def run_once():
model = DiffusionModel(n, m, ep, p, q)
search_tree = {}
nsim_total = 0
while not model.has_stabilized():
best_target, nsim = search(model, search_tree)
model.step(best_target)
nsim_total += nsim
return model.get_value(), nsim_total
def optimize():
log_dir = opt.logging_root
cond_mkdir(log_dir)
summaries_dir = os.path.join(log_dir, opt.experiment_name)
cond_mkdir(summaries_dir)
m = DiffusionModel(us_init=opt.us_init,
ua_init=opt.ua_init,
n_init=opt.n_init)
m.to(device)
optim = torch.optim.Adam(m.parameters(), lr=opt.lr, amsgrad=True)
writer = SummaryWriter(summaries_dir)
converged = False
converged_eps = opt.convergence
prev_loss = 1e6
for ii in range(opt.num_iters):
loss, model, data = objective(m)
optim.zero_grad()
# write summary
writer.add_scalar('mse', loss, ii)
if not ii % opt.steps_til_summary:
write_summary(writer, m, loss, model, data, ii)
print(f'{ii}: {loss.detach().cpu().numpy():03f}')
loss.backward()
optim.step()
# values should be non-negative
def clamp_nonnegative(m):
m.ua.data = torch.clamp(m.ua.data, min=0)
m.us.data = torch.clamp(m.us.data, min=0)
m.apply(clamp_nonnegative)
if torch.abs(loss - prev_loss) < converged_eps:
converged = True
break
prev_loss = loss.clone()
out = {
'us': m.us.detach().cpu().numpy().squeeze().item() * m.us_scale,
'ua': m.ua.detach().cpu().numpy().squeeze().item(),
't0': m.t0.detach().cpu().numpy().squeeze(),
'us_init': opt.us_init,
'ua_init': opt.ua_init,
'n_init': opt.n_init,
'mse': loss.detach().cpu().numpy().squeeze().item(),
'iters': ii,
'converged': converged,
'converged_eps': converged_eps
}
np.save(os.path.join(summaries_dir, 'out.npy'), out)
def run_once():
model = DiffusionModel(n, m, ep, p, q)
while not model.has_stabilized():
observed_infected, resistant = model.get_observed_infected(), model.get_resistant()
if model.has_stabilized(observed_infected, resistant):
# Observed model has stabilized; take no action for one round
target = None
else:
if strategy == 1:
target = random.choice(tuple(observed_infected))
elif strategy == 2:
target = get_best_target(model, observed_infected, resistant)
else:
target = random.choice(tuple(set(range(n)).difference(resistant)))
model.step(target)
return model.get_value()
def objective(m: DiffusionModel):
# get predicted diffusion model
model = []
for idx, thickness in enumerate(m.thicknesses):
d = 0.0254 * thickness
model.append(m.forward(d, m.ua, m.us, m.t0[idx]))
model = torch.cat(model, dim=0)
# convolve model with measured impulse response
model = F.conv1d(model, m.laser, padding=(m.laser.shape[-1]))
max_vals, _ = torch.max(model, dim=2, keepdim=True)
model = model / max_vals
# get data
data = m.data.clone()
# calculate error
mse = torch.sum((model - data)**2)
return mse, model, data
