Reduce GPU memory usage in CacheDataloader (#1730)

* Reduce GPU memory usage in CacheDataloader

* Move also masks and seg.masks, depths to cpu

* Fix linter issues
This commit is contained in:
Jonáš Kulhánek authored and GitHub committed 2023-06-07 14:14:24 -04:00
1 parent d4b04376ab
commit 3b0f7581d8
10 files changed
+29 -11

No files matched your search

@@ -389,10 +389,11 @@ class VanillaDataManager(DataManager, Generic[TDataset]):
self.config.data = self.config.dataparser.data
self.dataparser = self.dataparser_config.setup()
self.includes_time = self.dataparser.includes_time
self.train_dataparser_outputs = self.dataparser.get_dataparser_outputs(split="train")
self.train_dataparser_outputs: DataparserOutputs = self.dataparser.get_dataparser_outputs(split="train")
self.train_dataset = self.create_train_dataset()
self.eval_dataset = self.create_eval_dataset()
self.exclude_batch_keys_from_device = self.train_dataset.exclude_batch_keys_from_device
if self.train_dataparser_outputs is not None:
cameras = self.train_dataparser_outputs.cameras
@@ -452,6 +453,7 @@ class VanillaDataManager(DataManager, Generic[TDataset]):
num_workers=self.world_size * 4,
pin_memory=True,
collate_fn=self.config.collate_fn,
exclude_batch_keys_from_device=self.exclude_batch_keys_from_device,
)
self.iter_train_image_dataloader = iter(self.train_image_dataloader)
self.train_pixel_sampler = self._get_pixel_sampler(self.train_dataset, self.config.train_num_rays_per_batch)
@@ -475,6 +477,7 @@ class VanillaDataManager(DataManager, Generic[TDataset]):
num_workers=self.world_size * 4,
pin_memory=True,
collate_fn=self.config.collate_fn,
exclude_batch_keys_from_device=self.exclude_batch_keys_from_device,
)
self.iter_eval_image_dataloader = iter(self.eval_image_dataloader)
self.eval_pixel_sampler = self._get_pixel_sampler(self.eval_dataset, self.config.eval_num_rays_per_batch)
+1
View File
@@ -42,6 +42,7 @@ class InputDataset(Dataset):
scale_factor: The scaling factor for the dataparser outputs
"""
exclude_batch_keys_from_device: List[str] = ["image", "mask"]
cameras: Cameras
def __init__(self, dataparser_outputs: DataparserOutputs, scale_factor: float = 1.0):
@@ -31,6 +31,8 @@ class DepthDataset(InputDataset):
scale_factor: The scaling factor for the dataparser outputs.
"""
exclude_batch_keys_from_device = InputDataset.exclude_batch_keys_from_device + ["depth_image"]
def __init__(self, dataparser_outputs: DataparserOutputs, scale_factor: float = 1.0):
super().__init__(dataparser_outputs, scale_factor)
assert (
+2
View File
@@ -34,6 +34,8 @@ class SDFDataset(InputDataset):
scale_factor: The scaling factor for the dataparser outputs.
"""
exclude_batch_keys_from_device = InputDataset.exclude_batch_keys_from_device + ["depth", "normal"]
def __init__(self, dataparser_outputs: DataparserOutputs, scale_factor: float = 1.0):
super().__init__(dataparser_outputs, scale_factor)
@@ -32,6 +32,8 @@ class SemanticDataset(InputDataset):
dataparser_outputs: description of where and how to read input images.
"""
exclude_batch_keys_from_device = InputDataset.exclude_batch_keys_from_device + ["mask", "semantics"]
def __init__(self, dataparser_outputs: DataparserOutputs, scale_factor: float = 1.0):
super().__init__(dataparser_outputs, scale_factor)
assert "semantics" in dataparser_outputs.metadata.keys() and isinstance(self.metadata["semantics"], Semantics)
+8 -2
View File
@@ -21,7 +21,7 @@ import concurrent.futures
import multiprocessing
import random
from abc import abstractmethod
from typing import Any, Callable, Dict, Optional, Sized, Tuple, Union
from typing import Any, Callable, Dict, List, Optional, Sized, Tuple, Union
import torch
from rich.progress import track
@@ -55,8 +55,11 @@ class CacheDataloader(DataLoader):
num_times_to_repeat_images: int = -1,
device: Union[torch.device, str] = "cpu",
collate_fn: Callable[[Any], Any] = nerfstudio_collate,
exclude_batch_keys_from_device: Optional[List[str]] = None,
**kwargs,
):
if exclude_batch_keys_from_device is None:
exclude_batch_keys_from_device = ["image"]
self.dataset = dataset
assert isinstance(self.dataset, Sized)
@@ -67,6 +70,7 @@ class CacheDataloader(DataLoader):
self.device = device
self.collate_fn = collate_fn
self.num_workers = kwargs.get("num_workers", 0)
self.exclude_batch_keys_from_device = exclude_batch_keys_from_device
self.num_repeated = self.num_times_to_repeat_images # starting value
self.first_time = True
@@ -118,7 +122,9 @@ class CacheDataloader(DataLoader):
"""Returns a collated batch."""
batch_list = self._get_batch_list()
collated_batch = self.collate_fn(batch_list)
collated_batch = get_dict_to_torch(collated_batch, device=self.device, exclude=["image"])
collated_batch = get_dict_to_torch(
collated_batch, device=self.device, exclude=self.exclude_batch_keys_from_device
)
return collated_batch
def __iter__(self):