encA = EncoderA(args.wseed, zPrivate_dim=args.n_private)
decA = DecoderA(args.wseed, zPrivate_dim=args.n_private)
encB = EncoderB(args.wseed)
decB = DecoderB(args.wseed)
if CUDA:
encA.cuda()
decA.cuda()
encB.cuda()
decB.cuda()
cuda_tensors(encA)
cuda_tensors(decA)
cuda_tensors(encB)
cuda_tensors(decB)
optimizer = torch.optim.Adam(list(encB.parameters()) +
list(decB.parameters()) +
list(encA.parameters()) + list(decA.parameters()),
lr=args.lr)
def elbo(q,
pA,
pB,
lamb=1.0,
beta1=(1.0, 1.0, 1.0),
beta2=(1.0, 1.0, 1.0),
bias=1.0):
# from each of modality
reconst_loss_A, kl_A = probtorch.objectives.mws_tcvae.elbo(
q,
pA,
encA = EncoderA(args.wseed, zShared_dim=args.n_shared)
decA = DecoderA(args.wseed, zShared_dim=args.n_shared)
encB = EncoderB(args.wseed, zShared_dim=args.n_shared)
decB = DecoderB(args.wseed, zShared_dim=args.n_shared)
if CUDA:
encA.cuda()
decA.cuda()
encB.cuda()
decB.cuda()
cuda_tensors(encA)
cuda_tensors(decA)
cuda_tensors(encB)
cuda_tensors(decB)
optimizer = torch.optim.Adam(
list(encB.parameters()) + list(decB.parameters()) + list(encA.parameters()) + list(decA.parameters()),
lr=args.lr)
#
# def elbo(q, pA, pB, lamb=1.0, annealing_factor=1.0):
# muA_own = q['sharedA'].dist.loc.squeeze(0)
# stdA_own = q['sharedA'].dist.scale.squeeze(0)
# muB_own = q['sharedB'].dist.loc.squeeze(0)
# stdB_own = q['sharedB'].dist.scale.squeeze(0)
#
# # from each of modality
# reconst_loss_A = pA['images_own'].loss.mean()
# kl_A = 0.5 * torch.sum(1 + torch.log(stdA_own ** 2 + EPS) - muA_own ** 2 - stdA_own ** 2,
# dim=1).mean()
#
encA = EncoderA(num_pixels=4096, num_hidden=512, zPrivate_dim=args.n_private, zShared_dim=args.n_shared)
decA = DecoderA(num_pixels=4096, num_hidden=512, zPrivate_dim=args.n_private, zShared_dim=args.n_shared)
encB = EncoderB(num_pixels=4096, num_hidden=512, zPrivate_dim=args.n_private, zShared_dim=args.n_shared)
decB = DecoderB(num_pixels=4096, num_hidden=512, zPrivate_dim=args.n_private, zShared_dim=args.n_shared)
if CUDA:
encA.cuda()
decA.cuda()
encB.cuda()
decB.cuda()
cuda_tensors(encA)
cuda_tensors(decA)
cuda_tensors(encB)
cuda_tensors(decB)
optimizer = torch.optim.Adam(list(encB.parameters())+list(decB.parameters())+list(encA.parameters())+list(decA.parameters()),
lr=args.lr)
def elbo(iter, q, pA, pB, lamb=1.0, beta1=(1.0, 1.0, 1.0), beta2=(1.0, 1.0, 1.0), bias=1.0):
# from each of modality
reconst_loss_A, kl_A = probtorch.objectives.mws_tcvae.elbo(q, pA, pA['imagesA_sharedA'], latents=['privateA', 'sharedA'], sample_dim=0, batch_dim=1,
beta=beta1, bias=bias)
reconst_loss_B, kl_B = probtorch.objectives.mws_tcvae.elbo(q, pB, pB['imagesB_sharedB'], latents=['privateB', 'sharedB'], sample_dim=0, batch_dim=1,
beta=beta2, bias=bias)
if q['poe'] is not None:
# by POE
# 기대하는바:sharedA가 sharedB를 따르길. 즉 sharedA에만 digit정보가 있으며, 그 permutataion이 GT에서처럼 identity이기를
reconst_loss_poeA, kl_poeA = probtorch.objectives.mws_tcvae.elbo(q, pA, pA['imagesA_poe'], latents=['privateA', 'poe'], sample_dim=0, batch_dim=1,
beta=beta1, bias=bias)
# 의미 없음. rec 은 항상 0. 인풋이 항상 GT고 poe결과도 GT를 따라갈 확률이 크기 때문(학습 초반엔 A가 unif이라서, 학습 될수록 A가 B label을 잘 따를테니)