if args.b_distribution == 'none':
att_b_ = (att_b * 2 - 1) * args.thres_int
if args.b_distribution == 'uniform':
att_b_ = (att_b * 2 - 1) * \
torch.rand_like(att_b) * \
(2 * args.thres_int)
if args.b_distribution == 'truncated_normal':
att_b_ = (att_b * 2 - 1) * \
(torch.fmod(torch.randn_like(att_b), 2) + 2) / 4.0 * \
(2 * args.thres_int)
if (it + 1) % (args.n_d + 1) != 0:
errD = attgan.trainD(img_a, att_a, att_a_, att_b, att_b_)
add_scalar_dict(writer, errD, it + 1, 'D')
else:
errG = attgan.trainG(img_a, att_a, att_a_, att_b, att_b_)
add_scalar_dict(writer, errG, it + 1, 'G')
progressbar.say(epoch=epoch,
iter=it + 1,
d_loss=errD['d_loss'],
g_loss=errG['g_loss'])
if (it + 1) % args.save_interval == 0:
# To save storage space, I only checkpoint the weights of G.
# If you'd like to keep weights of G, D, optim_G, optim_D,
# please use save() instead of saveG().
attgan.saveG(
os.path.join('output', args.experiment_name, 'checkpoint',
'weights.{:d}.pth'.format(epoch)))
# attgan.save(os.path.join(
# 'output', args.experiment_name, 'checkpoint', 'weights.{:d}.pth'.format(epoch)
# 打乱顺序
idx = torch.randperm(len(label_org))
label_trg = label_org[idx].contiguous()
c_org = label_org.clone()
c_trg = label_trg.clone()
c_org = c_org.cuda() if args.gpu else c_org
c_trg = c_trg.cuda() if args.gpu else c_trg
attr_diff = c_trg - c_org
if (it + 1) % (args.n_d + 1) != 0:
errD = attgan.trainD(img_real, label_org, attr_diff)
add_scalar_dict(writer, errD, it + 1, 'D')
else:
errG = attgan.trainG(img_real, label_trg, attr_diff)
add_scalar_dict(writer, errG, it + 1, 'G')
progressbar.say(epoch=epoch,
iter=it + 1,
d_loss=errD['d_loss'],
g_loss=errG['g_loss'])
if (it + 1) % args.save_interval == 0:
# To save storage space, I only checkpoint the weights of G.
# If you'd like to keep weights of G, D, optim_G, optim_D,
# please use save() instead of saveG().
attgan.save(
os.path.join('output', args.experiment_name, 'checkpoint',
'weights.{:d}.pth'.format(epoch)))
if (it + 1) % args.sample_interval == 0: