def main(args):
# things with paths
expdname = args['dirs']['exp_dir']
expname = args['exp_name']
logdname = args['dirs']['log_dir']
resdirname = os.path.join(expdname, expname)
#logdirname = os.path.join(logdname, expname)
logdirname = resdirname
make_dir_if_not_exist(logdirname, remove=True)
make_dir_if_not_exist(resdirname)
npzfile = os.path.join(args['dirs']['data_dir'],
args['data']['name'],
get_npz_basename(**args['data']))
# write params (including all overrides) to experiment directory
with open(os.path.join(resdirname, 'opt.json'), 'w') as f:
opt_dumps = dumps(args, indent=4, sort_keys=True)
f.write(opt_dumps)
if args['data']['use_attr']:
args['model'].update(xdim=args['model']['xdim']+1)
#get dataset
data = Dataset(npzfile=npzfile, **args['data'], batch_size=args['train']['batch_size'])
# get model
if 'Weighted' in args['model']['class']:
A_weights = [1. / x for x in data.get_A_proportions()]
Y_weights = [1. / x for x in data.get_Y_proportions()]
AY_weights = [[1. / x for x in L] for L in data.get_AY_proportions()]
if 'Eqopp' in args['model']['class']:
#we only care about ppl with Y = 0 --- those who didn't get sick
AY_weights[0][1] = 0. #AY_weights[0][1]
AY_weights[1][1] = 0. #AY_weights[1][1]
args['model'].update(A_weights=A_weights, Y_weights=Y_weights, AY_weights=AY_weights)
model_class = getattr(models, args['model'].pop('class'))
print(args['model'])
model = model_class(**args['model'], batch_size=args['train']['batch_size'])
with tf.Session() as sess:
reslogger = ResultLogger(resdirname)
#create Trainer
trainer = Trainer(model, data, sess=sess, expdir=resdirname, logs_path=logdirname,
**args['optim'], **args['train'])
# training
trainer.train(**args['train'])
# test the trained model
tester = Tester(model, data, sess, reslogger)
tester.evaluate(args['train']['batch_size'])
# flush
tf.reset_default_graph()
# all done
with open(os.path.join(resdirname, 'done.txt'), 'w') as f:
f.write('done')
def load_dirs_config(args, sweep_name=None):
dirs_config_path = os.path.join(CONFIG_DIR, 'dirs', args['dirconf'])
dirs = json.load(open(dirs_config_path, 'r'))
if not sweep_name is None:
dirs['exp'] = os.path.join(dirs['exp'], sweep_name)
dirs['log'] = os.path.join(dirs['log'], sweep_name)
for nm in dirs: make_dir_if_not_exist(dirs[nm])
return dirs
def main(fnames, seed, save_dict):
np.random.seed(seed)
make_dir_if_not_exist(fnames['data_dir_out'])
npzname_out = os.path.join(fnames['data_dir_out'], fnames['data_fout_name'])
n = save_dict['X'].shape[0]
train_inds, valid_inds, test_inds = get_split_inds(n, test_pct=0.3, valid_pct=0.2)
split_save_dict = {}
for inds, phase in [(train_inds, 'train'), (valid_inds, 'valid'), (test_inds, 'test')]:
for t in save_dict:
split_save_dict['{}_{}'.format(t, phase)] = save_dict[t][inds]
save_tensors(split_save_dict, npzname_out)
def main(args, dirs, data_kwargs, model_kwargs):
#get dataset
data = Dataset(**data_kwargs)
print('Dataset loaded from {}.'.format(dirs['data']))
#adaptively set xdim in model
model_kwargs['xdim'] = data.get_data_dim()
#get model
if args['model'] == 'BinaryCFMLP':
model = BinaryCFMLP(**model_kwargs)
elif args['model'] == 'BinaryCFDoubleMLP':
model = BinaryCFDoubleMLP(**model_kwargs)
else:
raise Exception('bad model name')
print('Model loaded.')
with tf.Session() as sess:
print('Session created.')
resdirname = os.path.join(dirs['exp'], args['name'])
logdirname = os.path.join(dirs['log'], args['name'], 'tb_log')
ckptdirname = os.path.join(resdirname, 'checkpoints')
for d in [resdirname, logdirname, ckptdirname]:
make_dir_if_not_exist(d)
#create Trainer
trainer = Trainer(model, data, batch_size=args['batch_size'], sess=sess, logs_path=logdirname, \
checkpoint_path=ckptdirname, results_path=resdirname)
save_path = trainer.train(n_epochs=args['num_epochs'],
patience=args['patience'])
trainer.restore(save_path)
trainer.test()
#save args
args['resdir'] = resdirname
args_path = os.path.join(resdirname, 'args.json')
json.dump(args, open(args_path, 'w'))
def __init__(self, dname, saver=None):
self.dname = dname
make_dir_if_not_exist(self.dname)
self.ckptdir = os.path.join(self.dname, 'checkpoints')
make_dir_if_not_exist(self.ckptdir)
self.npzdir = os.path.join(self.dname, 'npz')
make_dir_if_not_exist(self.npzdir)
self.saver = saver if not saver is None else tf.train.Saver()
self.testcsv_name = os.path.join(self.dname, 'test_metrics.csv')
self.testcsv = open(self.testcsv_name, 'w')
parser = argparse.ArgumentParser(
description='Plot decompositions through training')
parser.add_argument('-n', '--name', help='experiment name', default='')
parser.add_argument('-dirs',
'--dirconf',
help='config file for dirs',
default='madras-vector.json')
parser.add_argument('-fd', '--figdir', help='dir for figs', default='figs')
args = vars(parser.parse_args())
# get params
dirs = load_dirs_config(args)
expdir = os.path.join(dirs['exp'], args['name'])
figdir = os.path.join(args['figdir'], args['name'])
make_dir_if_not_exist(figdir)
base_groups = {
'treatment-shift': ['L', 'L_do', 'L_t_shift'],
'value': ['V', 'V_sample', 'V_treat', 'V_star'],
'regret': [
'V_regret_databias', 'V_regret_fnlearn', 'V_regret_fnlearn_unbiased',
'V_regret_databias_optfn'
]
}
a0_groups = {
'A0-{}'.format(k): ['A0_{}'.format(m) for m in base_groups[k]]
for k in base_groups
}
a1_groups = {
'A1-{}'.format(k): ['A1_{}'.format(m) for m in base_groups[k]]
def main(args):
args = format_transfer_args_as_train_args(args)
# things with paths
expdname = args['dirs']['exp_dir']
expname = args['exp_name']
logdname = args['dirs']['log_dir']
resdirname = os.path.join(expdname, expname)
make_dir_if_not_exist(resdirname)
logdirname = os.path.join(logdname, expname)
make_dir_if_not_exist(logdirname)
npzfile = os.path.join(args['dirs']['data_dir'], args['data']['name'],
get_npz_basename(**args['data']))
# write params (including all overrides) to experiment directory
with open(os.path.join(resdirname, 'done.txt'), 'w') as f:
f.write('done')
npzdname, _ = get_repr_filename(args)
repr_phase = args['transfer'][
'repr_phase'] # this will be "Test" or "Valid"
y_indices = args['transfer']['y_indices'] if hasattr(
args['transfer']['y_indices'],
'__iter__') else [args['transfer']['y_indices']]
if args['transfer']['repr_name'] == 'default':
base_data = Dataset(npzfile=npzfile,
**args['data'],
batch_size=args['train']['batch_size'])
if repr_phase == 'Test':
reprs = base_data.x_test
attrs = base_data.attr_test
y = base_data.y_test
elif repr_phase == 'Valid':
reprs = base_data.x_valid
attrs = base_data.attr_valid
y = base_data.y_valid
else:
# load reprs from LAFTR training as input
repr_fname = os.path.join(npzdname, 'Z.npz')
repr_dat = np.load(repr_fname)
reprs = repr_dat['X']
attr_fname = os.path.join(npzdname, 'A.npz')
attr_dat = np.load(attr_fname)
attrs = attr_dat['X']
y_fname = os.path.join(npzdname, 'Y.npz')
y_dat = np.load(y_fname)
y = y_dat['X']
print('shapes', reprs.shape, attrs.shape, y.shape)
for label_index in y_indices: # this will either be a list of ints or just 'a'
data = TransferDataset(reprs,
attrs,
label_index,
npzfile=npzfile,
Y_loaded=y,
phase=repr_phase,
**args['data'],
batch_size=args['train']['batch_size'])
model = RegularizedFairClassifier(**args['model'])
with tf.Session() as sess:
reslogger = ResultLogger(resdirname)
# create Trainer
trainer = Trainer(model,
data,
sess=sess,
expdir=resdirname,
logs_path=logdirname,
**args['optim'],
**args['train'])
trainer.train(**args['train']) #, train_metric=train_metric)
# test model
tester = Tester(model, data, sess, reslogger)
tester.evaluate(args['train']['batch_size'])
# flush
tf.reset_default_graph()
# all done
with open(os.path.join(resdirname, 'done.json'), 'w') as f:
opt_dumps = dumps(args, indent=4, sort_keys=True)
f.write(opt_dumps)