def _process_stems(save_dir, encoder_stem, decoder_stem, losses_stem,
config_stem):
paths = [save_dir, encoder_stem, decoder_stem, losses_stem, config_stem]
save_dir, encoder_stem, decoder_stem, losses_stem, config_stem = check_path(
paths, path_type=Path)
encoder_base = encoder_stem.with_suffix('.h5')
decoder_base = decoder_stem.with_suffix('.h5')
losses_base = losses_stem.with_suffix('.dict')
config_base = config_stem.with_suffix('.json')
bases = encoder_base, decoder_base, losses_base, config_base
files = [save_dir / base for base in bases]
encoder_file, decoder_file, loss_file, config_file = check_path(
files, path_type=str)
return encoder_file, decoder_file, loss_file, config_file
def __init__(self,
images,
output_dir,
stem='reconstruction',
fmt='jpg',
batch_size=32):
super().__init__()
self.images = images
self.output_dir = check_path(output_dir, Path)
self.stem = stem
if fmt[0] != '.':
fmt = '.' + fmt
self.format = fmt
self.model_predict = None
self.batch_size = batch_size
def __init__(self,
csv_file,
data,
batch_size=512,
verbose=False,
overwrite=False):
super().__init__()
self.csv_file = check_path(csv_file, path_type=Path)
self.data = data
self.batch_size = batch_size
self.overwrite = overwrite
self.encoder = None
self.num_latents = None
self.csv_header = None
self.batch_means = None
self.batch_stds = None
self.verbose = verbose
def __init__(self,
image,
output_dir,
stem='traversal',
traversal_range=(-4, 4),
traversal_resolution=10):
super().__init__()
if image.ndim == 4:
assert (image.shape[0] == 1), (
'Traversal check can only be performed on a single image.')
else:
image = np.expand_dims(image, axis=0)
self.image = image
self.output_dir = check_path(output_dir, Path)
self.stem = stem
self.tcvae = None
self.traversal_range = traversal_range
self.traversal_resolution = traversal_resolution
def save(self,
save_dir,
encoder_stem='encoder',
decoder_stem='decoder',
losses_stem='losses',
config_stem='config',
overwrite=False):
"""
Saves model resources into a user-specified directory.
Parameters
----------
save_dir : str or pathlib.Path
The directory where model resources will be saved.
encoder_stem : str
The name of the saved encoder file without file extension.
decoder_stem : str
The name of the saved decoder file without file extension.
losses_stem : str
The name of the saved loss dictionary without file extension.
overwrite : bool
Whether to overwrite an already existing directory.
"""
save_dir = check_path(save_dir, Path)
if save_dir.exists() and overwrite:
rmtree(save_dir)
save_dir.mkdir()
encoder_file, decoder_file, loss_file, config_file = _process_stems(
save_dir, encoder_stem, decoder_stem, losses_stem, config_stem)
self.encoder.save(encoder_file)
self.decoder.save(decoder_file)
with open(loss_file, 'wb') as f:
pkl.dump(self.loss_dict, f)
config = {
'batch_size': self.batch_size,
'dataset_size': self.dataset_size
}
write_json(config_file, config)
def __init__(self, model, model_dir, monitor):
super().__init__()
self.model_ = model
self.model_dir = check_path(model_dir, Path)
self.monitor = monitor
self.best_value = float('inf')