"""
run from repo root
first arg is base config file
override individual params with -o then comma-separated (no spaces) list of args, e.g., -o exp_name=foo,data.seed=0,model.fair_coeff=2.0
set templates by name, e.g., --data adult
e.g.,
>>> python src/run_unf_clf.py conf/transfer/run_unf_clf.json -o train.n_epochs=5 --data adult --dirs local
"""
from codebase.config import process_config
from transfer_learn import get_repr_filename
from transfer_learn import main as train_transfer_classifier
from run_laftr import correct_repr_exists
import os
transfer_opt = process_config(verbose=False)
training_opt = transfer_opt.copy()
if transfer_opt['transfer']['repr_name'] == 'scratch':
print('training classifier from data directly')
train_transfer_classifier(transfer_opt)
else:
# this experiment gets its own reps and tranfsers
repr_name, transfer_name = os.path.split(transfer_opt['exp_name'])
assert 'transfer' in transfer_name, "expect a sweep over some transfer params"
training_opt.update(exp_name=repr_name)
transfer_opt['transfer'].update(repr_name=repr_name)
training_opt.pop(
'transfer'
) # can discard transfer params for purposes of learning the base rep
assert correct_repr_exists(
training_opt), "coudln't find pre-trained laftr representation :("
# all done
with open(os.path.join(resdirname, 'done.txt'), 'w') as f:
f.write('done')
if __name__ == '__main__':
"""
This script trains a LAFTR model. For the full evaluation used in the paper (first train LAFTR then evaluate on naive classifier) see `src/run_laftr.py`.
Instructions:
1) Run from repo root
2) First arg is base config file
3) Optionally override individual params with -o then comma-separated (no spaces) list of args, e.g., -o exp_name=foo,data.seed=0,model.fair_coeff=2.0
(the overrides must come before the named templates in steps 4--5)
4) Set templates by name, e.g., --data adult
5) Required templates are --data and --dirs
e.g.,
>>> python src/laftr.py conf/laftr/config.json -o train.n_epochs=10,model.fair_coeff=2. --data adult --dirs local
This command trains LAFTR on the Adult dataset for ten epochs with batch size 32.
Model and optimization parameters are specified by the config file conf/classification/config.json
Dataset specifications are read from conf/templates/data/adult.json
Directory specifications are read from conf/templates/dirs/local.json
Finally, two hyperparameters are overridden by the last command.
By using the -o flag we train for 10 epochs with fairness regulariztion coeff 2. instead of the default values from the config.json.
"""
from codebase.config import process_config
opt = process_config(verbose=False)
main(opt)
print('opt_filename', opt_filename)
print('done_filename_txt', done_filename_txt)
print('done_filename_json', done_filename_json)
if os.path.exists(done_filename_txt) or os.path.exists(done_filename_json):
if load_config(opt_filename) == opt:
return True
return False
if __name__ == '__main__':
"""
This script takes LAFTR hyperparameters and trains a LAFTR model if a model with those hyperparameters values hasn't already been trained.
But if it was previously trained (i.e. representations/encoder already exist) then we do nothing.
It is meant to be used as part of a larger experiment sweep.
See `src/laftr.py` for instructions on how to format arguments.
"""
from codebase.config import process_config
from laftr import main as learn_reps
training_opt = process_config(verbose=False)
training_opt.pop(
'transfer'
) # can discard transfer params for purposes of learning the base rep
if correct_repr_exists(training_opt):
print('pre-trained reps exist, NOT TRAINING')
else:
print('learning reps from scratch')
learn_reps(training_opt)