Explanation of the RepeatFactorTrainingSampler.repeat_factors_from_category_frequency(`repeat_thresh`)

📚 RepeatFactorTrainingSampler.repeat_factors_from_category_frequency

From the docs repeat_factors_from_category_frequency

repeat_thresh (float) – frequency threshold below which data is repeated. If the frequency is half of repeat_thresh, the image will be repeated twice.

In the source code i find these lines:

# 2. For each category c, compute the category-level repeat factor:
#    r(c) = max(1, sqrt(t / f(c)))

Now if f(c) = frequency = 0.5 and t = repeat_thresh = 1 then r(c) = 1.41

Can someone explain the docstring “If the frequency is half of repeat_thresh, the image will be repeated twice.” to me? Based on the example above i would expect that every image in c to be repeated 1.41 times, not 2.0 as the doc suggests.

[chawins]

Thanks for the answer! In my case, I created a custom RepeatFactorTrainingSampler just without the sqrt().

[marijnl]

I still dont really understand the docstring but this could be due to my limited understanding of the sampling method. The implementation references LVIS paper appendix B2 which gives a more in depth description of Mask R-CNN with Data Resampling.

If you want you can make the RepeatFactorTrainingSampler balanced by calculating repeat_factors yourself. Maybe thats of any help

import torch
from detectron2.data.samplers.distributed_sampler import RepeatFactorTrainingSampler

num_iterations = 10000
num_images = 10000
class_distribution = [0.8, 0.2]
images_classes = (torch.rand(num_images) > class_distribution[0])*1
# dont use in production! this likely ignores images due to stochastic rounding. Repeat factors should be >= 1
image_level_repeat_factors = torch.tensor([1-class_distribution[class_index] for class_index in images_classes])

sampler = RepeatFactorTrainingSampler(repeat_factors=image_level_repeat_factors)
sampled = []
for i, sample in enumerate(sampler):
    sampled.append(images_classes[sample])
    if i == num_iterations:
        break

print("original distribution", torch.histc(images_classes*1.0, 2)/num_images) # tensor([0.7970, 0.2030])
print("sampled distribution", torch.histc(torch.tensor(sampled)*1.0, 2) / num_iterations) #  tensor([0.4986, 0.5015])