BoTorch is designed in to be model-agnostic and only requries that a model conform to a minimal interface. This tutorial walks through an example of implementing the rank-weighted Gaussian process ensemble (RGPE) [Feurer, Letham, Bakshy ICML 2018 AutoML Workshop] and using the RGPE in BoTorch to do meta-learning across related optimization tasks.
import os
import torch
import math
torch.manual_seed(29)
device = torch.device("cuda:2" if torch.cuda.is_available() else "cpu")
dtype = torch.double
SMOKE_TEST = os.environ.get("SMOKE_TEST")
We consider optimizing the following 1-D synthetic function $$f(x, s_i) = \frac{1}{10}\bigg(x-1\bigg)\bigg(\sin(x+s_i)+\frac{1}{10}\bigg)$$ where $$s_i = \frac{(i+9)\pi}{8}$$ is a task-dependent shift parameter and $i$ is the task index $i \in [1, t]$.
In this tutorial, we will consider the scenario where we have collected data from 5 prior tasks (referred to as base tasks), which with a different task dependent shift parameter $s_i$.
The goal now is use meta-learning to improve sample efficiency when optimizing a 6th task.
First let's define a function for compute the shift parameter $s_i$ and set the shift amount for the target task.
NUM_BASE_TASKS = 5 if not SMOKE_TEST else 2
def task_shift(task):
"""
Fetch shift amount for task.
"""
return math.pi * task / 12.0
# set shift for target task
TARGET_SHIFT = 0.0
Then, let's define our function $f(x, s_i)$ and set bounds on $x$.
BOUNDS = torch.tensor([[-10.0], [10.0]], dtype=dtype, device=device)
def f(X, shift=TARGET_SHIFT):
"""
Torch-compatible objective function for the target_task
"""
f_X = X * torch.sin(X + math.pi + shift) + X / 10.0
return f_X
We sample data from a Sobol sequence to help ensure numerical stability when using a small amount of 1-D data. Sobol sequences help prevent us from sampling a bunch of training points that are close together.
from botorch.utils.sampling import draw_sobol_samples
from botorch.utils.transforms import normalize, unnormalize
noise_std = 0.05
# Sample data for each base task
data_by_task = {}
for task in range(NUM_BASE_TASKS):
num_training_points = 20
# draw points from a sobol sequence
raw_x = draw_sobol_samples(
bounds=BOUNDS,
n=num_training_points,
q=1,
seed=task + 5397923,
).squeeze(1)
# get observed values
f_x = f(raw_x, task_shift(task + 1))
train_y = f_x + noise_std * torch.randn_like(f_x)
train_yvar = torch.full_like(train_y, noise_std**2)
# store training data
data_by_task[task] = {
# scale x to [0, 1]
"train_x": normalize(raw_x, bounds=BOUNDS),
"train_y": train_y,
"train_yvar": train_yvar,
}
from matplotlib import pyplot as plt
%matplotlib inline
fig, ax = plt.subplots(1, 1, figsize=(12, 8))
x = torch.linspace(-10, 10, 51)
for task in data_by_task:
# plot true function and observed values for base runs
t = ax.plot(
unnormalize(data_by_task[task]["train_x"], bounds=BOUNDS).cpu().numpy(),
data_by_task[task]["train_y"].cpu().numpy(),
".",
markersize=10,
label=f"Observed task {task}",
)
ax.plot(
x.detach().numpy(),
f(x, task_shift(task + 1)).cpu().numpy(),
label=f"Base task {task}",
color=t[0].get_color(),
)
# plot true target function
ax.plot(
x.detach().numpy(),
f(x, TARGET_SHIFT).detach().numpy(),
"--",
label="Target task",
)
ax.legend(loc="lower right", fontsize=10)
plt.tight_layout()
First, let's define a helper function to fit a FixedNoiseGP with an fixed observed noise level.
from gpytorch.mlls import ExactMarginalLogLikelihood
from botorch.models import FixedNoiseGP
from botorch.fit import fit_gpytorch_mll
def get_fitted_model(train_X, train_Y, train_Yvar, state_dict=None):
"""
Get a single task GP. The model will be fit unless a state_dict with model
hyperparameters is provided.
"""
Y_mean = train_Y.mean(dim=-2, keepdim=True)
Y_std = train_Y.std(dim=-2, keepdim=True)
model = FixedNoiseGP(train_X, (train_Y - Y_mean) / Y_std, train_Yvar)
model.Y_mean = Y_mean
model.Y_std = Y_std
if state_dict is None:
mll = ExactMarginalLogLikelihood(model.likelihood, model).to(train_X)
fit_gpytorch_mll(mll)
else:
model.load_state_dict(state_dict)
return model
# Fit base model
base_model_list = []
for task in range(NUM_BASE_TASKS):
print(f"Fitting base model {task}")
model = get_fitted_model(
data_by_task[task]["train_x"],
data_by_task[task]["train_y"],
data_by_task[task]["train_yvar"],
)
base_model_list.append(model)
The main idea of the RGPE is to estimate the target function as weighted sum of the target model and the base models: $$\bar f(\mathbf x | \mathcal D) = \sum_{i=1}^{t} w_if^i(\mathbf x |\mathcal D_i)$$ Importantly, the ensemble model is also a GP: $$\bar f(\mathbf x | \mathcal D) \sim \mathcal N\bigg(\sum_{i=1}^{t} w_i\mu_i(\mathbf x), \sum_{i=1}^{t}w_i^2\sigma_i^2\bigg)$$
The weights $w_i$ for model $i$ are based on the the ranking loss between a draw from the model's posterior and the targets. Specifically, the ranking loss for model $i$ is: $$\mathcal L(f^i, \mathcal D_t) = \sum_{j=1}^{n_t}\sum_{k=1}^{n_t}\mathbb 1\bigg[\bigg(f^i\big(\mathbf x^t_j\big) < f^i\big(\mathbf x_k^t\big)\bigg)\oplus \big(y_j^t < y_k^t\big)\bigg]$$ where $\oplus$ is exclusive-or.
The loss for the target model is computing using leave-one-out cross-validation (LOOCV) and is given by: $$\mathcal L(f^t, \mathcal D_t) = \sum_{j=1}^{n_t}\sum_{k=1}^{n_t}\mathbb 1\bigg[\bigg(f^t_{-j}\big(\mathbf x^t_j\big) < f^t_{-j}\big(\mathbf x_k^t\big)\bigg)\oplus \big(y_j^t < y_k^t\big)\bigg]$$ where $f^t_{-j}$ model fitted to all data from the target task except training example $j$.
The weights are then computed as: $$w_i = \frac{1}{S}\sum_{s=1}^S\mathbb 1\big(i = \text{argmin}_{i'}l_{i', s}\big)$$
def roll_col(X, shift):
"""
Rotate columns to right by shift.
"""
return torch.cat((X[..., -shift:], X[..., :-shift]), dim=-1)
def compute_ranking_loss(f_samps, target_y):
"""
Compute ranking loss for each sample from the posterior over target points.
Args:
f_samps: `n_samples x (n) x n`-dim tensor of samples
target_y: `n x 1`-dim tensor of targets
Returns:
Tensor: `n_samples`-dim tensor containing the ranking loss across each sample
"""
n = target_y.shape[0]
if f_samps.ndim == 3:
# Compute ranking loss for target model
# take cartesian product of target_y
cartesian_y = torch.cartesian_prod(
target_y.squeeze(-1),
target_y.squeeze(-1),
).view(n, n, 2)
# the diagonal of f_samps are the out-of-sample predictions
# for each LOO model, compare the out of sample predictions to each in-sample prediction
rank_loss = (
(
(f_samps.diagonal(dim1=1, dim2=2).unsqueeze(-1) < f_samps)
^ (cartesian_y[..., 0] < cartesian_y[..., 1])
)
.sum(dim=-1)
.sum(dim=-1)
)
else:
rank_loss = torch.zeros(
f_samps.shape[0], dtype=torch.long, device=target_y.device
)
y_stack = target_y.squeeze(-1).expand(f_samps.shape)
for i in range(1, target_y.shape[0]):
rank_loss += (
(roll_col(f_samps, i) < f_samps) ^ (roll_col(y_stack, i) < y_stack)
).sum(dim=-1)
return rank_loss
Define a function to:
target_modeldef get_target_model_loocv_sample_preds(
train_x, train_y, train_yvar, target_model, num_samples
):
"""
Create a batch-mode LOOCV GP and draw a joint sample across all points from the target task.
Args:
train_x: `n x d` tensor of training points
train_y: `n x 1` tensor of training targets
target_model: fitted target model
num_samples: number of mc samples to draw
Return: `num_samples x n x n`-dim tensor of samples, where dim=1 represents the `n` LOO models,
and dim=2 represents the `n` training points.
"""
batch_size = len(train_x)
masks = torch.eye(len(train_x), dtype=torch.uint8, device=device).bool()
train_x_cv = torch.stack([train_x[~m] for m in masks])
train_y_cv = torch.stack([train_y[~m] for m in masks])
train_yvar_cv = torch.stack([train_yvar[~m] for m in masks])
state_dict = target_model.state_dict()
# expand to batch size of batch_mode LOOCV model
state_dict_expanded = {
name: t.expand(batch_size, *[-1 for _ in range(t.ndim)])
for name, t in state_dict.items()
}
model = get_fitted_model(
train_x_cv, train_y_cv, train_yvar_cv, state_dict=state_dict_expanded
)
with torch.no_grad():
posterior = model.posterior(train_x)
# Since we have a batch mode gp and model.posterior always returns an output dimension,
# the output from `posterior.sample()` here `num_samples x n x n x 1`, so let's squeeze
# the last dimension.
sampler = SobolQMCNormalSampler(sample_shape=torch.Size([num_samples]))
return sampler(posterior).squeeze(-1)
def compute_rank_weights(train_x, train_y, base_models, target_model, num_samples):
"""
Compute ranking weights for each base model and the target model (using
LOOCV for the target model). Note: This implementation does not currently
address weight dilution, since we only have a small number of base models.
Args:
train_x: `n x d` tensor of training points (for target task)
train_y: `n` tensor of training targets (for target task)
base_models: list of base models
target_model: target model
num_samples: number of mc samples
Returns:
Tensor: `n_t`-dim tensor with the ranking weight for each model
"""
ranking_losses = []
# compute ranking loss for each base model
for task in range(len(base_models)):
model = base_models[task]
# compute posterior over training points for target task
posterior = model.posterior(train_x)
sampler = SobolQMCNormalSampler(sample_shape=torch.Size([num_samples]))
base_f_samps = sampler(posterior).squeeze(-1).squeeze(-1)
# compute and save ranking loss
ranking_losses.append(compute_ranking_loss(base_f_samps, train_y))
# compute ranking loss for target model using LOOCV
# f_samps
target_f_samps = get_target_model_loocv_sample_preds(
train_x,
train_y,
train_yvar,
target_model,
num_samples,
)
ranking_losses.append(compute_ranking_loss(target_f_samps, train_y))
ranking_loss_tensor = torch.stack(ranking_losses)
# compute best model (minimum ranking loss) for each sample
best_models = torch.argmin(ranking_loss_tensor, dim=0)
# compute proportion of samples for which each model is best
rank_weights = (
best_models.bincount(minlength=len(ranking_losses)).type_as(train_x)
/ num_samples
)
return rank_weights
from botorch.models.gpytorch import GPyTorchModel
from gpytorch.models import GP
from gpytorch.distributions import MultivariateNormal
from gpytorch.lazy import PsdSumLazyTensor
from gpytorch.likelihoods import LikelihoodList
from torch.nn import ModuleList
class RGPE(GP, GPyTorchModel):
"""
Rank-weighted GP ensemble. Note: this class inherits from GPyTorchModel which provides an
interface for GPyTorch models in botorch.
"""
_num_outputs = 1 # metadata for botorch
def __init__(self, models, weights):
super().__init__()
self.models = ModuleList(models)
for m in models:
if not hasattr(m, "likelihood"):
raise ValueError(
"RGPE currently only supports models that have a likelihood (e.g. ExactGPs)"
)
self.likelihood = LikelihoodList(*[m.likelihood for m in models])
self.weights = weights
self.to(weights)
def forward(self, x):
weighted_means = []
weighted_covars = []
# filter model with zero weights
# weights on covariance matrices are weight**2
non_zero_weight_indices = (self.weights**2 > 0).nonzero()
non_zero_weights = self.weights[non_zero_weight_indices]
# re-normalize
non_zero_weights /= non_zero_weights.sum()
for non_zero_weight_idx in range(non_zero_weight_indices.shape[0]):
raw_idx = non_zero_weight_indices[non_zero_weight_idx].item()
model = self.models[raw_idx]
posterior = model.posterior(x)
# unstandardize predictions
posterior_mean = posterior.mean.squeeze(-1) * model.Y_std + model.Y_mean
posterior_cov = posterior.mvn.lazy_covariance_matrix * model.Y_std.pow(2)
# apply weight
weight = non_zero_weights[non_zero_weight_idx]
weighted_means.append(weight * posterior_mean)
weighted_covars.append(posterior_cov * weight**2)
# set mean and covariance to be the rank-weighted sum the means and covariances of the
# base models and target model
mean_x = torch.stack(weighted_means).sum(dim=0)
covar_x = PsdSumLazyTensor(*weighted_covars)
return MultivariateNormal(mean_x, covar_x)
from botorch.acquisition.monte_carlo import qNoisyExpectedImprovement
from botorch.sampling.normal import SobolQMCNormalSampler
from botorch.optim.optimize import optimize_acqf
# suppress GPyTorch warnings about adding jitter
import warnings
warnings.filterwarnings("ignore", "^.*jitter.*", category=RuntimeWarning)
best_rgpe_all = []
best_random_all = []
best_vanilla_nei_all = []
N_BATCH = 10 if not SMOKE_TEST else 2
NUM_POSTERIOR_SAMPLES = 256 if not SMOKE_TEST else 16
RANDOM_INITIALIZATION_SIZE = 3
N_TRIALS = 10 if not SMOKE_TEST else 2
MC_SAMPLES = 512 if not SMOKE_TEST else 32
N_RESTART_CANDIDATES = 512 if not SMOKE_TEST else 8
N_RESTARTS = 10 if not SMOKE_TEST else 2
Q_BATCH_SIZE = 1
# Average over multiple trials
for trial in range(N_TRIALS):
print(f"Trial {trial + 1} of {N_TRIALS}")
best_rgpe = []
best_random = []
best_vanilla_nei = []
# Initial random observations
raw_x = draw_sobol_samples(
bounds=BOUNDS, n=RANDOM_INITIALIZATION_SIZE, q=1, seed=trial
).squeeze(1)
train_x = normalize(raw_x, bounds=BOUNDS)
train_y_noiseless = f(raw_x)
train_y = train_y_noiseless + noise_std * torch.randn_like(train_y_noiseless)
train_yvar = torch.full_like(train_y, noise_std**2)
vanilla_nei_train_x = train_x.clone()
vanilla_nei_train_y = train_y.clone()
vanilla_nei_train_yvar = train_yvar.clone()
# keep track of the best observed point at each iteration
best_value = train_y.max().item()
best_rgpe.append(best_value)
best_random.append(best_value)
vanilla_nei_best_value = best_value
best_vanilla_nei.append(vanilla_nei_best_value)
# Run N_BATCH rounds of BayesOpt after the initial random batch
for iteration in range(N_BATCH):
target_model = get_fitted_model(train_x, train_y, train_yvar)
model_list = base_model_list + [target_model]
rank_weights = compute_rank_weights(
train_x,
train_y,
base_model_list,
target_model,
NUM_POSTERIOR_SAMPLES,
)
# create model and acquisition function
rgpe_model = RGPE(model_list, rank_weights)
sampler_qnei = SobolQMCNormalSampler(sample_shape=torch.Size([MC_SAMPLES]))
qNEI = qNoisyExpectedImprovement(
model=rgpe_model,
X_baseline=train_x,
sampler=sampler_qnei,
prune_baseline=False,
)
# optimize
candidate, _ = optimize_acqf(
acq_function=qNEI,
bounds=torch.tensor([[0.0], [1.0]], dtype=dtype, device=device),
q=Q_BATCH_SIZE,
num_restarts=N_RESTARTS,
raw_samples=N_RESTART_CANDIDATES,
)
# fetch the new values
new_x = candidate.detach()
new_y_noiseless = f(unnormalize(new_x, bounds=BOUNDS))
new_y = new_y_noiseless + noise_std * torch.randn_like(new_y_noiseless)
new_yvar = torch.full_like(new_y, noise_std**2)
# update training points
train_x = torch.cat((train_x, new_x))
train_y = torch.cat((train_y, new_y))
train_yvar = torch.cat((train_yvar, new_yvar))
random_candidate = torch.rand(1, dtype=dtype, device=device)
next_random_noiseless = f(unnormalize(random_candidate, bounds=BOUNDS))
next_random = next_random_noiseless + noise_std * torch.randn_like(
next_random_noiseless
)
next_random_best = next_random.max().item()
best_random.append(max(best_random[-1], next_random_best))
# get the new best observed value
best_value = train_y.max().item()
best_rgpe.append(best_value)
# Run Vanilla NEI for comparison
vanilla_nei_model = get_fitted_model(
vanilla_nei_train_x,
vanilla_nei_train_y,
vanilla_nei_train_yvar,
)
vanilla_nei_sampler = SobolQMCNormalSampler(
sample_shape=torch.Size([MC_SAMPLES])
)
vanilla_qNEI = qNoisyExpectedImprovement(
model=vanilla_nei_model,
X_baseline=vanilla_nei_train_x,
sampler=vanilla_nei_sampler,
)
vanilla_nei_candidate, _ = optimize_acqf(
acq_function=vanilla_qNEI,
bounds=torch.tensor([[0.0], [1.0]], dtype=dtype, device=device),
q=Q_BATCH_SIZE,
num_restarts=N_RESTARTS,
raw_samples=N_RESTART_CANDIDATES,
)
# fetch the new values
vanilla_nei_new_x = vanilla_nei_candidate.detach()
vanilla_nei_new_y_noiseless = f(unnormalize(vanilla_nei_new_x, bounds=BOUNDS))
vanilla_nei_new_y = vanilla_nei_new_y_noiseless + noise_std * torch.randn_like(
new_y_noiseless
)
vanilla_nei_new_yvar = torch.full_like(vanilla_nei_new_y, noise_std**2)
# update training points
vanilla_nei_train_x = torch.cat([vanilla_nei_train_x, vanilla_nei_new_x])
vanilla_nei_train_y = torch.cat([vanilla_nei_train_y, vanilla_nei_new_y])
vanilla_nei_train_yvar = torch.cat(
[vanilla_nei_train_yvar, vanilla_nei_new_yvar]
)
# get the new best observed value
vanilla_nei_best_value = vanilla_nei_train_y.max().item()
best_vanilla_nei.append(vanilla_nei_best_value)
best_rgpe_all.append(best_rgpe)
best_random_all.append(best_random)
best_vanilla_nei_all.append(best_vanilla_nei)
import numpy as np
best_rgpe_all = np.array(best_rgpe_all)
best_random_all = np.array(best_random_all)
best_vanilla_nei_all = np.array(best_vanilla_nei_all)
x = range(RANDOM_INITIALIZATION_SIZE, RANDOM_INITIALIZATION_SIZE + N_BATCH + 1)
fig, ax = plt.subplots(1, 1, figsize=(10, 6))
# Plot RGPE - NEI
ax.errorbar(
x,
best_rgpe_all.mean(axis=0),
yerr=1.96 * best_rgpe_all.std(axis=0) / math.sqrt(N_TRIALS),
label="RGPE - NEI",
linewidth=3,
capsize=5,
capthick=3,
)
# Plot FixedNoiseGP - NEI
ax.errorbar(
x,
best_vanilla_nei_all.mean(axis=0),
yerr=1.96 * best_vanilla_nei_all.std(axis=0) / math.sqrt(N_TRIALS),
label="FixedNoiseGP - NEI",
linewidth=3,
capsize=5,
capthick=3,
)
# Plot Random
ax.errorbar(
x,
best_random_all.mean(axis=0),
yerr=1.96 * best_random_all.std(axis=0) / math.sqrt(N_TRIALS),
label="Random",
linewidth=3,
capsize=5,
capthick=3,
)
ax.set_ylim(bottom=0)
ax.set_xlabel("Iteration", fontsize=12)
ax.set_ylabel("Best Observed Value", fontsize=12)
ax.set_title("Best Observed Value by Iteration", fontsize=12)
ax.legend(loc="lower right", fontsize=10)
plt.tight_layout()
