model.save()
if args.verbosity >= 1:
sys.stdout.write(' (model saved)')
sys.stdout.flush()
time_since_save = timedelta()
elif args.verbosity >= 1:
# sys.stdout.write(' ')
sys.stdout.flush()
before = datetime.now()
train = {name: 0.0 for name in train}
n = 0
else:
for iteration in range(iteration_start, iteration_end + 1):
generated = dataset.generate(n=args.batch_size,
mode='train',
noise_range=args.pixel_noise)
model(data=generated, optimize=True, dropout=dropout)
if iteration % args.evaluation_frequency == 0 or iteration == 1 or iteration == args.evaluation_frequency // 2 or iteration == iteration_end:
generated = dataset.generate(n=args.evaluation_size,
mode='train',
noise_range=args.pixel_noise)
train = model(query=query, data=generated)
generated = dataset.generate(n=args.evaluation_size,
mode='validation',
noise_range=args.pixel_noise)
validation = model(query=query, data=generated)
after = datetime.now()
if args.report_file:
with open(args.report_file, 'a') as filehandle:
filehandle.write(str(iteration))
time=datetime.now().strftime('%H:%M:%S'),
dtype=dataset.type,
name=dataset.name,
mode=(' ' + mode if mode else '')))
sys.stdout.write(' 0% 0/{parts} (time per part: n/a)'.format(
parts=num_parts))
sys.stdout.flush()
for part in range(1, num_parts + 1):
before = datetime.now()
if args.unmanaged and len(parts) == 1 and parts[0] == 1:
path = directory
else:
path = os.path.join(directory, 'part{}'.format(start + part))
generated = dataset.generate(n=args.instances,
mode=mode,
noise_range=args.pixel_noise,
include_model=args.include_model,
alternatives=True)
if generated is None:
assert False
elif tf_records_flag:
tf_util.write_records(dataset=dataset,
records=generated,
path=path)
else:
dataset.serialize(path=path,
generated=generated,
archive=args.archive,
concat_worlds=args.concatenate_images,
html=args.html)
# if args.html and dataset.type == 'agreement':
from shapeworld import dataset
dataset = dataset(dtype='agreement',
name='oneshape_simple_textselect',
config='load(../data/oneshape_simple_textselect)')
generated = dataset.generate(n=250, mode='train')
k = ['caption_str', 'texts_str', 'pred_items']
for l in generated:
print(l, type(generated[l]))
if l == 'target':
print(generated[l].shape)
for i in range(10):
print(
f'Prediction items: {generated[k[2]][i]}, caption: {generated[k[0]][i]}, texts: {generated[k[1]][i]}'
)
from shapeworld import dataset
import pprint
dataset = dataset(
dtype='agreement',
name='oneshape_simple_textselect',
)
generated = dataset.generate(n=30,
mode='train',
noise_range=0.1,
include_model=True)
report_file_dir = os.path.dirname(args.report_file)
if not os.path.isdir(report_file_dir):
os.makedirs(report_file_dir)
with open(args.report_file, 'w') as filehandle:
filehandle.write(
'iteration,train loss,train accuracy,validation loss,validation accuracy\n'
)
iteration_end = iteration_start + args.iterations - 1
if args.evaluate:
# evaluation
sys.stdout.write('{} evaluate model...\n'.format(
datetime.now().strftime('%H:%M:%S')))
sys.stdout.write(' ')
sys.stdout.flush()
generated = dataset.generate(n=args.evaluation_size, mode='train')
feed_dict = {
placeholder: generated[value]
for value, placeholder in feed_dict_assignment.items()
}
feed_dict[dropout] = 0.0
training_accuracy = session.run(fetches=accuracy,
feed_dict=feed_dict)
sys.stdout.write('training={:.3f}'.format(training_accuracy))
generated = dataset.generate(n=args.evaluation_size,
mode='validation')
feed_dict = {
placeholder: generated[value]
for value, placeholder in feed_dict_assignment.items()
}
feed_dict[dropout] = 0.0
elif args.directory_unmanaged:
assert not args.append
assert not args.parts or len(args.parts) == 1
modes = (args.mode,)
directories = (args.directory_unmanaged,)
parts = args.parts or (1,)
start_part = (0,)
else:
assert False
if len(parts) == 1 and parts[0] == 1:
sys.stdout.write('{} generate {} {}{} data...\n'.format(datetime.now().strftime('%H:%M:%S'), dataset.type, dataset.name, ' ' + modes[0] if modes[0] else ''))
sys.stdout.flush()
generated = dataset.generate(n=args.instances, mode=modes[0], noise=(not args.no_pixel_noise), include_model=args.include_model)
dataset.serialize(directory=directories[0], generated=generated, archive=args.archive)
sys.stdout.write('{} data generation completed!\n'.format(datetime.now().strftime('%H:%M:%S')))
sys.stdout.flush()
else:
for mode, directory, num_parts, start in zip(modes, directories, parts, start_part):
if args.captioner_statistics:
filehandle = open(os.path.join(directory, 'captioner_statistics.csv'), 'a' if args.append else 'w')
dataset.collect_captioner_statistics(filehandle=filehandle, append=args.append)
sys.stdout.write('{} generate {}{} data...\n'.format(datetime.now().strftime('%H:%M:%S'), dataset, ' ' + mode if mode else ''))
sys.stdout.write(' 0% 0/{} (time per part: n/a)'.format(num_parts))
sys.stdout.flush()
for part in range(num_parts):
before = datetime.now()
generated = dataset.generate(n=args.instances, mode=mode, noise=(not args.no_pixel_noise), include_model=args.include_model)
dataset.serialize(directory=directory, generated=generated, name='part{}'.format(start + part), archive=args.archive)