def main():
parser = argparse.ArgumentParser(
description=
'Visualize the Disentanglement of ProgressiveGAN w/ Hessian Penalty')
# Model/Dataset Parameters:
parser.add_argument(
'--models',
required=True,
nargs='+',
help=
'Either the number of experiment in results directory, or a direct path to a .pkl '
'network checkpoint. You can specify multiple experiments/pkls '
'to generate visuals for all of them with one call to this script.')
parser.add_argument(
'--snapshot_kimgs',
default=['latest'],
nargs='+',
help=
'network-snapshot-<snapshot_kimg>.pkl to evaluate. This should either be "latest" or '
'a list of length equal to exp_numbers (each model needs a snapshot_kimg). '
'Not used if you are passing-in direct paths to .pkl checkpoints using --models.'
)
parser.add_argument('--seeds',
type=int,
default=[0],
nargs='+',
help='Seed for sampling the latent noise')
parser.add_argument(
'--nz_per_vid',
default=12,
type=int,
help=
'Number of z components to visualize per-video. This controls the "height" of the '
'generated videos.')
parser.add_argument(
'--samples',
default=8,
type=int,
help=
'Number of z samples to use (=interp_batch_size). This controls the "width" of the '
'generated videos.')
parser.add_argument(
'--steps',
default=90,
type=int,
help='Number of frames in video (=granularity of interpolation).')
parser.add_argument(
'--extent',
default=2.0,
type=float,
help='How "far" to move the z components (from -extent to extent)')
parser.add_argument(
'--minibatch_size',
default=100,
type=int,
help=
'Batch size to use when generating frames. If you get memory errors, try decreasing this.'
)
parser.add_argument(
'--interpolate_pre_norm',
action='store_true',
default=False,
help=
'If specified, interpolations are performed before the first pixel norm layer in G.'
'You should use this when nz is small (e.g., CLEVR-U).')
parser.add_argument('--no_loop',
action='store_true',
default=False,
help='If specified, saved video will not "loop".')
parser.add_argument(
'--stills_only',
action='store_true',
default=False,
help='If specified, only save frames instead of an mp4 video.')
parser.add_argument(
'--n_frames_to_save',
type=int,
default=0,
help=
'Number of "flattened" frames from video to save to png (0=disable).')
parser.add_argument(
'--transpose',
action='store_true',
default=False,
help='If specified, flips columns with rows in the video.')
parser.add_argument(
'--pad_x',
default=0,
type=int,
help='Padding between samples in video. WARNING: This can '
'cause weird problems with the video when '
'nz_per_vid > 1, so be careful using this.')
parser.add_argument(
'--pad_y',
default=0,
type=int,
help='Padding between rows in video. WARNING: This can '
'cause weird problems with the video when '
'nz_per_vid > 1, so be careful using this.')
opt = parser.parse_args()
opt = EasyDict(vars(opt))
if opt.pad_x > 0 or opt.pad_y > 0:
print(
'Warning: Using non-zero pad_x or pad_y can '
'cause moviepy to take a long time to make the video. '
'Also, there might be problems viewing the video with some applications '
'such as QuickTime.')
if os.path.isdir(opt.models[0]):
opt.models = sorted(glob(f'{opt.models[0]}/*/*.pkl'))
model_paths = [find_model(model) for model in opt.models]
opt.models = [m1 if m1 else m2 for m1, m2 in zip(model_paths, opt.models)]
if len(opt.snapshot_kimgs) == 1 and len(opt.models) > 1:
assert opt.snapshot_kimgs[0] == 'latest'
opt.snapshot_kimgs = ['latest'] * len(opt.models)
else:
assert len(opt.snapshot_kimgs) == len(opt.models)
for model, snapshot_kimg in zip(opt.models, opt.snapshot_kimgs):
for seed in opt.seeds:
opt.exp_number = model
opt.snapshot_kimg = snapshot_kimg
opt.seed = seed
run(opt)
