Add TopK downselection for initial batch generation. (#2636)

Summary:
## Motivation

In order to get facebook/Ax#2938 over the line with initial candidate generation that obey the constraints we want to use the existing tooling within `botorch`. The hard coded logic currently in Ax uses topk to downselect the sobol samples. To make a change there that will not impact existing users we then need to implement topk downselection in `botorch`.

Pull Request resolved: #2636

Test Plan:
TODO:
- [x] add tests for initialize_q_batch_topk

## Related PRs

facebook/Ax#2938. (#2610 was initially intended to play part of this solution but then I realized that the pattern I wanted to use was conflating repeats and the batch dimension.)

Reviewed By: Balandat

Differential Revision: D66413947

Pulled By: saitcakmak

fbshipit-source-id: 39e71f5cc0468d554419fa25dd545d9ee25289dc

botorch/optim/__init__.py

@@ -22,7 +22,11 @@
     LinearHomotopySchedule,
     LogLinearHomotopySchedule,
 )
-from botorch.optim.initializers import initialize_q_batch, initialize_q_batch_nonneg
+from botorch.optim.initializers import (
+    initialize_q_batch,
+    initialize_q_batch_nonneg,
+    initialize_q_batch_topn,
+)
 from botorch.optim.optimize import (
     gen_batch_initial_conditions,
     optimize_acqf,
@@ -43,6 +47,7 @@
     "gen_batch_initial_conditions",
     "initialize_q_batch",
     "initialize_q_batch_nonneg",
+    "initialize_q_batch_topn",
     "OptimizationResult",
     "OptimizationStatus",
     "optimize_acqf",

botorch/optim/initializers.py

@@ -271,13 +271,15 @@ def gen_batch_initial_conditions(
         fixed_features: A map `{feature_index: value}` for features that
             should be fixed to a particular value during generation.
         options: Options for initial condition generation. For valid options see
-            `initialize_q_batch` and `initialize_q_batch_nonneg`. If `options`
-            contains a `nonnegative=True` entry, then `acq_function` is
-            assumed to be non-negative (useful when using custom acquisition
-            functions). In addition, an "init_batch_limit" option can be passed
-            to specify the batch limit for the initialization. This is useful
-            for avoiding memory limits when computing the batch posterior over
-            raw samples.
+            `initialize_q_batch_topn`, `initialize_q_batch_nonneg`, and
+            `initialize_q_batch`. If `options` contains a `topn=True` then
+            `initialize_q_batch_topn` will be used. Else if `options` contains a
+            `nonnegative=True` entry, then `acq_function` is assumed to be
+            non-negative (useful when using custom acquisition functions).
+            `initialize_q_batch` will be used otherwise. In addition, an
+            "init_batch_limit" option can be passed to specify the batch limit
+            for the initialization. This is useful for avoiding memory limits
+            when computing the batch posterior over raw samples.
         inequality constraints: A list of tuples (indices, coefficients, rhs),
             with each tuple encoding an inequality constraint of the form
             `\sum_i (X[indices[i]] * coefficients[i]) >= rhs`.
@@ -328,14 +330,24 @@ def gen_batch_initial_conditions(
     init_kwargs = {}
     device = bounds.device
     bounds_cpu = bounds.cpu()
-    if "eta" in options:
-        init_kwargs["eta"] = options.get("eta")
-    if options.get("nonnegative") or is_nonnegative(acq_function):
+
+    if options.get("topn"):
+        init_func = initialize_q_batch_topn
+        init_func_opts = ["sorted", "largest"]
+    elif options.get("nonnegative") or is_nonnegative(acq_function):
         init_func = initialize_q_batch_nonneg
-        if "alpha" in options:
-            init_kwargs["alpha"] = options.get("alpha")
+        init_func_opts = ["alpha", "eta"]
     else:
         init_func = initialize_q_batch
+        init_func_opts = ["eta"]
+
+    for opt in init_func_opts:
+        # default value of "largest" to "acq_function.maximize" if it exists
+        if opt == "largest" and hasattr(acq_function, "maximize"):
+            init_kwargs[opt] = acq_function.maximize
+
+        if opt in options:
+            init_kwargs[opt] = options.get(opt)
 
     q = 1 if q is None else q
     # the dimension the samples are drawn from
@@ -363,7 +375,9 @@ def gen_batch_initial_conditions(
                         X_rnd_nlzd = torch.rand(
                             n, q, bounds_cpu.shape[-1], dtype=bounds.dtype
                         )
-                    X_rnd = bounds_cpu[0] + (bounds_cpu[1] - bounds_cpu[0]) * X_rnd_nlzd
+                    X_rnd = unnormalize(
+                        X_rnd_nlzd, bounds_cpu, update_constant_bounds=False
+                    )
             else:
                 X_rnd = sample_q_batches_from_polytope(
                     n=n,
@@ -375,7 +389,8 @@ def gen_batch_initial_conditions(
                     equality_constraints=equality_constraints,
                     inequality_constraints=inequality_constraints,
                 )
-            # sample points around best
+
+            # sample additional points around best
             if sample_around_best:
                 X_best_rnd = sample_points_around_best(
                     acq_function=acq_function,
@@ -395,6 +410,8 @@ def gen_batch_initial_conditions(
                     )
             # Keep X on CPU for consistency & to limit GPU memory usage.
             X_rnd = fix_features(X_rnd, fixed_features=fixed_features).cpu()
+
+            # Append the fixed fantasies to the randomly generated points
             if fixed_X_fantasies is not None:
                 if (d_f := fixed_X_fantasies.shape[-1]) != (d_r := X_rnd.shape[-1]):
                     raise BotorchTensorDimensionError(
@@ -411,6 +428,9 @@ def gen_batch_initial_conditions(
                     ],
                     dim=-2,
                 )
+
+            # Evaluate the acquisition function on `X_rnd` using `batch_limit`
+            # sized chunks.
             with torch.no_grad():
                 if batch_limit is None:
                     batch_limit = X_rnd.shape[0]
@@ -423,16 +443,22 @@ def gen_batch_initial_conditions(
                     ],
                     dim=0,
                 )
+
+            # Downselect the initial conditions based on the acquisition function values
             batch_initial_conditions, _ = init_func(
                 X=X_rnd, acq_vals=acq_vals, n=num_restarts, **init_kwargs
             )
             batch_initial_conditions = batch_initial_conditions.to(device=device)
+
+            # Return the initial conditions if no warnings were raised
             if not any(issubclass(w.category, BadInitialCandidatesWarning) for w in ws):
                 return batch_initial_conditions
+
             if factor < max_factor:
                 factor += 1
                 if seed is not None:
                     seed += 1  # make sure to sample different X_rnd
+
     warnings.warn(
         "Unable to find non-zero acquisition function values - initial conditions "
         "are being selected randomly.",
@@ -1057,6 +1083,56 @@ def initialize_q_batch_nonneg(
     return X[idcs], acq_vals[idcs]
 
 
+def initialize_q_batch_topn(
+    X: Tensor, acq_vals: Tensor, n: int, largest: bool = True, sorted: bool = True
+) -> tuple[Tensor, Tensor]:
+    r"""Take the top `n` initial conditions for candidate generation.
+
+    Args:
+        X: A `b x q x d` tensor of `b` samples of `q`-batches from a `d`-dim.
+            feature space. Typically, these are generated using qMC.
+        acq_vals: A tensor of `b` outcomes associated with the samples. Typically, this
+            is the value of the batch acquisition function to be maximized.
+        n: The number of initial condition to be generated. Must be less than `b`.
+
+    Returns:
+        - An `n x q x d` tensor of `n` `q`-batch initial conditions.
+        - An `n` tensor of the corresponding acquisition values.
+
+    Example:
+        >>> # To get `n=10` starting points of q-batch size `q=3`
+        >>> # for model with `d=6`:
+        >>> qUCB = qUpperConfidenceBound(model, beta=0.1)
+        >>> X_rnd = torch.rand(500, 3, 6)
+        >>> X_init, acq_init = initialize_q_batch_topn(
+        ...     X=X_rnd, acq_vals=qUCB(X_rnd), n=10
+        ... )
+
+    """
+    n_samples = X.shape[0]
+    if n > n_samples:
+        raise RuntimeError(
+            f"n ({n}) cannot be larger than the number of "
+            f"provided samples ({n_samples})"
+        )
+    elif n == n_samples:
+        return X, acq_vals
+
+    Ystd = acq_vals.std(dim=0)
+    if torch.any(Ystd == 0):
+        warnings.warn(
+            "All acquisition values for raw samples points are the same for "
+            "at least one batch. Choosing initial conditions at random.",
+            BadInitialCandidatesWarning,
+            stacklevel=3,
+        )
+        idcs = torch.randperm(n=n_samples, device=X.device)[:n]
+        return X[idcs], acq_vals[idcs]
+
+    topk_out, topk_idcs = acq_vals.topk(n, largest=largest, sorted=sorted)
+    return X[topk_idcs], topk_out
+
+
 def sample_points_around_best(
     acq_function: AcquisitionFunction,
     n_discrete_points: int,

botorch/utils/feasible_volume.py

@@ -11,7 +11,7 @@
 import botorch.models.model as model
 import torch
 from botorch.logging import _get_logger
-from botorch.utils.sampling import manual_seed
+from botorch.utils.sampling import manual_seed, unnormalize
 from torch import Tensor
 
 
@@ -164,9 +164,10 @@ def estimate_feasible_volume(
     seed = seed if seed is not None else torch.randint(0, 1000000, (1,)).item()
 
     with manual_seed(seed=seed):
-        box_samples = bounds[0] + (bounds[1] - bounds[0]) * torch.rand(
+        samples_nlzd = torch.rand(
             (nsample_feature, bounds.size(1)), dtype=dtype, device=device
         )
+        box_samples = unnormalize(samples_nlzd, bounds, update_constant_bounds=False)
 
     features, p_feature = get_feasible_samples(
         samples=box_samples, inequality_constraints=inequality_constraints

botorch/utils/sampling.py

@@ -98,14 +98,12 @@ def draw_sobol_samples(
     batch_shape = batch_shape or torch.Size()
     batch_size = int(torch.prod(torch.tensor(batch_shape)))
     d = bounds.shape[-1]
-    lower = bounds[0]
-    rng = bounds[1] - bounds[0]
     sobol_engine = SobolEngine(q * d, scramble=True, seed=seed)
-    samples_raw = sobol_engine.draw(batch_size * n, dtype=lower.dtype)
-    samples_raw = samples_raw.view(*batch_shape, n, q, d).to(device=lower.device)
+    samples_raw = sobol_engine.draw(batch_size * n, dtype=bounds.dtype)
+    samples_raw = samples_raw.view(*batch_shape, n, q, d).to(device=bounds.device)
     if batch_shape != torch.Size():
         samples_raw = samples_raw.permute(-3, *range(len(batch_shape)), -2, -1)
-    return lower + rng * samples_raw
+    return unnormalize(samples_raw, bounds, update_constant_bounds=False)
 
 
 def draw_sobol_normal_samples(

botorch/utils/transforms.py

@@ -66,17 +66,18 @@ def _update_constant_bounds(bounds: Tensor) -> Tensor:
     return bounds
 
 
-def normalize(X: Tensor, bounds: Tensor) -> Tensor:
+def normalize(X: Tensor, bounds: Tensor, update_constant_bounds: bool = True) -> Tensor:
     r"""Min-max normalize X w.r.t. the provided bounds.
 
-    NOTE: If the upper and lower bounds are identical for a dimension, that dimension
-    will not be scaled. Such dimensions will only be shifted as
-    `new_X[..., i] = X[..., i] - bounds[0, i]`. This avoids division by zero issues.
-
     Args:
         X: `... x d` tensor of data
         bounds: `2 x d` tensor of lower and upper bounds for each of the X's d
             columns.
+        update_constant_bounds: If `True`, update the constant bounds in order to
+            avoid division by zero issues. When the upper and lower bounds are
+            identical for a dimension, that dimension will not be scaled. Such
+            dimensions will only be shifted as
+            `new_X[..., i] = X[..., i] - bounds[0, i]`.
 
     Returns:
         A `... x d`-dim tensor of normalized data, given by
@@ -89,21 +90,27 @@ def normalize(X: Tensor, bounds: Tensor) -> Tensor:
         >>> bounds = torch.stack([torch.zeros(3), 0.5 * torch.ones(3)])
         >>> X_normalized = normalize(X, bounds)
     """
-    bounds = _update_constant_bounds(bounds=bounds)
+    bounds = (
+        _update_constant_bounds(bounds=bounds) if update_constant_bounds else bounds
+    )
     return (X - bounds[0]) / (bounds[1] - bounds[0])
 
 
-def unnormalize(X: Tensor, bounds: Tensor) -> Tensor:
+def unnormalize(
+    X: Tensor, bounds: Tensor, update_constant_bounds: bool = True
+) -> Tensor:
     r"""Un-normalizes X w.r.t. the provided bounds.
 
-    NOTE: If the upper and lower bounds are identical for a dimension, that dimension
-    will not be scaled. Such dimensions will only be shifted as
-    `new_X[..., i] = X[..., i] + bounds[0, i]`, matching the behavior of `normalize`.
-
     Args:
         X: `... x d` tensor of data
         bounds: `2 x d` tensor of lower and upper bounds for each of the X's d
             columns.
+        update_constant_bounds: If `True`, update the constant bounds in order to
+            avoid division by zero issues. When the upper and lower bounds are
+            identical for a dimension, that dimension will not be scaled. Such
+            dimensions will only be shifted as
+            `new_X[..., i] = X[..., i] + bounds[0, i]`. This is the inverse of
+            the behavior of `normalize` when `update_constant_bounds=True`.
 
     Returns:
         A `... x d`-dim tensor of unnormalized data, given by
@@ -116,7 +123,9 @@ def unnormalize(X: Tensor, bounds: Tensor) -> Tensor:
         >>> bounds = torch.stack([torch.zeros(3), 0.5 * torch.ones(3)])
         >>> X = unnormalize(X_normalized, bounds)
     """
-    bounds = _update_constant_bounds(bounds=bounds)
+    bounds = (
+        _update_constant_bounds(bounds=bounds) if update_constant_bounds else bounds
+    )
     return X * (bounds[1] - bounds[0]) + bounds[0]