diff --git a/timm/data/loader.py b/timm/data/loader.py
index 99cf132f..9d5d51cc 100644
--- a/timm/data/loader.py
+++ b/timm/data/loader.py
@@ -5,6 +5,7 @@ https://github.com/NVIDIA/apex/commit/d5e2bb4bdeedd27b1dfaf5bb2b24d6c000dee9be#d
 
 Hacked together by / Copyright 2020 Ross Wightman
 """
+import warnings
 
 import torch.utils.data
 import numpy as np
@@ -245,11 +246,25 @@ class MultiEpochsDataLoader(torch.utils.data.DataLoader):
         self._DataLoader__initialized = True
         self.iterator = super().__iter__()
 
+        self._items_to_consume = 0
+        self._last_consumed_item = -1
+
     def __len__(self):
         return len(self.batch_sampler.sampler)
 
     def __iter__(self):
-        for i in range(len(self)):
+        # If a previous call to this function didn't consume all the corresponding items, the next call it's gonna
+        # continue the previous one, which is incorrect. So we consume the remaining items.
+        # It's inefficient but I can't come up with a better way to do this.
+        # Probably avoiding this class completely is more efficient if you're skipping a bunch of items in every epoch.
+        for _ in range(self._items_to_consume - 1 - self._last_consumed_item):
+            warnings.warn("Consuming the rest of the sampler items from a previous call. "
+                          "Consider not using MultiEpochsDataLoader as it may take a lot of time.")
+            next(self.iterator)
+
+        self._items_to_consume = len(self)
+        for i in range(self._items_to_consume):
+            self._last_consumed_item = i
             yield next(self.iterator)