NVIDIA/DALI

ExternalSource and fn.readers.file

Open

#3,811 opened on Apr 11, 2022

View on GitHub
 (9 comments) (0 reactions) (1 assignee)C++ (670 forks)auto 404
external sourcehelp wanted

Repository metrics

Stars
 (5,722 stars)
PR merge metrics
 (PR metrics pending)

Description

  1. CodeBase1: I need to use the weighted random sampler for ImageNet1000, so, i checke this issue. link. Then, I followed this link to get my own codes, including Sampler, ExternalSourcePipeline and DALIGenericIterator.

  2. CodeBase2: Also, I check this ImageNet1000-resnet link.

However, the difference in efficiency of loading data is surprisingly large!!!

I'm not sure. The possible reason may be the difference between fn.readers.file in link and np.fromfile in link

for _ in range(self.batch_size):
    jpeg_filename, label = self.files[self.i % self.n].split(' ')
    batch.append(np.fromfile(self.images_dir + jpeg_filename, dtype = np.uint8))  # we can use numpy
    labels.append(torch.tensor([int(label)], dtype = torch.uint8)) # or PyTorch's native tensors
    self.i += 1

How should I achieve an efficiency similar to CodeBase2 with weighted random sampler.

Contributor guide

Research direction
Compare the data loading performance between ExternalSource with np.fromfile and fn.readers.file in DALI. Investigate the impact of batch size, prefetch queue depth, and use of external source iterators. Check if the weighted random sampler introduces overhead and explore ways to integrate it efficiently with DALI's pipeline.
Tech stack
pythonpytorch
Domain
backendmachine learning
Issue type
Performance
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-3 hours
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Mostly clear
Prerequisites
PythonPyTorchDALI
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
40