Skip to content

Latest commit

 

History

History
83 lines (59 loc) · 3.98 KB

data-cleaning-tutorial-tips.md

File metadata and controls

83 lines (59 loc) · 3.98 KB
Raw

Data cleaning tutorial tips

Suggested datasets

  1. Imagenette 2 - precomputed results!
  2. Oxford Pets - precomputed results!
  3. PC parts
  4. Caltech 101 - remove background category
  5. 10 big cats of the wild
  6. Try your own between 1'000 and 5'000 images, with resolution > 128 x 128 (subsample if necessary)

Suggested platforms

  • Google Colab is the most stable option. Note that GPU runtimes may not be available on free plans, and they may not be very performant.
  • A virtual environment is likely the best choice for those who have a GPU and experience with training models.

Managing time

  • FastDup can handle up to 100k images in a few minutes. Recommended for everyone as first step, possibly on multiple datasets.
  • We provide models and features for Imagenette 2 and Oxford Pets. You should be able to run these by the end of the workshop even with CPU.
  • SelfClean is the newest tool, and it needs self-supervised pre-training. Since we exceptionally start training from pretrained backbones, it should work reasonably well even with a few epochs. Your best second step for custom datasets after FastDup, but if you do not have a good GPU it may still not finish in time.
  • CleanLab needs to train several supervised models for cross-validation folds, so it is the slowest option. Very interesting for looking at label errors in your own dataset, but it may a good homework unless you have access to great computing resources. P.S. choosing a smaller model than ResNet50 may help with speed.
  • Alternatively, CleanVision is a package by CleanLab which is considerably faster and very easy to use, but with limited features.

Known issues

  • FastDup does not install smoothly on Windows machines.
  • SelfClean will only run with single CPU on Apple ARM machines.

Mounting Google Drive on Colab

This can be useful as datasets do not have to be downloaded again in case the runtime disconnects.

import os
try:
    from google.colab import drive
    mount_point = os.path.join(os.path.sep, "content", "drive")
    drive.mount(mount_point)
except ImportError:
    pass

FastDup

Docs

The docs for FastDup are not so easy to find, so here is the link:

https://visual-layer.readme.io/docs/v1-api

Tricks

fd.run(num_images=N) only runs on N images to reduce runtime

Grayscale images

If your custom dataset has single-channel grayscale images, these need to be converted to color before feeding them into the usual pipelines. Also, according to how your images are normalized, you may need a scaling factor (with integer // or floating point / division) to correctly visualize them. Here is a custom image loader to be used in the ImageFolder class of the torch.datasets library. Finally, take into account that different tools will want channel first or channel last format!

Custom Image Loader

from PIL import Image
SCALING_FACTOR = 256
def image_loader(path: str):
    image = Image.open(path)
    rgb_image = Image.new("RGB", image.size)
    return np.repeat(np.array(image)[:, :, None], 3, axis=-1) // SCALING_FACTOR  # Channel last
    # return np.repeat(np.array(image)[None, :, :], 3, axis=0) // SCALING_FACTOR  # Channel first

In case you get bored...

Is content so far too basic, or models training?

  • Try out different models with FastDup, for instance using a DINOv2 models only requires setting a string option!
  • Request a trial period to play with CleanStudio, the online tool by CleanLab.
  • Have a look at our SelfClean paper to find out more advanced details on cleaning!