdb_path, map_size=1099511627776 * 2, subdir=os.path.isdir(lmdb_path))
+    with env.begin(write=True) as txn:
+        for i in tqdm(range(len(dataset))):
+            buffer = BytesIO()
+            torch.save(dataset[i], buffer)
+            txn.put("{}".format(i).encode(), buffer.getvalue())
+        txn.put(b"classes_list", pickle.dumps(dataset.classes_list))
+        txn.put(b"__len__", pickle.dumps(len(dataset)))
+
+
+def main():
+    data_transform = torchvision.transforms.Compose([
+        torchvision.transforms.Resize([int(224 * 1.15), int(224 * 1.15)]),
+        torchvision.transforms.CenterCrop(224),
+        torchvision.transforms.ToTensor(),
+        torchvision.transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
+    ])
+    origin_dataset = ImprovedImageFolder("/data/few-shot/CUB_200_2011/CUB_200_2011/images", transform=data_transform)
+    dataset_to_lmdb(origin_dataset, "/data/few-shot/lmdb/CUB_200_2011/data.lmdb")
+
+
+if __name__ == '__main__':
+    main()
+