def main():
args = parse_args()
input_size = (settings.reduced_image_channels,
settings.reduced_image_width, settings.reduced_image_height)
vae = VAE(input_size=input_size,
latent_dim=settings.vae_latent_dim).to(settings.device)
savefile = Path(args.savefile)
if savefile.exists():
vae.load_state_dict(torch.load(f'{savefile}'))
vae.eval()
run(vae, savefile)
def eval(config, testloader):
storage = {
# 'll_precision': None, 'll_recall': None,
'log_densities': None,
'params': None,
'ground_truth': None
}
input_dim = testloader.dataset.input_dim_
vae = VAE(input_dim, config, checkpoint_directory=None)
vae.to(config['model']['device'])
if args.restore_filename is not None:
vae.restore_model(args.restore_filename, epoch=None)
vae.eval()
precisions, recalls, all_log_densities = [], [], []
# z sample sizes: 100
for i in range(100):
print("evaluation round {}".format(i))
_, _, precision, recall, log_densities, ground_truth = vae.evaluate(
testloader)
precisions.append(precision)
recalls.append(recall)
all_log_densities.append(np.expand_dims(log_densities, axis=1))
print(mean_confidence_interval(precisions))
print(mean_confidence_interval(recalls))
all_log_densities = np.concatenate(all_log_densities, axis=1)
# log sum exponential
storage['log_densities'] = logsumexp(all_log_densities,
axis=1) - np.log(100)
storage['ground_truth'] = ground_truth
# storage['ll_precision'] = mean_confidence_interval(precisions)
# storage['ll_recall'] = mean_confidence_interval(recalls)
# storage['params'] = self._get_parameters(testloader)
pkl_filename = './results/test/{}{}/{}.pkl'.format(config['model']['name'], \
config['model']['config_id'], args.restore_filename)
os.makedirs(os.path.dirname(pkl_filename), exist_ok=True)
with open(pkl_filename, 'wb') as _f:
pickle.dump(storage, _f, pickle.HIGHEST_PROTOCOL)
