[BUG] Datasets do not pass `min_tiles` filter when training GANs

[siddhir]

Description

GAN training io issue likely because it takes in the filters dictionary but not the .filter() function when constructing this initial dataset for io. Filtering after this step doesn’t seem to work and neither does including min_tiles=21 within the initial P.dataset(….min_tiles=21, filters={….}) call. Given I couldn’t figure out a way to include the min_tiles within the dictionary, I basically duplicated the annotation file and included another filter called “MIN_TILE” which excluded all WSIs with <21 extracted tiles.

To Reproduce

Steps to reproduce the behavior:

1. Use min_tiles argument when constructing any dataset for GAN

def main(): P = sf.Project('/mnt/data/PROJECTS/Title') P.annotations = './qc_plan.csv' full_dataset = P.dataset( tile_px=299, tile_um=302, filters={'site': ['XXXXXXXXXXX], 'progressor':[XXXXX], 'QC_Sid': ['QC_PASS'], }, min_tiles=21 )

train_df = full_dataset.filter({'site': [XXXXXXX]}) 

P.gan_train(
    dataset=train_df,
    model='stylegan3',
    cfg='stylegan2',
    gpus=4,
    batch=128,
    kimg=10000,
    resize=256,
    tile_labels= "label",
    exp_label="label"

)

if name == 'main': mp.freeze_support() main()

Expected behavior

Should take in the min_tiles argument when constructing dataset

Environment:

• Slideflow Version (e.g., 1.0):
• OS (e.g., Ubuntu):
• How you installed Slideflow (pip, source):
• Python version:
• CUDA/cuDNN version:
• GPU models and configuration:
• Any other relevant information:

Additional context

[jamesdolezal]

This should be fixed in the latest dev update (f939885). Give it a try and see if it works - as a reminder, you'll need to update both the main slideflow repository as well as the GAN submodules:

git checkout dev 
git pull
git submodule update --recursive

[jamesdolezal]

For planning - this will be included in the 2.2 release. I can't put it in the next patch release due to somewhat complex dependencies on the StyleGAN submodules which are being updated for 2.2.

[jamesdolezal]

Released in 2.2 - thanks!