diff --git a/nerfstudio/data/datamanagers/full_images_datamanager.py b/nerfstudio/data/datamanagers/full_images_datamanager.py index 9c37ce619..868064a7f 100644 --- a/nerfstudio/data/datamanagers/full_images_datamanager.py +++ b/nerfstudio/data/datamanagers/full_images_datamanager.py @@ -315,7 +315,7 @@ class FullImageDatamanager(DataManager, Generic[TDataset]): Pretends to be the dataloader for evaluation, it returns a list of (camera, data) tuples """ image_indices = [i for i in range(len(self.eval_dataset))] - data = deepcopy(self.cached_eval) + data = [d.copy() for d in self.cached_eval] _cameras = deepcopy(self.eval_dataset.cameras).to(self.device) cameras = [] for i in image_indices: @@ -350,6 +350,9 @@ class FullImageDatamanager(DataManager, Generic[TDataset]): self.train_unseen_cameras = self.sample_train_cameras() data = self.cached_train[image_idx] + # We're going to copy to make sure we don't mutate the cached dictionary. + # This can cause a memory leak: https://github.com/nerfstudio-project/nerfstudio/issues/3335 + data = data.copy() data["image"] = data["image"].to(self.device) assert len(self.train_cameras.shape) == 1, "Assumes single batch dimension" @@ -375,7 +378,8 @@ class FullImageDatamanager(DataManager, Generic[TDataset]): # Make sure to re-populate the unseen cameras list if we have exhausted it if len(self.eval_unseen_cameras) == 0: self.eval_unseen_cameras = [i for i in range(len(self.eval_dataset))] - data = deepcopy(self.cached_eval[image_idx]) + data = self.cached_eval[image_idx] + data = data.copy() data["image"] = data["image"].to(self.device) assert len(self.eval_dataset.cameras.shape) == 1, "Assumes single batch dimension" camera = self.eval_dataset.cameras[image_idx : image_idx + 1].to(self.device)