def end_log_game_action(self, game_name):
import json
if game_name not in self._game_info: return
game = self._game_info.get(game_name)
if 'actions' in game:
file_path = os.path.join(self.logdir, 'game_info')
misc.ensure_dir(file_path)
file_path = os.path.join(file_path, game_name + '.actions.json')
misc.save_file(file_path, json.dumps(game['actions']))
del self._game_info[game_name]
def save_structure_to_files(regime):
'''Parse the metadata to compute a mapping between structures and their corresponding datapoints.'''
print('Saving structure metadata')
data_dir = '../data/' + regime + '/'
data_files = os.listdir(data_dir)
data_files = [data_dir + data_file for data_file in data_files]
structure_to_files = {}
for count, data_file in enumerate(data_files):
if count % 10000 == 0:
print(count, '/', len(data_files))
data = np.load(data_file)
# TODO Serialize structure instead of using string representation
structure = parse_structure(data['relation_structure']).to_str()
if structure not in structure_to_files:
structure_to_files[structure] = [data_file]
else:
structure_to_files[structure].append(data_file)
if os.path.exists('save_state/' + regime):
os.mkdir('save_state/' + regime)
misc.save_file(structure_to_files,
'save_state/' + regime + '/structure_to_files.pkl')
return structure_to_files
def save_normalization_stats(regime, batch_size=100):
'''Compute the mean and standard deviation jointly across all channels.'''
print('Saving normalization stats')
data_dir = '../data/' + regime + '/'
data_files = os.listdir(data_dir)
data_files = [data_dir + data_file for data_file in data_files]
train_files = [
data_file for data_file in data_files if 'train' in data_file
]
loader = torch.utils.data.DataLoader(PreprocessDataset(
train_files, 'image'),
batch_size=batch_size)
print('Computing x_mean')
sum = 0
n = 0
count = 0
for x in loader:
sum += x.sum()
n += x.numel()
count += batch_size
if count % 100000 == 0:
print(count, '/', len(train_files))
x_mean = float(sum / n)
print('Computing x_sd')
sum = 0
n = 0
count = 0
for x in loader:
sum += ((x - x_mean)**2).sum()
n += x.numel()
count += batch_size
if count % 100000 == 0:
print(count, '/', len(train_files))
x_sd = float(np.sqrt(sum / n))
misc.save_file((x_mean, x_sd),
'save_state/' + regime + '/normalization_stats.pkl')
return x_mean, x_sd
def write_wav(self, full_out_file=None):
"""
Write sound data to a WAV-file.
Parameters
----------
fullOutFile : string
Path- and file-name of the outfile. If none is given,
the user is asked interactively to choose a folder/name
for the outfile.
Returns
-------
None :
Examples
--------
>>> mySound = Sound()
>>> mySound.read_sound('test.wav')
>>> mySound.write_wav()
"""
if full_out_file is None:
(out_file,
out_dir) = misc.save_file(FilterSpec='*.wav',
DialogTitle='Write sound to ...',
DefaultName='')
full_out_file = os.path.join(out_dir, out_file)
if full_out_file is None:
print('Output discarded.')
return 0
else:
(out_file, out_dir) = os.path.split(full_out_file)
write(str(full_out_file), int(self.rate), self.data)
print('Sounddata written to ' + out_file + ', with a sample rate of ' +
str(self.rate))
print('OutDir: ' + out_dir)
return full_out_file
def save_train_info(self):
import json
train_info_path = os.path.join(self.logdir, 'train_info.txt')
misc.save_file(train_info_path, json.dumps(self.train_info))