def middle_validate(net, trainloader, device, name=""):
print("validating vae and net ...")
known_energy, unknown_energy = [], []
# loading vae model
vae = VanillaVAE(in_channels=1, latent_dim=args.latent_dim)
vae = vae.to(device)
if device == 'cuda':
vae = torch.nn.DataParallel(vae)
cudnn.benchmark = True
if os.path.isfile(args.vae_resume):
vae_checkpoint = torch.load(args.vae_resume)
vae.load_state_dict(vae_checkpoint['net'])
print('==> Resuming vae from checkpoint, loaded..')
with torch.no_grad():
for batch_idx, (inputs, targets) in enumerate(trainloader):
inputs, targets = inputs.to(device), targets.to(device)
sampled = sampler(vae, device, args)
out_known = net(inputs)
out_unkown = net(sampled)
known_energy.append(out_known["pnorm"])
unknown_energy.append(out_unkown["pnorm"])
progress_bar(batch_idx, len(trainloader))
known_energy = torch.cat(known_energy, dim=0)
unknown_energy = torch.cat(unknown_energy, dim=0)
energy_hist_sperate(known_energy, unknown_energy, args, name)
return {
# unkown is smaller than known
"vae": vae,
"mid_known": known_energy.median().data,
"mid_unknown": unknown_energy.median().data
}
def mixup_validate(net, trainloader, mixuploader, device, stage="1"):
print("validating mixup ...")
known_energy, unknown_energy = [], []
with torch.no_grad():
batch_idx = -1
for (inputs, targets), (inputs_bak,
targets_bak) in zip(trainloader, mixuploader):
batch_idx += 1
inputs, targets = inputs.to(device), targets.to(device)
inputs_bak, targets_bak = inputs_bak.to(device), targets_bak.to(
device)
mixed = mixup(inputs, targets, inputs_bak, targets_bak, args)
out_known = net(inputs)
out_unkown = net(mixed)
known_energy.append(out_known["energy"])
unknown_energy.append(out_unkown["energy"])
progress_bar(batch_idx, len(trainloader))
known_energy = torch.cat(known_energy, dim=0)
unknown_energy = torch.cat(unknown_energy, dim=0)
energy_hist_sperate(known_energy, unknown_energy, args,
"mixup_stage" + stage)
return {
# unkown is smaller than known
"mid_known": known_energy.median().data,
"mid_unknown": unknown_energy.median().data
}
