How can we run training? about clipseg HOT 8 CLOSED

Thanks for the question! Training is implemented in train_loop in experiment_setup.py with the training configuration being specified in experiments/phrasecut.yaml.
We plan to describe training in more detail later.

from clipseg.

BinaryIoU is missing in metrics.py

from clipseg.

@soskek Thanks for pointing this out. We will address these issues when we update the codebase.

from clipseg.

Thank you!
I’m looking forward to using the full codebase!

from clipseg.

I wrote some missing features for reminders.

e.g. madpack (?) may be missing.

https://github.com/timojl/clipseg/blob/master/datasets/phrasecut.py#L2

from madpack.transforms.spatial import random_crop_slices

from clipseg.

TrainingLogger in general_utils.py seems broken or old.

https://github.com/timojl/clipseg/blob/master/experiment_setup.py#L162

    with TrainingLogger(log_dir=config.name, interval=150, model=model, config=tracker_config, utilization_iters=500) as logger:

caused

TypeError: __init__() got an unexpected keyword argument 'interval'

and

AttributeError: __enter__

from clipseg.

@timojl Could you provide training command to reproduce results ?

from clipseg.

The code base has been updated. Training and evaluation should be straightforward now and the problems pointed out by @soskek (thanks, again) should be resolved.

from clipseg.