def test_sgvb(self):
μ, ρ1, ρ2, σ = 1.5, 0.2, 0.1, 1.5
torch.manual_seed(123)
params = dict(μ=μ, ρ1=ρ1, ρ2=ρ2, σ=σ)
model = AR2(input_length=100)
y = model.simulate(**params)
fit = sgvb(model, y, max_iters=200, quiet=True)
summ = fit.summary()
def test_training_loop(self):
torch.manual_seed(123)
m = LocalLevelModel(input_length=20)
y, z = m.simulate(γ=0., η=2., σ=1.5, ρ=0.85)
self.assertEqual(20, len(y))
self.assertEqual(20, len(z))
fit = sgvb(m, y, max_iters=8, quiet=True)
self.assertIsInstance(fit, MVNPosterior)
def test_plots(self):
torch.manual_seed(123)
m = UnivariateGaussian()
N, μ0, σ0 = 100, 5., 5.
y = m.simulate(N=N, μ=μ0, σ=σ0)
fit = sgvb(m, y, max_iters=100, quiet=True)
patch("ptvi.model.plt.show", fit.plot_marg_post("μ"))
patch("ptvi.model.plt.show", fit.plot_data())
patch("ptvi.model.plt.show", fit.plot_elbos())
def test_sgvb_gpu_double(self):
μ, ρ1, ρ2, σ = 1.5, 0.2, 0.1, 1.5
torch.manual_seed(123)
params = dict(μ=μ, ρ1=ρ1, ρ2=ρ2, σ=σ)
model = AR2(input_length=100, dtype=torch.float64, device=cuda)
y = model.simulate(**params)
self.assertEqual(y.dtype, torch.float64)
self.assertEqual(y.device.type, cuda.type)
fit = sgvb(model, y, max_iters=10, quiet=True)
summ = fit.summary()
def test_smoke_test_sgvb(self):
model = UnivariateGaussian()
torch.manual_seed(123)
N, μ0, σ0 = 100, 5., 5.
y = model.simulate(N=N, μ=μ0, σ=σ0)
fit = sgvb(model,
y,
max_iters=2**4,
num_draws=1,
sim_entropy=True,
quiet=True)
self.assertIsInstance(fit, MVNPosterior)
def test_training_loop_double_gpu(self):
torch.manual_seed(123)
m = LocalLevelModel(input_length=20, dtype=torch.float64, device=cuda)
y, z = m.simulate(γ=0., η=2., σ=1.5, ρ=0.85)
self.assertEqual(20, len(y))
self.assertEqual(y.device.type, cuda.type)
self.assertEqual(y.dtype, torch.float64)
self.assertEqual(20, len(z))
self.assertEqual(z.device.type, cuda.type)
self.assertEqual(z.dtype, torch.float64)
fit = sgvb(m, y, max_iters=8, quiet=True)
self.assertIsInstance(fit, MVNPosterior)
def test_smoke_test_sgvb_gpu_if_available(self):
model = UnivariateGaussian(device=cuda)
torch.manual_seed(123)
N, μ0, σ0 = 100, 5., 5.
y = model.simulate(N=N, μ=μ0, σ=σ0)
self.assertEqual(y.device.type, cuda.type)
fit = sgvb(model,
y,
max_iters=2**4,
num_draws=1,
sim_entropy=True,
quiet=True)
self.assertIsInstance(fit, MVNPosterior)
def test_training(self):
fll = FilteredLocalLevelModel(input_length=50)
true_params = dict(γ=0., η=2., ρ=0.95, σ=1.5)
algo_seed, data_seed = 123, 123
torch.manual_seed(data_seed)
y, z = fll.simulate(**true_params)
self.assertIsInstance(y, torch.Tensor)
self.assertEqual(y.shape, (50, ))
self.assertIsInstance(z, torch.Tensor)
self.assertEqual(z.shape, (50, ))
torch.manual_seed(algo_seed)
fit = sgvb(fll, y, max_iters=8, quiet=True)
self.assertIsInstance(fit, MVNPosterior)
def test_plots(self):
torch.manual_seed(123)
m = LocalLevelModel(input_length=20)
y, z = m.simulate(γ=0., η=2., σ=1.5, ρ=0.85)
fit = sgvb(m, y, max_iters=100, quiet=True)
patch("ptvi.model.plt.show", fit.plot_sample_paths())
patch("ptvi.model.plt.show", fit.plot_pred_ci(fc_steps=2, true_y=y))
patch("ptvi.model.plt.show", fit.plot_marg_post("η"))
patch("ptvi.model.plt.show", fit.plot_data())
patch("ptvi.model.plt.show", fit.plot_elbos())
patch("ptvi.model.plt.show",
fit.plot_latent(true_z=z, include_data=True))
def test_plots_and_forecasts_gpu(self):
μ, ρ1, ρ2, σ = 1.5, 0.2, 0.1, 1.5
torch.manual_seed(123)
params = dict(μ=μ, ρ1=ρ1, ρ2=ρ2, σ=σ)
model = AR2(input_length=20, dtype=torch.float64, device=cuda)
y = model.simulate(**params)
fit = sgvb(model, y, max_iters=10, quiet=True)
patch("ptvi.model.plt.show", fit.plot_sample_paths())
patch("ptvi.model.plt.show", fit.plot_sample_paths(fc_steps=2))
patch("ptvi.model.plt.show", fit.plot_pred_ci(fc_steps=2, true_y=y))
patch("ptvi.model.plt.show", fit.plot_marg_post("σ"))
patch("ptvi.model.plt.show", fit.plot_data())
patch("ptvi.model.plt.show", fit.plot_elbos())
def test_outputs(self):
torch.manual_seed(123)
m = LocalLevelModel(input_length=20)
y, z = m.simulate(γ=0., η=2., σ=1.5, ρ=0.85)
# we can't do much better than smoke test sampling methods
fit = sgvb(m, y, max_iters=100, quiet=True)
ss = fit.sample_paths(N=10, fc_steps=0)
self.assertEqual(ss.shape, (10, 20))
self.assertEqual(0, torch.sum(torch.isnan(ss)))
ss = fit.sample_paths(N=10, fc_steps=10)
self.assertEqual(ss.shape, (10, 20 + 10))
self.assertEqual(0, torch.sum(torch.isnan(ss)))
summ = fit.summary()
self.assertTrue(all(summ.index == ["γ", "η", "σ", "ρ"]))
def conditional(ctx, datafile, t, outfile, n, a, b, c, maxiters):
"""Forecast stoch vol model and compute log score, conditional on T observations.
Example:
stochvol --data_seed=123 --algo_seed=123 conditional experiment.csv 200 SV00200.json --N=100 --a=1. --b=0. --c=0.8
"""
assert t > 1
start_date, start_time = str(datetime.today()), time()
click.echo(_DIVIDER)
click.echo("Stochastic volatility model: conditional score estimation")
click.echo(_DIVIDER)
true_params = dict(a=a, b=b, c=c)
algo_seed = ctx.obj["algo_seed"]
data_seed = ctx.obj["data_seed"]
data = pd.read_csv(datafile)
click.echo(f"Started at: {start_date}")
click.echo(f"Reading {t}/{len(data)} observations from {datafile}.")
click.echo(f"True parameters assumed to be a={a}, b={b}, c={c}")
# draw N variates from p(y_T+1 | z_T+1, a, b, c)
click.echo(
f"Drawing {n} variates from p(y_T+1, z_T+1 | z_T, a, b, c) with "
f"data_seed={data_seed}"
)
torch.manual_seed(data_seed)
a, b, c = map(torch.tensor, (a, b, c))
z_next = b + c * data["z"][t - 1] + Normal(0, 1).sample((n,))
y_next = Normal(0, torch.exp(a) * torch.exp(z_next / 2)).sample()
y_next_list = y_next.cpu().numpy().squeeze().tolist() # for saving
# perform inference
y = data["y"][:t]
model = SVModel(input_length=t)
click.echo(repr(model))
torch.manual_seed(algo_seed)
fit = sgvb(model, y, max_iters=maxiters)
click.echo("Inference summary:")
click.echo(fit.summary(true=true_params))
click.echo(f"Generating {n} forecast draws from q...")
# filter to get p(z_T | y, θ) then project z_{T+1}, z_{T+2}, ...
forecast, fc_draws = fit.forecast(steps=1)
fc_draws_list = fc_draws.squeeze().tolist()
dens = forecast.pdf(y_next)
scores = np.log(dens[dens > 0])
score = np.mean(scores)
score_se = np.std(scores)
click.echo(f"Forecast log score = {score:.4f} nats (sd = {score_se:.4f}, n = {n})")
click.echo(f"Writing results to {outfile} in JSON format.")
y_list = data["y"][:t].tolist()
z_list = data["z"][:t].tolist()
summary = {
"method": "VSMC",
"algo_seed": algo_seed,
"data_seed": data_seed,
"datafile": datafile,
"t": t,
"outfile": outfile,
"fc_draws": fc_draws_list,
"score": score,
"score_se": score_se,
"n": n,
"y_next": y_next_list,
"start_date": start_date,
"elapsed": time() - start_time,
"true_params": true_params,
"full_length": len(data),
"max_iters": maxiters,
"inference_results": str(fit.summary()),
"y": y_list,
"z": z_list,
}
with open(outfile, "w", encoding="utf8") as ofilep:
json.dump(summary, ofilep, indent=4, sort_keys=True)
click.echo(f"Done in {time() - start_time:.1f} seconds.")
click.echo(_DIVIDER)
def test_training_loop(self):
m = StochVolModel(input_length=50)
y, b = m.simulate(λ=0.5, σ=0.5, φ=0.95)
fit = sgvb(m, y, max_iters=5, quiet=True)
self.assertIsInstance(fit, MVNPosterior)
