Dataset setup for reported Resnet Benchmarks

[afzalxo]

Hi, I am trying to understand the image serialization parameters utilized to generate the dataset files for which the accuracy is reported on the rn18 and rn50 here (https://github.com/libffcv/ffcv-imagenet). The given example uses the following notation:

# Serialize images with:
# - 500px side length maximum
# - 50% JPEG encoded
# - quality=90 JPEGs
./write_imagenet.sh 500 0.50 90

And the reported dataset setup parameters are listed as

• ResNet-50 training: 50% JPEG 500px side length
• ResNet-18 training: 10% JPEG 400px side length

This means, for example, that for resnet-50, 50% of the dataset is JPEG encoded, remaining raw (i.e., write_mode=proportion). If this is the case, what is the JPEG quality used for encoding (i.e., last input parameter given to write_imagenet.sh file)? Is it 90 as shown in the example?