dry_run = True
else:
dry_run = False
cartesian_product = []
for data_type, vals in runs_dic.items():
task_format = vals['task_format']
for split in vals['splits']:
input_file = vals['input_file']
input_file = input_file.format(split)
for encoding_format in vals['encode_format']:
for task in vals['tasks']:
for control in ['false', 'true']:
if control == 'true':
out_dir = vals['out_dir'].format(
encoding_format, task + '_control', split)
else:
out_dir = vals['out_dir'].format(
encoding_format, task, split)
cartesian_product.append([
input_file, f'models/lm/{task}/{encoding_format}/',
out_dir, task_format, encoding_format, control
])
parallelize(nodes,
cartesian_product,
'amnesic_probing/runs/encode/run_layer_encode.sh',
on_gpu=True,
dry_run=dry_run)
if arguments['--dry_run']:
dry_run = True
else:
dry_run = False
cartesian_product = []
for data_type, vals in runs_dic.items():
base_dir = vals['base_dir']
vecs = base_dir + '/' + vals['vecs']
text = base_dir + '/' + vals['text']
task_type = vals['task_type']
for masking in ['normal', 'masked']:
base_dir = vals['base_dir'].format(masking)
vecs = base_dir + '/' + vals['vecs']
text = base_dir + '/' + vals['text']
task_type = vals['task_type']
for label in vals['labels']:
data_label = label
if task_type != 'task':
data_label = 'np_start'
cartesian_product.append([
vecs, f'{base_dir}/{data_label}.pickle', text,
f'models/lm/{label}/{masking}/layer:last/', task_type
])
parallelize(nodes,
cartesian_product,
'amnesic_probing/runs/evaluate/run_eval_per_dim.sh',
on_gpu=True,
dry_run=dry_run)
arguments = docopt(__doc__)
if arguments['--dry_run']:
dry_run = True
else:
dry_run = False
cartesian_product = []
for data_type, vals in runs_dic.items():
for masking in ['normal', 'masked']:
layer_dir = vals['layers_dir'].format(masking)
for task in vals['labels']:
base_dir = vals['base_dir'].format(masking, task)
text_file = f'{base_dir}/tokens.pickle'
task_type = vals['task_type']
# using the task labels in the regular case, or random labels from the same data
# when the task is generated on the fly (e.g. word_len)
data_label = task
if task_type != 'task':
data_label = 'np_start'
cartesian_product.append([layer_dir,
base_dir,
f'{base_dir}/{data_label}.pickle',
text_file,
task_type])
parallelize(nodes, cartesian_product,
'amnesic_probing/runs/evaluate/run_layer_wise_deprobe.sh',
on_gpu=False, dry_run=dry_run)
base_dir = vals['base_dir'].format(masking)
vecs = base_dir + '/' + vals['vecs']
for task in vals['task']:
for label in vals['labels']:
output_dir = 'models/lm/{0}/{1}/layer:{2}/'
if task == 'task':
task_name = label
else:
task_name = task
# running over all the model layers
if 'layers' in vals:
for layer in vals['layers']:
cartesian_product.append([
vecs.format(layer),
f'{base_dir}/{label}.pickle',
output_dir.format(task_name, masking,
layer), task, balanced
], )
else:
cartesian_product.append([
vecs, f'{base_dir}/{label}.pickle',
output_dir.format(task_name, masking, 'last'),
task, balanced
])
parallelize(nodes,
cartesian_product,
'amnesic_probing/runs/core/run_deprobe.sh',
on_gpu=False,
dry_run=dry_run)
'dep_p', 'dep_poss', 'dep_adp', 'dep_amod', 'dep_nsubj', 'dep_dep',
'dep_dobj', 'dep_cc', 'dep_conj', 'dep_advmod', 'dep_ROOT',
'dep_ccomp', 'dep_aux', 'dep_xcomp', 'dep_neg'
],
},
}
if __name__ == '__main__':
arguments = docopt(__doc__)
if arguments['--dry_run']:
dry_run = True
else:
dry_run = False
cartesian_product = []
for data_type, vals in runs_dic.items():
for label in vals['labels']:
for masking in ['normal', 'masked']:
train_dir = vals['train_dir'].format(masking)
cartesian_product.append([
train_dir, f'models/lm/{label}/{masking}/layer:last/',
f'models/lm/{label}/{masking}/layer:last/P.npy'
])
parallelize(nodes,
cartesian_product,
'amnesic_probing/runs/core/run_ft_reg.sh',
on_gpu=True,
dry_run=dry_run)