Skip to content

Commit

Permalink
made opt test optimization problems easier. reorganized opt tests to …
Browse files Browse the repository at this point in the history 
…do more class-wide precomputing. reorganized opt tests to use UnivariateSampler.
  • Loading branch information
@bwpriest
bwpriest committed Dec 19, 2023
1 parent b18c898 commit 1047098
Show file tree
Hide file tree
Showing 5 changed files with 152 additions and 196 deletions.
143 changes: 72 additions & 71 deletions MuyGPyS/_test/optimize.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -5,33 +5,32 @@

from absl.testing import parameterized

import MuyGPyS._src.math as mm

from MuyGPyS._src.mpi_utils import _consistent_chunk_tensor
from MuyGPyS._test.utils import _check_ndarray, _sq_rel_err
from MuyGPyS._test.gp import benchmark_sample_full, BenchmarkGP
from MuyGPyS._test.utils import (
_check_ndarray,
_sq_rel_err,
_basic_nn_kwarg_options,
)
from MuyGPyS._test.sampler import UnivariateSampler
from MuyGPyS.gp.deformation import Isotropy, l2
from MuyGPyS.gp.kernels import Matern
from MuyGPyS.gp.hyperparameter import ScalarParam
from MuyGPyS.gp.noise import HomoscedasticNoise
from MuyGPyS.gp.tensors import pairwise_tensor, crosswise_tensor
from MuyGPyS.neighbors import NN_Wrapper
from MuyGPyS.optimize.batch import sample_batch


class BenchmarkTestCase(parameterized.TestCase):
@classmethod
def setUpClass(cls):
super(BenchmarkTestCase, cls).setUpClass()
cls.data_count = 1001
cls.its = 13
cls.sim_train = dict()
cls.xs = mm.linspace(-10.0, 10.0, cls.data_count).reshape(
cls.data_count, 1
)
cls.train_features = cls.xs[::2, :]
cls.test_features = cls.xs[1::2, :]
cls.train_count, _ = cls.train_features.shape
cls.test_count, _ = cls.test_features.shape
cls.data_count = 501
cls.its = 7
cls.train_ratio = 0.50
cls.batch_count = 150
cls.nn_count = 10

cls.feature_count = 1
cls.response_count = 1
cls.length_scale = 1e-2
Expand All @@ -51,17 +50,14 @@ def setUpClass(cls):
}

cls.params = {
"length_scale": ScalarParam(1e-1, (1e-2, 1e0)),
"smoothness": ScalarParam(0.78, (1e-1, 2e0)),
"length_scale": ScalarParam(0.05, (1e-2, 1e0)),
"smoothness": ScalarParam(2.0, (1e-1, 5)),
"noise": HomoscedasticNoise(1e-5, (1e-8, 1e-2)),
}

cls.setUpGP()
cls.simulate()

@classmethod
def setUpGP(cls):
cls.gp = BenchmarkGP(
cls.sampler = UnivariateSampler(
data_count=cls.data_count,
train_ratio=cls.train_ratio,
kernel=Matern(
smoothness=ScalarParam(cls.params["smoothness"]()),
deformation=Isotropy(
Expand All @@ -71,79 +67,84 @@ def setUpGP(cls):
),
noise=HomoscedasticNoise(cls.params["noise"]()),
)
cls.gp.scale._set(5.0)
cls.sampler.gp.scale._set(5.0)

@classmethod
def simulate(cls):
cls.ys = mm.zeros((cls.its, cls.data_count, cls.response_count))
cls.train_responses = mm.zeros(
(cls.its, cls.train_count, cls.response_count)
)
cls.test_responses = mm.zeros(
(cls.its, cls.test_count, cls.response_count)
cls.train_features, cls.test_features = cls.sampler.features()
cls.train_count = cls.train_features.shape[0]
cls.test_count = cls.test_features.shape[0]

cls.train_responses_list = list()
cls.test_responses_list = list()
for _ in range(cls.its):
train_responses, test_responses = cls.sampler.sample()
cls.train_responses_list.append(train_responses)
cls.test_responses_list.append(test_responses)

cls.nbrs_lookup = NN_Wrapper(
cls.train_features, cls.nn_count, **_basic_nn_kwarg_options[0]
)
cls.batch_indices_list = list()
cls.batch_nn_indices_list = list()
for _ in range(cls.its):
batch_indices, batch_nn_indices = sample_batch(
cls.nbrs_lookup, cls.batch_count, cls.train_count
)
cls.batch_indices_list.append(batch_indices)
cls.batch_nn_indices_list.append(batch_nn_indices)

cls.batch_crosswise_diffs_list = list()
cls.batch_pairwise_diffs_list = list()
cls.batch_targets_list = list()
cls.batch_nn_targets_list = list()
for i in range(cls.its):
ys = benchmark_sample_full(
cls.gp, cls.test_features, cls.train_features
batch_crosswise_diffs = crosswise_tensor(
cls.train_features,
cls.train_features,
cls.batch_indices_list[i],
cls.batch_nn_indices_list[i],
)
batch_pairwise_diffs = pairwise_tensor(
cls.train_features,
cls.batch_nn_indices_list[i],
)
cls.train_responses = mm.assign(
cls.train_responses,
ys[cls.test_count :, :],
i,
slice(None),
slice(None),
batch_targets = _consistent_chunk_tensor(
cls.train_responses_list[i][cls.batch_indices_list[i]]
)
cls.test_responses = mm.assign(
cls.test_responses,
ys[: cls.test_count, :],
i,
slice(None),
slice(None),
batch_nn_targets = _consistent_chunk_tensor(
cls.train_responses_list[i][cls.batch_nn_indices_list[i]]
)

cls.batch_crosswise_diffs_list.append(batch_crosswise_diffs)
cls.batch_pairwise_diffs_list.append(batch_pairwise_diffs)
cls.batch_targets_list.append(batch_targets)
cls.batch_nn_targets_list.append(batch_nn_targets)

def _optim_chassis(
self,
muygps,
name,
itr,
nbrs_lookup,
batch_count,
loss_fn,
opt_fn,
opt_kwargs,
loss_kwargs=dict(),
) -> float:
batch_indices, batch_nn_indices = sample_batch(
nbrs_lookup, batch_count, self.train_count
)
batch_crosswise_diffs = crosswise_tensor(
self.train_features,
self.train_features,
batch_indices,
batch_nn_indices,
)
batch_pairwise_diffs = pairwise_tensor(
self.train_features,
batch_nn_indices,
)
batch_targets = _consistent_chunk_tensor(
self.train_responses[itr, batch_indices, :]
)
batch_nn_targets = _consistent_chunk_tensor(
self.train_responses[itr, batch_nn_indices, :]
)
muygps = opt_fn(
muygps,
batch_targets,
batch_nn_targets,
batch_crosswise_diffs,
batch_pairwise_diffs,
self.batch_targets_list[itr],
self.batch_nn_targets_list[itr],
self.batch_crosswise_diffs_list[itr],
self.batch_pairwise_diffs_list[itr],
loss_fn=loss_fn,
loss_kwargs=loss_kwargs,
**opt_kwargs,
verbose=False, # TODO True,
verbose=False,
)
estimate = muygps.kernel._hyperparameters[name]()
print(
f"\toptimizes {name} with {estimate} (target {self.params[name]()})"
f" -- sq rel err = {_sq_rel_err(self.params[name](), estimate)}"
)
return _sq_rel_err(self.params[name](), estimate)

def _check_ndarray(self, *args, **kwargs):
Expand Down
2 changes: 1 addition & 1 deletion MuyGPyS/_test/utils.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -35,7 +35,7 @@
{
"random_state": 1,
"init_points": 3,
"n_iter": 10,
"n_iter": 7,
"allow_duplicate_points": True,
},
],
Expand Down
52 changes: 23 additions & 29 deletions experimental/GPsolvetensors.ipynb
View file

Large diffs are not rendered by default.

Loading

0 comments on commit 1047098

Please sign in to comment.