Home

This page lists different examples of code usage.

• Default experiment

  Launch the experiment with default parameters:

  python experiment.py

  Default main parameters are:

  • models: VggNet
  • datasets: CIFAR10
  • methods: least confidence sampling
  • Ks: 200
  • number of trainings: 5
  • number of epochs: 10

  Therefore, you will have only one result.

• Experiment several active learning methods

  Launch the experiment with margin and entropy sampling:

  python experiment.py --methods "margin,entropy"

  Main parameters are now:

  • models: VggNet
  • datasets: CIFAR10
  • methods: margin sampling, entropy sampling
  • Ks: 200
  • number of trainings: 5
  • number of epochs: 10

  This will launch 2 active learnings, one per method, so the final plot will contain 2 results.

• Experiment several active learning methods and models

  Launch the experiment with margin and entropy sampling on AlexNet and ResNet:

  python experiment.py --methods "margin,entropy" --models "AlexNet,ResNet"

  Main parameters are now:

  • models: AlexNet, ResNet
  • datasets: CIFAR10
  • methods: margin sampling, entropy sampling
  • Ks: 200
  • number of trainings: 5
  • number of epochs: 10

  This will launch 4 active learnings, one per method and model.

• Experiment several active learning methods, models, and numbers of samples to add to next training set

  Launch the experiment with margin and entropy sampling on AlexNet and ResNet, with k=100 and k=30:

  python experiment.py --methods "margin,entropy" --models "AlexNet,ResNet" --Ks "100,30"

  Main parameters are now:

  • models: AlexNet, ResNet
  • datasets: CIFAR10
  • methods: margin sampling, entropy sampling
  • Ks: 100, 30
  • number of trainings: 5
  • number of epochs: 10

  This will launch 8 active learnings, one per method and model and k.

• Experiment with a reduced pool set

  The image datasets are really big: 50000 images for CIFAR10 pool set (train + validation). So it is possible to truncate the pool set to 2000 images for example with:

  python experiment.py --pool-length 2000

  Note that this reduces the pool set, not the test set.