def model_inverse(test_y, test_z):
x_test = model([test_z], c=[test_y], rev=True)
return x_test
### TRAINING CONFIGURATION AND HYPERPARAMETERS ###
######################################################
c = {
# GENERAL STUFF
'suffix':
f'{model.name}_conditional_cinn-2', # identifier for trained models and outputs
'device': 'cuda', # 'cuda' for GPU, 'cpu' for CPU
'interactive_visualization':
True, # requires visdom package to be installed
# DATA
'ndim_x': n_parameters,
'ndim_y': n_observations,
'ndim_z': n_parameters,
'data_model': model(),
'vis_y_target': (0.52466008, 0.21816375, 2.29708147),
# MODEL ARCHITECTURE
'n_blocks': 2,
'hidden_layer_sizes': 151, # 200k
'init_scale': 0.005,
# TRAINING HYPERPARAMETERS
'n_epochs': 50, # total number of epochs
'max_batches_per_epoch':
100, # iterations per epoch (or if training data exhausted)
'batch_size': 10000,
'n_test': 100000,
'n_train': 1000000,
'lr_init': 0.01, # initial learning rate
def model_inverse(test_z):
return model(test_z, rev=True)
def sample_conditional(y, z_x=None):
if z_x is None:
z_x = torch.randn(y.shape[0], n_parameters).to(device)
z_y, _ = model([y, z_x])
y, x = model([z_y, z_x], rev=True)
return x
def sample_joint(n_samples):
return model([
torch.randn(n_samples, n_observations).to(device),
torch.randn(n_samples, n_parameters).to(device)
],
rev=True)
def model_inverse(test_y, test_z):
z_y, z_x = model(
[test_y,
torch.randn(test_y.shape[0], n_parameters).to(c.device)])
y_test, x_test = model([z_y, test_z], rev=True)
return x_test
