def run_calculate_weighting_command( self ): target_file_path, message = PySide.QtGui.QFileDialog( ).getSaveFileName( None, 'Save Location' ) data, message = methods.determine_weighting( self.transform, root_bone = self.root_bone, tolerance = self.tolerance ) if not data: self.warning( message ) self.file_path_field.setText( 'Data invalid' ) return False methods.save_json( target_file_path, data ) if os.path.exists( target_file_path ): self.file_path = target_file_path self.file_path_field.setText( target_file_path ) return True
def run_skin_calculation( transform, root_bone, tolerance = -1, file_path = '' ): """ Main method call for the skin converter to calculate skin weights from the transform's shape node. Returns [ data, message ] Data : skin calculation data Message : message about the results. If data == False, message holds more info. """ data, message = methods.determine_weighting( transform, root_bone, tolerance = tolerance ) # Data is valid, so if a file path was provided too, save the data out. if file_path and data: methods.save_json( file_path, data ) return data, message
def annotate_data(source_file, annotations_file, tag=False, compounds=False):
"Function to annotate existing coco data"
data = load_json(source_file)
for entry in data:
raw_description = entry['caption']
doc = nlp.tokenizer(raw_description)
entry['tokenized'] = [tok.orth_ for tok in doc]
if tag:
# Call the tagger on the document.
nlp.tagger(doc)
entry['tagged'] = [(tok.orth_, tok.tag_) for tok in doc]
if compounds:
list_of_compounds = compounds_from_doc(doc)
entry['compounds'] = list_of_compounds
save_json(data, annotations_file)
return data
def run_calculate_weighting_command(self):
target_file_path, message = PySide.QtGui.QFileDialog().getSaveFileName(
None, 'Save Location')
data, message = methods.determine_weighting(self.transform,
root_bone=self.root_bone,
tolerance=self.tolerance)
if not data:
self.warning(message)
self.file_path_field.setText('Data invalid')
return False
methods.save_json(target_file_path, data)
if os.path.exists(target_file_path):
self.file_path = target_file_path
self.file_path_field.setText(target_file_path)
return True
def run_all(args):
"Run all metrics on the data and save JSON files with the results."
# Annotate generated data.
annotated = annotate_data(args.source_file,
args.annotations_file,
tag=True,
compounds=True)
# Load training data. (For computing novelty.)
train_data = load_json('./Data/COCO/Processed/tokenized_train2014.json')
train_descriptions = [
entry['caption'] for entry in train_data['annotations']
]
# Load annotated data.
sentences = sentences_from_file(args.annotations_file)
# Analyze the data.
stats = system_stats(sentences)
# Get raw descriptions.
gen_descriptions = [
entry['caption'] for entry in load_json(args.source_file)
]
extra_stats = sentence_stats(train_descriptions, gen_descriptions)
stats.update(extra_stats)
# Save statistics data.
save_json(stats, args.stats_file)
################################
# Global recall
train_stats = load_json('./Data/COCO/Processed/train_stats.json')
val_stats = load_json('./Data/COCO/Processed/val_stats.json')
train = set(train_stats['types'])
val = set(val_stats['types'])
learnable = train & val
gen = set(stats['types'])
recalled = gen & val
coverage = {
"recalled": recalled,
"score": len(recalled) / len(learnable),
"not_in_val": gen - learnable
}
coverage['omissions'] = most_frequent_omissions(
coverage['recalled'],
val_stats, # Use validation set as reference.
n=None)
val_count_list = get_count_list(val_stats)
coverage['percentiles'] = percentiles(val_count_list, recalled)
save_json(coverage, args.global_coverage_file)
####################################
# Local recall
val_index = index_from_file('./Data/COCO/Processed/tagged_val2014.json',
tagged=True,
lower=True)
generated = {entry['image_id']: entry['tokenized'] for entry in annotated}
local_recall_res = dict(scores=local_recall_scores(generated, val_index),
counts=local_recall_counts(generated, val_index))
save_json(local_recall_res, args.local_coverage_file)
##################################
# Nouns pps
npdata = {
'pp_data': pp_stats(annotated),
'compound_data': compound_stats(annotated)
}
save_json(npdata, args.noun_pp_file)
for entry in coverage_results.values():
entry['omissions'] = most_frequent_omissions(
entry['recalled'],
val_stats, # Use validation set as reference.
n=None) # Rank everything
# Add percentile scores.
val_count_list = get_count_list(val_stats)
for entry in coverage_results.values():
recalled = entry['recalled']
entry['percentiles'] = percentiles(val_count_list, recalled)
plot_percentiles(coverage_results)
# Save the data
save_json(coverage_results, './Data/Output/global_recall.json')
# Show a table with the results.
table = tabulate(tabular_data=[
(system, entry['score']) for system, entry in coverage_results.items()
],
headers=['System', 'Coverage'],
tablefmt='latex_booktabs',
floatfmt='.2f')
print(table)
with open('./Data/Output/global_recall_table.txt', 'w') as f:
f.write(table)
f.write('\n\n')
f.write(
f'The limit is: {limit}. This means {size_limit} words in Val cannot be learned.'
'Shetty-et-al-2016': 'Shetty et al. 2016',
'Shetty-et-al-2017': 'Shetty et al. 2017',
'Tavakoli-et-al-2017': 'Tavakoli et al. 2017',
'Vinyals-et-al-2017': 'Vinyals et al. 2017',
'Wu-et-al-2016': 'Wu et al. 2016',
'Zhou-et-al-2017': 'Zhou et al. 2017'
}
system2color = dict(zip(sorted(system2label), my_palette))
val_index = index_from_file('./Data/COCO/Processed/tagged_val2014.json',
tagged=True,
lower=True)
systems = [
'Dai-et-al-2017', 'Liu-et-al-2017', 'Mun-et-al-2017',
'Shetty-et-al-2016', 'Shetty-et-al-2017', 'Tavakoli-et-al-2017',
'Vinyals-et-al-2017', 'Wu-et-al-2016', 'Zhou-et-al-2017'
]
all_results = dict()
for system in systems:
print('Processing:', system)
generated = name_to_mapping(system)
system_results = dict(scores=local_recall_scores(generated, val_index),
counts=local_recall_counts(generated, val_index))
all_results[system] = system_results
plot_scores(all_results)
save_json(all_results, './Data/Output/local_recall.json')
train_data = load_json('./Data/COCO/Processed/tokenized_train2014.json')
train_descriptions = [entry['caption'] for entry in train_data['annotations']]
for folder in [
'Dai-et-al-2017', 'Liu-et-al-2017', 'Mun-et-al-2017',
'Shetty-et-al-2016', 'Shetty-et-al-2017', 'Tavakoli-et-al-2017',
'Vinyals-et-al-2017', 'Wu-et-al-2016', 'Zhou-et-al-2017'
]:
print('Processing:', folder)
# Define source and target.
base = './Data/Systems/'
source = base + folder + '/Val/annotated.json'
target = base + folder + '/Val/stats.json'
# Load data.
sentences = sentences_from_file(source)
# Process data.
stats = system_stats(sentences)
# Get raw descriptions.
gen_descriptions = [entry['caption'] for entry in load_json(source)]
extra_stats = sentence_stats(train_descriptions, gen_descriptions)
stats.update(extra_stats)
# Save data.
save_json(stats, target)
from methods import parallel_sentences_from_file, parallel_stats, load_json, save_json, sentence_stats
train = parallel_sentences_from_file('./Data/COCO/Processed/tokenized_train2014.json',
tagged=False, # Don't load tags.
lower=True) # Lowercase all descriptions.
val = parallel_sentences_from_file('./Data/COCO/Processed/tagged_val2014.json',
tagged=False, # Don't load tags.
lower=True) # Lowercase all descriptions.
# Compute stats for train and val data.
train_stats = parallel_stats(train)
val_stats = parallel_stats(val)
# Extra stats.
train_data = load_json('./Data/COCO/Processed/tokenized_train2014.json')
train_descriptions = [entry['caption'] for entry in train_data['annotations']]
val_data = load_json('./Data/COCO/Processed/tagged_val2014.json')
val_descriptions = [entry['caption'] for entry in val_data['annotations']]
extra_stats = sentence_stats(train_descriptions, val_descriptions)
val_stats.update(extra_stats)
# Save data to file.
save_json(train_stats, './Data/COCO/Processed/train_stats.json')
save_json(val_stats, './Data/COCO/Processed/val_stats.json')
'Vinyals-et-al-2017':
'\citeauthor{vinyals2017show} (\citeyear{vinyals2017show})',
'Wu-et-al-2016': '\citeauthor{wu2017image} (\citeyear{wu2017image})',
'Zhou-et-al-2017':
'\citeauthor{zhou2017watch} (\citeyear{zhou2017watch})'
}
data = [val_row] + system_rows
headers = [
'', '2', '3', '4', 'Ratio', 'Types-2', '1', '2', '3', '4', '5',
'Ratio', 'Types-1'
]
table = tabulate(data, headers, tablefmt='latex_booktabs')
# Modify table
# table = table.replace('{lrrrrrrrrrr}','{lrrrcrrrrrc}')
table = table.replace('& 0.3 &', '& 0.30 &')
table = table.replace(
'\\toprule',
'\\toprule \n & \multicolumn{3}{c}{Compound length} & \multicolumn{2}{c}{Compound stats} & \multicolumn{5}{c}{Prepositional phrase depth} & \multicolumn{2}{c}{PP stats}\\\\\n \cmidrule(lr){2-4} \cmidrule(lr){5-6} \cmidrule(lr){7-11} \cmidrule(lr){12-13}\n'
)
for system, cite in systems.items():
table = table.replace(system, cite)
# Print and save.
print(table)
with open('./Data/Output/nouns_pps_table.txt', 'w') as f:
f.write(table)
save_json(all_data, './Data/Output/nouns_pps.json')
