# Confusing with the Impact of \`num\_workers\` on Image Agumentation Speed with PIL.Image and Tensor

**URL:** <https://discuss.pytorch.org/t/confusing-with-the-impact-of-num-workers-on-image-agumentation-speed-with-pil-image-and-tensor/199993>\
**Category:** data\
**Created:** [April 1, 2024, 4:16pm UTC](https://discuss.pytorch.org/t/confusing-with-the-impact-of-num-workers-on-image-agumentation-speed-with-pil-image-and-tensor/199993 "2024-04-01T16:16:55Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![HELLORPG](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/hellorpg/32/53892_2.png) [@HELLORPG](https://discuss.pytorch.org/u/HELLORPG)\
**Post date:** [April 1, 2024, 4:16pm UTC](https://discuss.pytorch.org/t/confusing-with-the-impact-of-num-workers-on-image-agumentation-speed-with-pil-image-and-tensor/199993/1 "2024-04-01T16:16:55Z")

</div>

In my code, I need to augment a clip (image sequence) during an iteration. I try to use `torch.Tensor` as my data type because the [doc](https://pytorch.org/vision/stable/transforms.html#performance-considerations) recommends I do so. However, I found it may lead negative impact on speed.

Here is my demo code, using the newest PyTorch version:

```python
from torch.utils.data import DataLoader, Dataset
import torch
from PIL import Image
from torchvision.transforms import v2
import copy
import random
import time

class MyDatasets(Dataset):
    def __init__ (self, data_type="tensor"):
        self.data_type = data_type

    def __len__ (self):
        return 50

    def __getitem__ (self, index):
        if self.data_type == "tensor":
            _ = torch.randn(3, 800, 1440)
            clip = _[None, ...].repeat(40, 1, 1, 1)
        else:
            _ = Image.new("RGB", (1440, 800))
            clip = []
            for i in range(40):
                clip.append(copy.deepcopy(_))
        # # Stack:
        # if isinstance(clip, torch.Tensor):
        # clip = torch
        # Horizontal flip:
        if isinstance(clip, torch.Tensor):
            clip = v2.functional.hflip(clip)
        else:
            clip = [v2.functional.hflip(_) for _ in clip]
        # Resize:
        if isinstance(clip, torch.Tensor):
            clip = v2.functional.resize(clip, [720, 1280])
        else:
            clip = [v2.functional.resize(_, [720, 1280]) for _ in clip]
        # Crop:
        if isinstance(clip, torch.Tensor):
            clip = v2.functional.crop(clip, 50, 50, 480, 800)
        else:
            clip = [v2.functional.crop(_, 50, 50, 480, 800) for _ in clip]
        # To Tensor
        if isinstance(clip, list):
            clip = [v2.functional.to_tensor(_) for _ in clip]
            clip = torch.stack(clip, dim=0)
        return clip

if __name__ == ' __main__':
    dataset = MyDatasets("tensor")
    dataloader = DataLoader(dataset, batch_size=1, num_workers=4)
    begin_time = time.time()
    for data in enumerate(dataloader):
        # print(data)
        pass
    print("Cost: ", time.time() - begin_time)
    # For PIL:
    dataset = MyDatasets("pil")
    dataloader = DataLoader(dataset, batch_size=1, num_workers=4)
    begin_time = time.time()
    for data in enumerate(dataloader):
        # print(data)
        pass
    print("Cost: ", time.time() - begin_time)

```

Run this script, I’ll get the outputs following:

```bash
Tensor Cost: 9.852455615997314
PIL Cost: 7.140314817428589

```

More complex augmentations may cause a larger gap.

However, if I set `num_workers=0` in my data loader, I’ll get:

```bash
Tensor Cost: 11.182294607162476
PIL Cost: 24.352734088897705

```

I totally do not understand this phenomenon. Could anyone here help me understand why this happens?  
Thanks a lot.

---

<div class="post-metadata">

**Author:** ![HELLORPG](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/hellorpg/32/53892_2.png) [@HELLORPG](https://discuss.pytorch.org/u/HELLORPG)\
**Post date:** [April 1, 2024, 4:33pm UTC](https://discuss.pytorch.org/t/confusing-with-the-impact-of-num-workers-on-image-agumentation-speed-with-pil-image-and-tensor/199993/2 "2024-04-01T16:33:40Z")

</div>

Are `torchvision.transforms` with `torch.Tensor` aggressive against multiprocessing?