def init_models(train_obs_mean, architecture, device):
"""Args:
train_obs_mean: tensor of shape [obs_dim]
architecture: linear_1, linear_2, linear_3 or non_linear
device: torch.device
Returns: generative_model, inference_network
"""
if architecture[:len('linear')] == 'linear':
num_stochastic_layers = int(architecture[-1])
generative_model = models.GenerativeModel(
num_stochastic_layers=num_stochastic_layers,
num_deterministic_layers=0,
device=device, train_obs_mean=train_obs_mean)
inference_network = models.InferenceNetwork(
num_stochastic_layers=num_stochastic_layers,
num_deterministic_layers=0,
device=device, train_obs_mean=train_obs_mean)
elif architecture == 'non_linear':
generative_model = models.GenerativeModel(
num_stochastic_layers=1,
num_deterministic_layers=2,
device=device, train_obs_mean=train_obs_mean)
inference_network = models.InferenceNetwork(
num_stochastic_layers=1,
num_deterministic_layers=2,
device=device, train_obs_mean=train_obs_mean)
if device.type == 'cuda':
generative_model.cuda()
inference_network.cuda()
return generative_model, inference_network
def load_models(model_folder='.', iteration=None, load_mws_memory=False):
"""Returns: generative_model, inference network
"""
if iteration is None:
suffix = ''
else:
suffix = iteration
generative_model_path = os.path.join(model_folder,
'gen{}.pt'.format(suffix))
inference_network_path = os.path.join(model_folder,
'inf{}.pt'.format(suffix))
if os.path.exists(generative_model_path):
args = load_object(get_args_path(model_folder))
generative_model = models.GenerativeModel(
args.init_mixture_logits,
softmax_multiplier=args.softmax_multiplier, device=args.device
).to(device=args.device)
inference_network = models.InferenceNetwork(
args.num_mixtures, args.relaxed_one_hot, args.temperature,
args.device).to(device=args.device)
generative_model.load_state_dict(torch.load(generative_model_path))
print_with_time('Loaded from {}'.format(generative_model_path))
inference_network.load_state_dict(torch.load(inference_network_path))
print_with_time('Loaded from {}'.format(inference_network_path))
if load_mws_memory:
mws_memory_path = os.path.join(model_folder,
'mws_mem{}.pkl'.format(suffix))
mws_memory = load_object(mws_memory_path)
return generative_model, inference_network, mws_memory
else:
return generative_model, inference_network
else:
return None, None
def evaluate_logp_sleep(
generative_model,
data_loader,
test_data_loader,
lstm_hidden_size,
obs_embedding_dim,
num_samples,
test_num_particles,
num_iterations,
log_interval,
):
device = next(generative_model.parameters()).device
inference_network = models.InferenceNetwork(
generative_model.num_primitives,
lstm_hidden_size,
generative_model.num_rows,
generative_model.num_cols,
generative_model.num_arcs,
obs_embedding_dim,
use_alphabet=generative_model.use_alphabet,
).to(device)
logp, losses = train_sleep_(
generative_model,
inference_network,
data_loader,
test_data_loader,
num_samples,
test_num_particles,
num_iterations,
log_interval,
)
return logp, losses
def init_models(pcfg_path):
"""Returns: generative_model, inference_network, true_generative_model"""
grammar, true_production_probs = read_pcfg(pcfg_path)
generative_model = models.GenerativeModel(grammar)
inference_network = models.InferenceNetwork(grammar)
true_generative_model = models.GenerativeModel(grammar,
true_production_probs)
return generative_model, inference_network, true_generative_model
def init(num_data, num_dim, true_cluster_cov, device):
prior_loc = torch.zeros(num_dim, device=device)
prior_cov = torch.eye(num_dim, device=device)
generative_model = models.GenerativeModel(num_data, prior_loc, prior_cov,
device).to(device)
inference_network = models.InferenceNetwork(num_data, num_dim).to(device)
true_generative_model = models.GenerativeModel(num_data, prior_loc,
prior_cov, device,
true_cluster_cov).to(device)
return (generative_model, inference_network, true_generative_model)
def init_models(args):
"""Returns: generative_model, inference_network, true_generative_model"""
generative_model = models.GenerativeModel(
args.init_mixture_logits, softmax_multiplier=args.softmax_multiplier,
device=args.device).to(device=args.device)
inference_network = models.InferenceNetwork(
args.num_mixtures, args.relaxed_one_hot, args.temperature,
args.device).to(device=args.device)
true_generative_model = models.GenerativeModel(
args.true_mixture_logits, softmax_multiplier=args.softmax_multiplier,
device=args.device).to(device=args.device)
return generative_model, inference_network, true_generative_model
def load_models(model_folder='.'):
"""Returns: generative_model, inference network
"""
generative_model_path = os.path.join(model_folder, 'gen.pt')
inference_network_path = os.path.join(model_folder, 'inf.pt')
pcfg_path_path = os.path.join(model_folder, 'pcfg_path.txt')
with open(pcfg_path_path) as f:
pcfg_path = f.read()
grammar, _ = read_pcfg(pcfg_path)
generative_model = models.GenerativeModel(grammar)
inference_network = models.InferenceNetwork(grammar)
generative_model.load_state_dict(torch.load(generative_model_path))
print_with_time('Loaded from {}'.format(generative_model_path))
inference_network.load_state_dict(torch.load(inference_network_path))
print_with_time('Loaded from {}'.format(inference_network_path))
return generative_model, inference_network
def init(run_args, device):
generative_model = models.GenerativeModel(
run_args.num_primitives,
run_args.initial_max_curve,
run_args.big_arcs,
run_args.p_lstm_hidden_size,
run_args.num_rows,
run_args.num_cols,
run_args.num_arcs,
run_args.likelihood,
run_args.p_uniform_mixture,
use_alphabet=run_args.condition_on_alphabet,
).to(device)
inference_network = models.InferenceNetwork(
run_args.num_primitives,
run_args.q_lstm_hidden_size,
run_args.num_rows,
run_args.num_cols,
run_args.num_arcs,
run_args.obs_embedding_dim,
run_args.q_uniform_mixture,
use_alphabet=run_args.condition_on_alphabet,
).to(device)
optimizer = init_optimizer(
generative_model,
inference_network,
1,
)
stats = Stats([], [], [], [], [], [], [], [])
if "mws" in run_args.algorithm:
memory = init_memory(
run_args.num_train_data,
run_args.memory_size,
generative_model.num_arcs,
generative_model.num_primitives,
device,
)
else:
memory = None
return generative_model, inference_network, optimizer, memory, stats