def get_opt():
opt = Namespace()
# Model
opt.CAPACITY = 32
opt.NUM_COMPONENTS = 10
opt.GMM_NUM_COMPONENTS = 10
# Training
opt.LR = 0.001
opt.NUM_ITER = 3000
opt.CUDA = True
opt.REC_FREQ = 10
# Meta
opt.ALPHA_LR = 0.1
# opt.ALPHA_NUM_ITER = 500
# opt.ALPHA_NUM_ITER = 50
opt.ALPHA_NUM_ITER = 10
opt.FINETUNE_LR = 0.001
opt.FINETUNE_NUM_ITER = 10
opt.PARAM_DISTRY = lambda mean: normal(mean, 2, opt.NUM_SAMPLES)
opt.PARAM_SAMPLER = lambda: np.random.uniform(-4, 4)
# Sampling
opt.NUM_SAMPLES = 1000
opt.TRAIN_DISTRY = lambda: normal(0, 2, opt.NUM_SAMPLES)
opt.TRANS_DISTRY = lambda: normal(random.randint(-4, 4), 2, opt.NUM_SAMPLES
)
return opt
with torch.no_grad():
_X = torch.tensor([[1.]])
_Y = torch.tensor([[0.]])
if opt.CUDA:
_X, _Y = _X.to(encoder.theta.device), _Y.to(encoder.theta.device)
_X_d, _Y_d = decoder(_X, _Y)
_X_de, _Y_de = encoder(_X_d, _Y_d)
print(f"Initial (A, B) = {_X.item()}, {_Y.item()}")
print(f"Decoded (X, Y) = {_X_d.item()}, {_Y_d.item()}")
print(f"Encoded (U, V) = {_X_de.item()}, {_Y_de.item()}")
opt = Namespace()
# Model
opt.CAPACITY = 32
opt.NUM_COMPONENTS = 10
opt.GMM_NUM_COMPONENTS = 10
# Training
opt.LR = 0.01
opt.NUM_ITER = 20
opt.NUM_META_ITER = 1000
opt.ENCODER_LR = 0.01
opt.ALPHA_LR = 0.001
opt.CUDA = True
opt.REC_FREQ = 10
opt.ALPHA_INIT = 0.
opt.USE_BASELINE = True
# Fine tuning