def test_extract_parse_actions():
tree = ParentedTree.fromstring('(ROOT (satellite:attribution (text 0)) (nucleus:span (satellite:condition (text 1)) (nucleus:span (nucleus:span (nucleus:same-unit (text 2)) (nucleus:same-unit (satellite:temporal (text 3)) (nucleus:span (text 4)))) (satellite:conclusion (text 5)))))')
# I think the tree above would be for something
# like this silly little example:
# "John said that if Bob bought this excellent book,
# then before the end of next week Bob would finish it,
# and therefore he would be happy."
actions = extract_parse_actions(tree)
num_shifts = len([x for x in actions if x.type == 'S'])
assert num_shifts == 6
assert actions[0].type == 'S'
assert actions[1].type == 'U'
assert actions[1].label == 'satellite:attribution'
assert actions[2].type == 'S'
def test_reconstruct_training_examples():
'''
This code goes through the training data and makes sure
that the actions extracted from the trees can be used to
reconstruct those trees from a list of EDUs.
'''
train_path = 'rst_discourse_tb_edus_TRAINING_TRAIN.json'
with open(train_path) as f:
data = json.load(f)
rst_parser = Parser(max_acts=1, max_states=1, n_best=1)
for doc_dict in data:
tree_orig = ParentedTree.fromstring(doc_dict['rst_tree'])
actions = extract_parse_actions(tree_orig)
tree2 = next(rst_parser.parse(doc_dict,
gold_actions=actions,
make_features=False))['tree']
logging.info('test_reconstruct_training_examples verified tree for {}'.format(doc_dict['path_basename']))
assert tree2 == tree_orig
def main():
import argparse
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('train_file',
help='Path to JSON training file.',
type=argparse.FileType('r'))
parser.add_argument('eval_file',
help='Path to JSON dev or test file for ' +
'tuning/evaluation.',
type=argparse.FileType('r'))
parser.add_argument('model_path',
help='Prefix for the path to where the model should be'
' stored. A suffix with the C value will be added.')
parser.add_argument('-w',
'--working_path',
help='Path to where intermediate files should be ' +
'stored',
default='working')
parser.add_argument('-C',
'--C_values',
help='comma-separated list of model complexity ' +
'parameter settings to evaluate.',
default=','.join([str(2.0**x) for x in range(-4, 5)]))
parser.add_argument('-v',
'--verbose',
help='Print more status information. For every ' +
'additional time this flag is specified, ' +
'output gets more verbose.',
default=0,
action='count')
parser.add_argument('-s',
'--single_process',
action='store_true',
help='Run in a single process for all hyperparameter' +
' grid points, to simplify debugging.')
args = parser.parse_args()
if os.path.exists(args.working_path):
raise IOError("{} already exists. Stopping here to avoid the "
"possibility of overwriting files that are currently "
"being used.".format(args.working_path))
os.makedirs(args.working_path)
parser = Parser(1, 1, 1)
# Convert verbose flag to actually logging level
log_levels = [logging.WARNING, logging.INFO, logging.DEBUG]
log_level = log_levels[min(args.verbose, 2)]
# Make warnings from built-in warnings module get formatted more nicely
logging.captureWarnings(True)
logging.basicConfig(format=('%(asctime)s - %(name)s - %(levelname)s - ' +
'%(message)s'),
level=log_level)
logger = logging.getLogger(__name__)
logger.info('Extracting examples')
train_data = json.load(args.train_file)
eval_data = json.load(args.eval_file)
train_examples = []
for doc_dict in train_data:
path_basename = doc_dict['path_basename']
logging.info('Extracting examples for {}'.format(path_basename))
tree = ParentedTree.fromstring(doc_dict['rst_tree'])
collapse_rst_labels(tree)
actions = extract_parse_actions(tree)
for i, (action_str, feats) in \
enumerate(parser.parse(doc_dict, gold_actions=actions)):
example_id = "{}_{}".format(path_basename, i)
example = {"x": Counter(feats), "y": action_str, "id": example_id}
train_examples.append(example)
# print("{} {}".format(action_str, " ".join(feats)))
# train and evaluate a model for each value of C
best_labeled_f1 = -1.0
best_C = None
# train and evaluate models with different C values in parallel
C_values = [float(x) for x in args.C_values.split(',')]
partial_train_and_eval_model = partial(train_and_eval_model,
args.working_path, args.model_path,
eval_data)
# Make the SKLL jsonlines feature file
train_path = os.path.join(args.working_path, 'rst_parsing.jsonlines')
with open(train_path, 'w') as train_file:
for example in train_examples:
train_file.write('{}\n'.format(json.dumps(example)))
if args.single_process:
all_results = [
partial_train_and_eval_model(C_value) for C_value in C_values
]
else:
n_workers = len(C_values)
with ProcessPoolExecutor(max_workers=n_workers) as executor:
all_results = executor.map(partial_train_and_eval_model, C_values)
for C_value, results in zip(C_values, all_results):
results["C"] = C_value
print(json.dumps(sorted(results.items())))
if results["labeled_f1"] > best_labeled_f1:
best_labeled_f1 = results["labeled_f1"]
best_C = C_value
print("best labeled F1 = {}, with C = {}".format(best_labeled_f1, best_C))
def main():
import argparse
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('train_file',
help='Path to JSON training file.',
type=argparse.FileType('r'))
parser.add_argument('eval_file',
help='Path to JSON dev or test file for ' +
'tuning/evaluation.',
type=argparse.FileType('r'))
parser.add_argument('model_path',
help='Prefix for the path to where the model should be'
' stored. A suffix with the C value will be added.')
parser.add_argument('-w', '--working_path',
help='Path to where intermediate files should be ' +
'stored', default='working')
parser.add_argument('-C', '--C_values',
help='comma-separated list of model complexity ' +
'parameter settings to evaluate.',
default=','.join([str(2.0 ** x)
for x in range(-4, 5)]))
parser.add_argument('-v', '--verbose',
help='Print more status information. For every ' +
'additional time this flag is specified, ' +
'output gets more verbose.',
default=0, action='count')
parser.add_argument('-s', '--single_process', action='store_true',
help='Run in a single process for all hyperparameter' +
' grid points, to simplify debugging.')
args = parser.parse_args()
if os.path.exists(args.working_path):
raise IOError("{} already exists. Stopping here to avoid the "
"possibility of overwriting files that are currently "
"being used.".format(args.working_path))
os.makedirs(args.working_path)
parser = Parser(1, 1, 1)
# Convert verbose flag to actually logging level
log_levels = [logging.WARNING, logging.INFO, logging.DEBUG]
log_level = log_levels[min(args.verbose, 2)]
# Make warnings from built-in warnings module get formatted more nicely
logging.captureWarnings(True)
logging.basicConfig(format=('%(asctime)s - %(name)s - %(levelname)s - ' +
'%(message)s'), level=log_level)
logger = logging.getLogger(__name__)
logger.info('Extracting examples')
train_data = json.load(args.train_file)
eval_data = json.load(args.eval_file)
train_examples = []
for doc_dict in train_data:
path_basename = doc_dict['path_basename']
logging.info('Extracting examples for {}'.format(path_basename))
tree = ParentedTree.fromstring(doc_dict['rst_tree'])
collapse_rst_labels(tree)
actions = extract_parse_actions(tree)
for i, (action_str, feats) in \
enumerate(parser.parse(doc_dict, gold_actions=actions)):
example_id = "{}_{}".format(path_basename, i)
example = {"x": Counter(feats), "y": action_str, "id": example_id}
train_examples.append(example)
# print("{} {}".format(action_str, " ".join(feats)))
# train and evaluate a model for each value of C
best_labeled_f1 = -1.0
best_C = None
# train and evaluate models with different C values in parallel
C_values = [float(x) for x in args.C_values.split(',')]
partial_train_and_eval_model = partial(train_and_eval_model,
args.working_path, args.model_path,
eval_data)
# Make the SKLL jsonlines feature file
train_path = os.path.join(args.working_path, 'rst_parsing.jsonlines')
with open(train_path, 'w') as train_file:
for example in train_examples:
train_file.write('{}\n'.format(json.dumps(example)))
if args.single_process:
all_results = [partial_train_and_eval_model(C_value)
for C_value in C_values]
else:
n_workers = len(C_values)
with ProcessPoolExecutor(max_workers=n_workers) as executor:
all_results = executor.map(partial_train_and_eval_model, C_values)
for C_value, results in zip(C_values, all_results):
results["C"] = C_value
print(json.dumps(sorted(results.items())))
if results["labeled_f1"] > best_labeled_f1:
best_labeled_f1 = results["labeled_f1"]
best_C = C_value
print("best labeled F1 = {}, with C = {}".format(best_labeled_f1, best_C))