def main():
usage = "%prog results_file"
parser = OptionParser(usage=usage)
(options, args) = parser.parse_args()
results_file = args[0]
results = pd.DataFrame(columns=('masked', 'test', 'valid', 'dir'))
lines = fh.read_text(results_file)
for i, line in enumerate(lines):
if i > 0:
parts = line.split()
date = parts[0]
time = parts[1]
name = parts[2]
masked = float(parts[3])
test = float(parts[4])
name_parts = name.split('_')
run_num = int(name_parts[-1])
if run_num < 41:
if test > 0:
valid = parts[5][1:-1]
else:
valid = masked
#results.loc[run_num, 'iteration'] = run_num
results.loc[run_num, 'masked'] = masked
results.loc[run_num, 'test'] = test
results.loc[run_num, 'valid'] = valid
results.loc[run_num, 'dir'] = name
results.to_csv(results_file + 'results.csv', columns=results.columns)
sorted = results.sort('masked')
print sorted
print "best by masked"
print sorted.values[-1, :]
print "best by valid"
sorted = results.sort('valid')
print sorted.values[-1, :]
def parse_xml_output(xml_filename):
raw_xml = fh.read_text(xml_filename)
sentences = []
sentiments = []
dependencies = []
dependency_tuples = []
entities = []
coref = []
groups = []
detailed_groups = []
coref_indices = []
coref_heads = []
line_index = 0
sentence_id = 0
while line_index < len(raw_xml):
line = raw_xml[line_index].lstrip()
match = re.search('<sentence id="(\d+)" line="\d+" sentimentValue="(\d+)" sentiment="(.*)"', line)
if match is not None:
line_index, sentence, sent_dependencies, sent_dependency_tuples,\
sent_entities = parse_sentence(raw_xml, line_index)
sentences.append(sentence)
sentiments.append([match.group(3)])
dependencies.append(sent_dependencies)
dependency_tuples.append(sent_dependency_tuples)
entities.append(sent_entities)
if re.search('<coreference>', line) is not None:
line_index, coref, groups, detailed_groups, coref_indices,\
coref_heads = parse_coref(raw_xml, line_index, len(sentences))
line_index += 1
# deal with the cases where there are no entities
if len(coref) == 0:
for s in sentences:
coref.append([])
return sentences, sentiments, dependencies, dependency_tuples, entities, coref, groups, detailed_groups,\
coref_indices, coref_heads
def test_over_time(project_dir, subset, config_file, model_type, field, train_start, train_end, test_start, test_end, n_train=None, n_calib=0, penalty='l2', suffix='', loss='log', objective='f1', do_ensemble=True, dh=300, label='label', intercept=True, n_dev_folds=5, average='micro', seed=None, alpha_min=0.01, alpha_max=1000.0, n_alphas=8, sample_labels=False, group_identical=False, annotated_subset=None, nonlinearity='tanh', init_lr=1e-2, min_epochs=2, max_epochs=50, patience=5, tol=1e-4, list_size=1, repeats=1, oracle=False, lower=None, interactive=False, stoplist_file=None, cshift=False, n_cshift=None, do_cfm=True, do_platt=True, dropout=0.0, min_test=None, test_prop=None, verbose=False):
# Just run a regular model, one per year, training on the past, and save the reults
if seed is not None:
seed = int(seed)
np.random.seed(seed)
log = {
'project': project_dir,
'subset': subset,
'config_file': config_file,
'model_type': model_type,
'field': field,
'train_start': train_start,
'train_end': train_end,
'test_start': test_start,
'test_end': test_end,
'n_train': n_train,
'n_calib': n_calib,
'penalty': penalty,
'cshift': cshift,
'n_cshift': n_cshift,
'suffix': suffix,
'loss': loss,
'objective': objective,
'do_ensemble': do_ensemble,
'dh': dh,
'label': label,
'intercept': intercept,
'n_dev_folds': n_dev_folds,
'average': average,
'seed': seed,
'alpha_min': alpha_min,
'alpha_max': alpha_max,
'n_alphas': n_alphas,
'sample_labels': sample_labels,
'group_identical': group_identical,
'annotated_subset': annotated_subset,
'nonlinearity': nonlinearity,
'init_lr': init_lr,
'min_epochs': min_epochs,
'max_epochs': max_epochs,
'patience': patience,
'tol': tol,
'interactive': interactive,
'stoplist_file': stoplist_file,
'list_size': list_size
}
model_basename = make_model_basename(log)
# save the experiment parameters to a log file
logfile = os.path.join(dirs.dir_logs(project_dir), model_basename + '.json')
fh.makedirs(dirs.dir_logs(project_dir))
fh.write_to_json(log, logfile)
# load the features specified in the config file
config = fh.read_json(config_file)
feature_defs = []
for f in config['feature_defs']:
feature_defs.append(features.parse_feature_string(f))
# load the file that contains metadata about each item
metadata_file = os.path.join(dirs.dir_subset(project_dir, subset), 'metadata.csv')
metadata = fh.read_csv_to_df(metadata_file)
field_vals = list(set(metadata[field].values))
field_vals.sort()
print("Splitting data according to %s", field)
print("Values:", field_vals)
print("\nTesting on %s to %s" % (test_start, test_end))
# first, split into training and non-train data based on the field of interest
all_items = list(metadata.index)
test_selector_all = (metadata[field] >= int(test_start)) & (metadata[field] <= int(test_end))
test_subset_all = metadata[test_selector_all]
test_items_all = test_subset_all.index.tolist()
n_test_all = len(test_items_all)
if min_test is not None:
if n_test_all < min_test:
print("Not enough test samples; exiting")
return
if train_end is None:
if train_start is None:
train_selector_all = metadata[field] < int(test_start)
else:
train_selector_all = (metadata[field] < int(test_start)) & (metadata[field] >= train_start)
else:
if train_start is None:
train_selector_all = metadata[field] <= int(train_end)
else:
train_selector_all = (metadata[field] <= int(train_end)) & (metadata[field] >= train_start)
train_subset_all = metadata[train_selector_all]
train_items_all = list(train_subset_all.index)
n_train_all = len(train_items_all)
# only keep the items in the train and test sets
all_items = train_items_all + test_items_all
print("Train: %d, Test: %d (labeled and unlabeled)" % (n_train_all, n_test_all))
# load all labels
label_dir = dirs.dir_labels(project_dir, subset)
labels_df = fh.read_csv_to_df(os.path.join(label_dir, label + '.csv'), index_col=0, header=0)
labels_df = labels_df.loc[all_items]
# if desired, attempt to learn weights for the training data using techniques for covariate shift
if cshift:
print("Training a classifier for covariate shift")
# start by learning to discriminate train from non-train data
# Label items based on whether they come from train or test
train_test_labels = np.zeros((len(all_items), 2), dtype=int)
train_test_labels[:n_train_all, 0] = 1
train_test_labels[n_train_all:, 1] = 1
if np.sum(train_test_labels[:, 0]) < np.sum(train_test_labels[:, 1]):
cshift_pos_label = 0
else:
cshift_pos_label = 1
train_test_labels_df = pd.DataFrame(train_test_labels, index=all_items, columns=[0, 1])
if n_cshift is not None and len(all_items) >= n_cshift:
print("Taking a random sample of %d items for reweighting" % n_cshift)
#np.random.shuffle(all_items)
cshift_items = np.random.choice(all_items, size=n_cshift, replace=False)
else:
print("Using all train items")
cshift_items = all_items
print(train_test_labels_df.loc[cshift_items].mean(axis=0))
# create a cshift model using the same specifiction as our model below (e.g. LR/MLP, etc.)
model_name = model_basename + '_' + str(test_start) + '-' + str(test_end) + 'cshift'
model, dev_f1, dev_acc, dev_cal, dev_cal_overall = train.train_model_with_labels(project_dir, model_type, loss, model_name, subset, train_test_labels_df, feature_defs, items_to_use=cshift_items, penalty=penalty, alpha_min=alpha_min, alpha_max=alpha_max, n_alphas=n_alphas, intercept=intercept, n_dev_folds=n_dev_folds, save_model=True, do_ensemble=False, dh=dh, seed=seed, pos_label=cshift_pos_label, verbose=False)
print("cshift results: %0.4f f1, %0.4f acc" % (dev_f1, dev_acc))
#X_cshift, features_concat = predict.load_data(project_dir, model_name, subset, items_to_use=all_items)
X_cshift, features_concat = predict.load_data(project_dir, model_name, subset, items_to_use=all_items)
cshift_pred_probs = model.predict_probs(X_cshift)
f_items = features_concat.get_items()
assert len(f_items) == len(all_items)
for i in range(len(all_items)):
assert all_items[i] == f_items[i]
cshift_pred_probs_df = pd.DataFrame(cshift_pred_probs, index=features_concat.get_items(), columns=range(2))
# display the min and max probs
print("Min: %0.6f" % cshift_pred_probs_df[1].values[:n_train_all].min())
print("Mean: %0.6f" % cshift_pred_probs_df[1].values[:n_train_all].mean())
print("Max: %0.6f" % cshift_pred_probs_df[1].values[:n_train_all].max())
# HACK: need to prevent 0s in prob(y=0|x)
p_train_values = cshift_pred_probs_df[0].values
threshold = 0.01
p_train_values[p_train_values < threshold] = threshold
print("After thresholding")
print("Min: %0.6f" % p_train_values[:n_train_all].min())
print("Mean: %0.6f" % p_train_values[:n_train_all].mean())
print("Max: %0.6f" % p_train_values[:n_train_all].max())
# use the estimated probability of each item being a training item to compute item weights
weights = n_train_all / float(n_test_all) * (1.0/p_train_values - 1)
weights_df_all = pd.DataFrame(weights, index=all_items)
# print a summary of the weights from just the training items
print("Min weight: %0.4f" % weights[:n_train_all].min())
print("Ave weight: %0.4f" % weights[:n_train_all].mean())
print("Max weight: %0.4f" % weights[:n_train_all].max())
# print a summary of all weights
#print("Min weight: %0.4f" % weights.min())
#print("Ave weight: %0.4f" % weights.mean())
#print("Max weight: %0.4f" % weights.max())
# create a data frame with this information
else:
weights_df_all = None
# find the labeled items
print("Subsetting items with labels")
label_sums_df = labels_df.sum(axis=1)
labeled_item_selector = label_sums_df > 0
labels_df = labels_df[labeled_item_selector]
n_labeled_items, n_classes = labels_df.shape
print("%d labeled items" % n_labeled_items)
labeled_items = set(labels_df.index)
train_items_labeled = [i for i in train_items_all if i in labeled_items]
test_items = [i for i in test_items_all if i in labeled_items]
#n_train = len(train_items)
n_test = len(test_items)
for r in range(repeats):
# set seed very explicily here to make sure experiments are comparable
if seed is not None:
seed += 1
np.random.seed(seed)
print("* Starting repetition %d *" % r)
model_name = model_basename + '_' + str(test_start) + '-' + str(test_end) + '_' + str(r)
if n_train is not None and len(train_items_labeled) >= n_train:
np.random.shuffle(train_items_labeled)
train_items = np.random.choice(train_items_labeled, size=n_train, replace=False)
else:
print("Using all train items")
train_items = train_items_labeled
n_train_r = len(train_items)
# now, choose a calibration set
if n_calib > 0 and n_test >= n_calib:
np.random.shuffle(test_items)
calib_items = np.random.choice(test_items, size=n_calib, replace=False)
elif n_test < n_calib:
print("Error: Only %d labeled test instances available" % n_test)
calib_items = test_items
else:
calib_items = []
if weights_df_all is not None:
weights_df = weights_df_all[labeled_item_selector]
else:
weights_df = None
print("Labeled train: %d, test: %d" % (n_train_r, n_test))
# create a data frame to hold a summary of the results
output_df = pd.DataFrame([], columns=['N', 'training data', 'test data', 'cal', 'estimate', 'MAE', '95lcl', '95ucl', 'contains_test'])
test_labels_df = labels_df.loc[test_items]
# do a fake adjustment of the test label proportions
if test_prop is not None:
test_prop = float(test_prop)
test_label_values = test_labels_df.values
test_label_props = test_label_values[:, 1] / (test_label_values[:, 1] + test_label_values[:, 0])
order = list(np.argsort(test_label_props))
true_prop = np.mean(test_label_props)
if test_prop < true_prop:
i = 0
running = test_label_props[order[i]]
new_test_items = [test_items[order[i]]]
i += 1
while (running / i) <= test_prop:
running += test_label_props[order[i]]
new_test_items.append(test_items[order[i]])
i += 1
print("Taking %d test_items" % len(new_test_items))
test_items = new_test_items[:]
else:
order.reverse()
i = 0
running = test_label_props[order[i]]
new_test_items = [test_items[order[i]]]
i += 1
while (running / i) >= test_prop:
running += test_label_props[order[i]]
new_test_items.append(test_items[order[i]])
i += 1
print("Taking %d test_items" % len(new_test_items))
test_items = new_test_items[:]
test_labels_df = labels_df.loc[test_items]
test_label_values = test_labels_df.values
test_label_props = test_label_values[:, 1] / (test_label_values[:, 1] + test_label_values[:, 0])
print("New props = %0.3f" % np.mean(test_label_props))
# if instructed, sample labels in proportion to annotations (to simulate having one label per item)
if sample_labels:
print("Sampling labels")
# normalize the labels
temp = labels_df.values / np.array(labels_df.values.sum(axis=1).reshape((n_labeled_items, 1)), dtype=float)
samples = np.zeros([n_labeled_items, n_classes], dtype=int)
for i in range(n_labeled_items):
index = np.random.choice(np.arange(n_classes), size=1, p=temp[i, :])
samples[i, index] = 1
sampled_labels_df = pd.DataFrame(samples, index=labels_df.index, columns=labels_df.columns)
else:
sampled_labels_df = labels_df
train_labels_df = sampled_labels_df.loc[train_items].copy()
if n_calib > 0:
calib_labels_df = sampled_labels_df.loc[calib_items].copy()
else:
calib_labels_df = None
# get the true proportion of labels in the test OR non-training data (calibration and test combined)
target_props, target_estimate, target_std = get_estimate_and_std(test_labels_df, use_n_annotations=True)
output_df.loc['target'] = [n_test, 'test', 'test', 'n/a', target_estimate, 0, target_estimate - 2 * target_std, target_estimate + 2 * target_std, np.nan]
# get the same estimate from training data
train_props, train_estimate, train_std = get_estimate_and_std(train_labels_df, use_n_annotations=True)
print("Train props:", train_props, train_estimate)
train_rmse = np.abs(train_estimate - target_estimate)
train_contains_test = target_estimate > train_estimate - 2 * train_std and target_estimate < train_estimate + 2 * train_std
output_df.loc['train'] = [n_train_r, 'train', 'test', 'n/a', train_estimate, train_rmse, train_estimate - 2 * train_std, train_estimate + 2 * train_std, train_contains_test]
# get the same estimate from training data
if n_calib > 0:
calib_props, calib_estimate, calib_std = get_estimate_and_std(calib_labels_df, use_n_annotations=True)
# compute the error of this estimate
calib_rmse = np.abs(calib_estimate - target_estimate)
calib_contains_test = target_estimate > calib_estimate - 2 * calib_std and target_estimate < calib_estimate + 2 * calib_std
output_df.loc['calib'] = [n_calib, 'calib', 'test', 'n/a', calib_estimate, calib_rmse, calib_estimate - 2 * calib_std, calib_estimate + 2 * calib_std, calib_contains_test]
else:
calib_estimate = 0.0
calib_std = 1.0
output_df.loc['calib'] = [n_calib, 'calib', 'test', 'n/a', np.nan, np.nan, np.nan, np.nan, np.nan]
if train_estimate > 0.5:
pos_label = 0
else:
pos_label = 1
print("Using %d as the positive label" % pos_label)
results_df = pd.DataFrame([], columns=['f1', 'acc', 'mae', 'estimated calibration'])
# Now train a model on the training data, saving the calibration data for calibration
if stoplist_file is not None:
stoplist = fh.read_text(stoplist_file)
stoplist = {s.strip() for s in stoplist}
print(stoplist)
else:
stoplist = None
print("Training a LR model")
model, dev_f1, dev_acc, dev_cal_mae, dev_cal_est = train.train_model_with_labels(project_dir, model_type, 'log', model_name, subset, sampled_labels_df, feature_defs, weights_df=weights_df, items_to_use=train_items, penalty=penalty, alpha_min=alpha_min, alpha_max=alpha_max, n_alphas=n_alphas, intercept=intercept, objective=objective, n_dev_folds=n_dev_folds, do_ensemble=do_ensemble, dh=dh, seed=seed, pos_label=pos_label, vocab=None, group_identical=group_identical, nonlinearity=nonlinearity, init_lr=init_lr, min_epochs=min_epochs, max_epochs=max_epochs, patience=patience, do_cfm=do_cfm, do_platt=do_platt, lower=lower, stoplist=stoplist, dropout=dropout, verbose=verbose)
results_df.loc['cross_val'] = [dev_f1, dev_acc, dev_cal_mae, dev_cal_est]
X_test, features_concat = predict.load_data(project_dir, model_name, subset, items_to_use=test_items)
test_predictions = model.predict(X_test)
test_predictions_df = pd.DataFrame(test_predictions, index=features_concat.get_items(), columns=[label])
test_pred_probs = model.predict_probs(X_test)
_, n_labels = test_pred_probs.shape
test_pred_probs_df = pd.DataFrame(test_pred_probs, index=features_concat.get_items(), columns=range(n_labels))
f1_test, acc_test = evaluate_predictions.evaluate_predictions(test_labels_df, test_predictions_df, test_pred_probs_df, pos_label=pos_label, average=average)
true_test_vector = np.argmax(test_labels_df.as_matrix(), axis=1)
test_cal_est = evaluation.evaluate_calibration_rmse(true_test_vector, test_pred_probs_df.as_matrix(), min_bins=1, max_bins=1)
test_cc_estimate, test_pcc_estimate = model.predict_proportions(X_test)
test_cc_mae = np.mean(np.abs(test_cc_estimate[1] - target_estimate))
test_pcc_mae = np.mean(np.abs(test_pcc_estimate[1] - target_estimate))
results_df.loc['test'] = [f1_test, acc_test, test_pcc_mae, test_cal_est]
output_df.loc['CC'] = [n_train_r, 'train', 'test', 'n/a', test_cc_estimate[1], test_cc_mae, np.nan, np.nan, np.nan]
output_df.loc['PCC'] = [n_train_r, 'train', 'test', 'n/a', test_pcc_estimate[1], test_pcc_mae, np.nan, np.nan, np.nan]
test_acc_estimate_internal, test_acc_ms_estimate_internal = model.predict_proportions(X_test, do_cfm=do_cfm)
test_acc_rmse_internal = np.abs(test_acc_estimate_internal[1] - target_estimate)
test_acc_ms_rmse_internal = np.abs(test_acc_ms_estimate_internal[1] - target_estimate)
output_df.loc['ACC_internal'] = [n_train_r, 'train', 'test', 'n/a', test_acc_estimate_internal[1], test_acc_rmse_internal, np.nan, np.nan, np.nan]
output_df.loc['MS_internal'] = [n_train_r, 'train', 'nontrain', 'predicted', test_acc_ms_estimate_internal[1], test_acc_ms_rmse_internal, np.nan, np.nan, np.nan]
test_platt1_estimate, test_platt2_estimate = model.predict_proportions(X_test, do_platt=do_platt)
test_platt1_rmse = np.abs(test_platt1_estimate[1] - target_estimate)
test_platt2_rmse = np.abs(test_platt2_estimate[1] - target_estimate)
output_df.loc['PCC_platt1'] = [n_train_r, 'train', 'test', 'n/a', test_platt1_estimate[1], test_platt1_rmse, np.nan, np.nan, np.nan]
output_df.loc['PCC_platt2'] = [n_train_r, 'train', 'nontrain', 'predicted', test_platt2_estimate[1], test_platt2_rmse, np.nan, np.nan, np.nan]
if n_calib > 0:
cc_plus_cal_estimate = (test_cc_estimate[1] + calib_estimate) / 2.0
pcc_plus_cal_estimate = (test_pcc_estimate[1] + calib_estimate) / 2.0
cc_plus_cal_mae = np.mean(np.abs(cc_plus_cal_estimate - target_estimate))
pcc_plus_cal_mae = np.mean(np.abs(pcc_plus_cal_estimate - target_estimate))
#output_df.loc['CC_plus_cal'] = [n_train, 'train', 'test', 'n/a', cc_plus_cal_estimate, cc_plus_cal_mae, np.nan, np.nan, np.nan]
output_df.loc['PCC_plus_cal'] = [n_train_r, 'train', 'test', 'n/a', pcc_plus_cal_estimate, pcc_plus_cal_mae, np.nan, np.nan, np.nan]
results_df.to_csv(os.path.join(dirs.dir_models(project_dir), model_name, 'accuracy.csv'))
output_df.to_csv(os.path.join(dirs.dir_models(project_dir), model_name, 'results.csv'))
"""
def parse_xml_files(xml_filelist_filename, output_dir):
filelist = fh.read_text(xml_filelist_filename)
parsed_files = {}
sentiments = {}
dependencies = {}
dependency_tuples = {}
entities = {}
coref = {}
coref_entities = {}
coref_heads = {}
all_groups = {}
jk_grams = {}
amalgram_pairs = {}
for file in filelist:
file = file.rstrip('\n')
print file
# peel off both .txt and .xml
basename = fh.get_basename_wo_ext(fh.get_basename_wo_ext(file))
sentences, doc_sentiments, doc_dependencies, doc_dependency_tuples, doc_entities, doc_coref, groups, _,\
doc_coref_entities, doc_coref_heads = parse_xml_output(file)
parsed_files[basename] = sentences
sentiments[basename] = doc_sentiments
dependencies[basename] = doc_dependencies
dependency_tuples[basename] = doc_dependency_tuples
entities[basename] = doc_entities
coref[basename] = doc_coref
coref_entities[basename] = doc_coref_entities
coref_heads[basename] = doc_coref_heads
doc_jk_grams, doc_jk_indices = find_jk_grams(sentences)
jk_grams[basename] = doc_jk_grams
# output documents to amalgram format
#amalgram_dir = os.path.join(dirs.data_amalgram_dir, 'input')
#if not os.path.exists(amalgram_dir):
# os.makedirs(amalgram_dir)
tagged_sents = ['\n'.join([t['word'] + '\t' + t['POS'] for t in s]) + '\n' for s in sentences]
# save word/tag pairs for amalgram
tagged_sents = [[(t['word'], t['POS']) for t in s] for s in sentences]
amalgram_pairs[basename] = tagged_sents
# uncomment for extracting story elements...
parsed_dir = os.path.join(output_dir, 'parsed')
if not os.path.exists(parsed_dir):
os.makedirs(parsed_dir)
parsed_filename = os.path.join(parsed_dir, basename + '.json')
fh.write_to_json(sentences, parsed_filename, sort_keys=False)
sentiment_filename = fh.make_filename(output_dir, 'sentiments', 'json')
fh.write_to_json(sentiments, sentiment_filename, sort_keys=False)
dependencies_filename = fh.make_filename(output_dir, 'dependency_tuple_ids', 'json')
fh.write_to_json(dependency_tuples, dependencies_filename, sort_keys=False)
coref_filename = fh.make_filename(output_dir, 'entities', 'json')
fh.write_to_json(coref, coref_filename, sort_keys=False)
jkgrams_filename = fh.make_filename(output_dir, 'jkgrams', 'json')
fh.write_to_json(jk_grams, jkgrams_filename, sort_keys=False)
coref_heads_filename = fh.make_filename(output_dir, 'coref_heads', 'json')
fh.write_to_json(coref_heads, coref_heads_filename, sort_keys=False)
amalgram_keys = amalgram_pairs.keys()
amalgram_keys.sort()
amalgram_data_file = os.path.join(dirs.data_amalgram_dir, 'input.txt')
with codecs.open(amalgram_data_file, 'w', encoding='utf-8') as output_file:
for k in amalgram_keys:
sents = amalgram_pairs[k]
for s in sents:
for p in s:
output_file.write(p[0] + '\t' + p[1] + '\n')
output_file.write('\n')
for k in amalgram_keys:
amalgram_data_file = os.path.join(dirs.data_amalgram_dir, k + '.txt')
with codecs.open(amalgram_data_file, 'w', encoding='utf-8') as output_file:
sents = amalgram_pairs[k]
for s in sents:
for p in s:
output_file.write(p[0] + '\t' + p[1] + '\n')
output_file.write('\n')
amalgram_index_file = os.path.join(dirs.data_amalgram_dir, 'index.txt')
with codecs.open(amalgram_index_file, 'w', encoding='utf-8') as output_file:
for k in amalgram_keys:
sents = amalgram_pairs[k]
for s in sents:
output_file.write(k + '\n')
#all_groups_filename = fh.make_filename(output_dir, 'all_groups', 'json')
#fh.write_to_json(all_groups, all_groups_filename)
return parsed_files, dependencies
