start_time = time.time()
counter = 1
mse_loss = torch.nn.MSELoss(reduction='mean')
mse_loss_sum = torch.nn.MSELoss(reduction='sum')
latent_mse_loss = torch.nn.MSELoss(reduction='sum')
from covNetModel import Encoder
from covNetModel import Decoder
num_samp = opt.num_samp
encoder = Encoder(z_dim, opt.c_dim, opt.img_size)
decoder = Decoder(z_dim, opt.c_dim, opt.img_size)
transNet = TransOpt()
sampler_c = Sample_c()
# Load a pretrained file - make sure the model parameters fit the parameters of rht pre-trained models
if data_use == 'rotDigits':
checkpoint = torch.load(
'./pretrained_models/rotMNIST/network_batch32_zdim10.pt')
elif data_use == 'natDigits':
checkpoint = torch.load(
'./pretrained_models/natMNIST/network_batch32_zdim6.pt')
# Or specify a network that's been trained independently
#checkpoint = torch.load(save_folder + 'network_batch' + str(batch_orig) + '_zdim' + str(opt.z_dim) + '_step' + str(opt.stepUse) + '.pt')
encoder.load_state_dict(checkpoint['model_state_dict_encoder'])
decoder.load_state_dict(checkpoint['model_state_dict_decoder'])
sample_X, sample_labels, _ = load_mnist("test")
data_X, data_y, y_orig = load_mnist("train")
sample_labels = data_y
sample_X_torch = torch.from_numpy(data_X)
sample_X_torch = sample_X_torch.permute(0, 3, 1, 2)
sample_X_torch = sample_X_torch.float()
batch_idxs = len(data_X) // opt.batch_size
anchor_X, anchor_y = select_mnist_anchors(data_X, data_y, opt.num_anchor)
anchor_X_torch = torch.from_numpy(anchor_X)
anchor_X_torch = anchor_X_torch.permute(0, 3, 1, 2)
anchor_X_torch = anchor_X_torch.float()
encoder = Encoder(z_dim, opt.c_dim, opt.img_size)
decoder = Decoder(z_dim, opt.c_dim, opt.img_size)
anchor_orig = 1
num_class = 10
transNet = TransOpt()
sampler_c = Sample_c()
# Initialize weights
encoder.apply(weights_init_normal)
decoder.apply(weights_init_normal)
## Initialize dictionary
Psi = Variable(torch.mul(torch.randn(N_use, M, dtype=torch.double), 0.01),
requires_grad=True)
# Load previously trained networks - replace checkpoint_folder with a checkpoint that is useful to you
if load_checkpoint == 1:
checkpoint_folder = './results/checkpoint_folder/'
checkpoint = torch.load(checkpoint_folder +
'network_batch32_zdim10_step31000.pt')
os.makedirs(save_folder)
if data_use == 'rotDigits':
sample_dir = './' + opt.model + opt.data_use + '_batch' + str(
opt.batch_size) + '_zdim' + str(
opt.z_dim) + '_zeta' + str(zeta) + '_gamma' + str(
opt.gamma) + '_scale' + str(scale) + '_M' + str(
M) + '_' + opt.training_phase + '_samples/'
if not os.path.exists(sample_dir):
os.makedirs(sample_dir)
# Define the loss functions
mse_loss = torch.nn.MSELoss(reduction='mean')
latent_mse_loss = torch.nn.MSELoss(reduction='sum')
transNet = TransOpt()
# Load data
nTrain = 4000
nTest = 500
noise_std = 0.01
if data_use == 'concen_circle':
from fullyConnectedModel import Encoder
from fullyConnectedModel import Decoder
mapMat = np.random.uniform(-1, 1, (x_dim, z_dim))
if opt.training_phase == 'AE_train':
sio.savemat(
save_folder + 'mapMat_circleHighDim_z' + str(z_dim) + '_x' +
str(x_dim) + '.mat', {'mapMat': mapMat})
else:
sample_labels = data_y
sample_X_torch = torch.from_numpy(data_X)
sample_X_torch = sample_X_torch.permute(0,3,1,2)
sample_X_torch = sample_X_torch.float()
batch_idxs = len(data_X) // opt.batch_size
anchor_X,anchor_y = select_mnist_anchors(data_X,data_y,opt.num_anchor)
anchor_X_torch = torch.from_numpy(anchor_X)
anchor_X_torch= anchor_X_torch.permute(0,3,1,2)
anchor_X_torch = anchor_X_torch.float()
encoder = Encoder(z_dim,opt.c_dim,opt.img_size)
decoder = Decoder(z_dim,opt.c_dim,opt.img_size)
anchor_orig = 1
num_class = 10
transNet = TransOpt()
sampler_c = Sample_c()
# Initialize weights
encoder.apply(weights_init_normal)
decoder.apply(weights_init_normal)
## Initialize dictionary
Psi = Variable(torch.mul(torch.randn(N_use, M, dtype=torch.double),0.01), requires_grad=True)
# Load previously trained networks - replace checkpoint_folder with a checkpoint that is useful to you
if load_checkpoint == 1:
checkpoint_folder = './results/checkpoint_folder/'
checkpoint = torch.load(checkpoint_folder + 'network_batch32_zdim10_step31000.pt')
encoder.load_state_dict(checkpoint['model_state_dict_encoder'])