Why multiple threads are spawned even with num_workers=0 in pytorch dataloader?

data
1

Versions:
torch: 1.6.0
torchvision: 0.7.0

Created a custom dataloader to apply custom augmentations on the image and transform it to a tensor. Even without enabling custom augmentations and setting num_workers=0 in the Dataloader object (used default way as well by not setting anything), it is still consuming around 15-20 threads (not consuming the whole thread but 9% of it). I am not sure what is causing this issue as I want to keep the resources free to be used by compute intensive processes of other users.

Tried:

  1. Setting OMP and other env variables to restrict the use of threads, using os.environ module in Python
  2. Tried setting combinations of num_workers and pin_memory params
  3. Optimized the code to avoid copying ops to CPU and using numpy for calculation of metrics there.
  4. I had installed torch using conda at the time of env creation but installed torch using pip due to some issues later. Can that cause an issue?

@ptrblck: Can you help me out in understanding as to what is actually causing this or does the API has some internal operations which are segregated over the threads, by default?!

2

I would check what operations are actually using multiple threads—you could do this by moving the dataloader operations outside of a dataloader and seeing if they are still using multiple threads. This would help rule out whether the problem is with the dataloader or the operations it is performing.

3 
self.transform = transforms.Compose([transforms.ToTensor()])
image = self.transform(image) 
label = torch.tensor(0) if class_=='normal' else torch.tensor(1)
return image, label

This is the only operation getting performed inside the dataloader.

4

Note that ToTensor() can call a few operations under the hood such as rounding, dtype conversion, and memory layout transformation, so I would check if any of those individual operations show as using multiple threads on your system:
https://pytorch.org/vision/main/_modules/torchvision/transforms/functional.html#to_tensor