def build_custom_post_net(batch_theta, batch_x):
flow_lik, flow_post = func.set_up_networks(model.prior.base_dist.low,
model.prior.base_dist.high,
dim_post=model.nbr_params)
return flow_post
# w_sim_wrapper = lambda param: torch.as_tensor(func.whiten(simulator_wrapper(param), whiteness_params))
def simulator(params):
N = params.shape[0]
data = torch.zeros(params.shape[0], 19)
for i in range(N):
data[i, :] = torch.as_tensor(
func.whiten(simulator_wrapper(params[i, :]), whiteness_params))
return data
flow_lik, flow_post = func.set_up_networks(model.prior.low,
model.prior.high,
dim_post=model.nbr_params)
# setting for not exteded:
# decay_rate_post = 0.95
# prob_prior_decay_rate = 0.9
# 1000, 10000
# setting for exteded:
# decay_rate_post = 0.9
# prob_prior_decay_rate = 0.9
# 2000, 10000
optimizer_lik = torch.optim.Adam(flow_lik.parameters())
optimizer_post = torch.optim.Adam(flow_post.parameters(),
lr=0.001,
def build_custom_like_net(batch_theta, batch_x):
flow_lik, flow_post = func.set_up_networks(seed)
return flow_lik
s_of_theta = model.model_sim(theta, True)
return func.normalize_summary_stats(s_of_theta, m_s_of_prior, s_s_of_prior)
hyper_params = [0.001, 0.001, 0.9, 0.9] # lr_like, lr_post, gamma_post, gamma
if lambda_val > 0:
hyper_params[-1] = lambda_val
if hp_tuning >= 2:
hyper_params = func.sample_hp("snpla", hp_tuning)
print(hyper_params)
flow_lik, flow_post = func.set_up_networks()
optimizer_lik = torch.optim.Adam(flow_lik.parameters(), lr=hyper_params[0])
optimizer_post = torch.optim.Adam(flow_post.parameters(), lr=hyper_params[1], weight_decay=0.0) # used def value before
decay_rate_post = hyper_params[2] # no adaptation of Adam's base rate
nbr_rounds = 5
prob_prior_decay_rate = hyper_params[3]
prob_prior = snpla.calc_prob_prior(nbr_rounds, prob_prior_decay_rate)
print(prob_prior)
nbr_lik = [1000, 1000, 1000, 1000, 1000] # , 2000, 2000]
nbr_epochs_lik = [50, 50, 50, 50, 50]
batch_size = 25
batch_size_post = 2000
nbr_post = [10000, 10000, 10000, 10000, 10000] # , 10000, 10000]
