def test_datasetprovider_awkward_exporter_single_tensor_same_shape():
"""
The core assumption for simplest cases is that one
system will be featurized as a single tensor, and that
all the tensors will be of the same shape across systems.
If that's the case:
- DatasetProvider.to_numpy() will work and will return
a X, y tuple of arrays
- DatasetProvider.to_dict_of_arrays() will work and will
return a dict of arrays with X, y keys.
Note that `.to_numpy()` won't work if the core assumptions
are broken. For those cases, `.to_dict_of_arrays()` is
recommended instead.
"""
from kinoml.core.ligands import RDKitLigand
from kinoml.features.ligand import MorganFingerprintFeaturizer
from kinoml.features.core import Concatenated, TupleOfArrays
import awkward as ak
conditions = AssayConditions()
systems = [
System([RDKitLigand.from_smiles(smi)]) for smi in ("CCCCC", "CCCCCCCC")
]
measurements = [
BaseMeasurement(50, conditions=conditions, system=systems[0]),
BaseMeasurement(30, conditions=conditions, system=systems[1]),
]
dataset = DatasetProvider(measurements=measurements)
featurizer1 = MorganFingerprintFeaturizer(radius=2, nbits=512)
featurizer2 = MorganFingerprintFeaturizer(radius=2, nbits=1024)
concatenate = Concatenated([featurizer1, featurizer2], axis=1)
aggregated = TupleOfArrays([concatenate])
aggregated.featurize(systems)
for system in systems:
assert system.featurizations["last"][0].shape[0] == (1024 + 512)
# With a single tensor per system, we build a unified X tensor
# First dimension in X and y must match
X, y = dataset.to_numpy()
# This extra dimension here V
# comes from the TupleOfArrays aggregation
assert X.shape[:3] == (2, 1, (1024 + 512))
assert X.shape[0] == y.shape[0]
# With dict_of_arrays and single tensor per system,
# the behavior is essentially the same
(Xa, ), ya = dataset.to_awkward()
assert ak.to_numpy(Xa).shape == (2, (1024 + 512))
assert ak.to_numpy(ya).shape == (2, )
def test_measurements():
from kinoml.core.measurements import BaseMeasurement, PercentageDisplacementMeasurement
from kinoml.core.conditions import AssayConditions
from kinoml.core.components import MolecularComponent
from kinoml.core.systems import System
conditions = AssayConditions()
system = System([MolecularComponent()])
measurement = BaseMeasurement(50, conditions=conditions, system=system)
assert isinstance(measurement, BaseMeasurement)
assert measurement == BaseMeasurement(50, conditions=conditions, system=system)
assert measurement != BaseMeasurement(10, conditions=conditions, system=system)
def test_datasetprovider():
conditions = AssayConditions()
measurements = [
BaseMeasurement(50,
conditions=conditions,
system=System([MolecularComponent()])),
BaseMeasurement(30,
conditions=conditions,
system=System([MolecularComponent()])),
]
dataset = DatasetProvider(measurements=measurements)
assert len(dataset.conditions) == 1
assert next(iter(dataset.conditions)) == conditions
def test_datasetprovider():
from kinoml.datasets.core import DatasetProvider
from kinoml.core.systems import System
from kinoml.core.components import MolecularComponent
from kinoml.core.measurements import BaseMeasurement
from kinoml.core.conditions import AssayConditions
from kinoml.features.core import BaseFeaturizer
conditions = AssayConditions()
measurements = [
BaseMeasurement(50,
conditions=conditions,
system=System([MolecularComponent()])),
BaseMeasurement(30,
conditions=conditions,
system=System([MolecularComponent()])),
]
provider = DatasetProvider(measurements=measurements,
featurizers=[BaseFeaturizer()])
assert len(provider.conditions) == 1
assert next(iter(provider.conditions)) == conditions
def test_datasetprovider_awkward_exporter_multiple_subtensors():
"""
When we use an aggregator like TupleOfArrays, which
breaks the one system -> one tensor assumption,
we need to use the `.dict_of_arrays()` exporter.
This creates a dictionary of arrays, where each key is
autogenerated like `X_s{int}_a{int}`, where `s` is the
system index and `a` is the array index.
"""
from kinoml.core.ligands import RDKitLigand
from kinoml.features.ligand import MorganFingerprintFeaturizer
from kinoml.features.core import TupleOfArrays
import awkward as ak
conditions = AssayConditions()
systems = [
System([RDKitLigand.from_smiles(smi)]) for smi in ("CCCCC", "CCCCCCCC")
]
measurements = [
BaseMeasurement(50, conditions=conditions, system=systems[0]),
BaseMeasurement(30, conditions=conditions, system=systems[1]),
]
dataset = DatasetProvider(measurements=measurements)
featurizer1 = MorganFingerprintFeaturizer(radius=2, nbits=512)
featurizer2 = MorganFingerprintFeaturizer(radius=2, nbits=1024)
aggregated = TupleOfArrays([featurizer1, featurizer2])
aggregated.featurize(dataset.systems)
for system in systems:
assert len(system.featurizations["last"]) == 2
assert system.featurizations["last"][0].shape[0] == 512
assert system.featurizations["last"][1].shape[0] == 1024
(x1, x2), y = dataset.to_awkward()
assert len(x1) == len(x2) == len(y)
assert ak.to_numpy(x1).shape == (2, 512)
assert ak.to_numpy(x2).shape == (2, 1024)
def test_datasetprovider_awkward_exporter_single_tensor_different_shapes():
"""
When a featurizer returns arrays of different shape for each
system, one can either choose between:
A. Pad them to the same dimension with PadFeaturizer,
and apply the Concatenated aggregator.
B. Keep them separate and export them with `.dict_of_arrays()`,
This creates a dictionary of arrays, where each key is
autogenerated like `X_s{int}`, where `s` is the
system index.
"""
from kinoml.core.ligands import RDKitLigand
from kinoml.features.ligand import OneHotSMILESFeaturizer
from kinoml.features.core import TupleOfArrays
import awkward as ak
conditions = AssayConditions()
smiles = ("CCCCC", "CCCCCCCC")
systems = [System([RDKitLigand.from_smiles(smi)]) for smi in smiles]
measurements = [
BaseMeasurement(50, conditions=conditions, system=systems[0]),
BaseMeasurement(30, conditions=conditions, system=systems[1]),
]
dataset = DatasetProvider(measurements=measurements)
featurizer = OneHotSMILESFeaturizer()
aggregated = TupleOfArrays([featurizer])
aggregated.featurize(dataset.systems)
for system, smi in zip(systems, smiles):
assert system.featurizations["last"][0].shape == (53, len(smi))
# X is returned as single-item list thanks to TupleOfArrays
(X, ), y = dataset.to_awkward()
assert X.type.length == len(y)
assert ak.to_numpy(X[0]).shape == (53, len(smiles[0]))
assert ak.to_numpy(X[1]).shape == (53, len(smiles[1]))
def test_datasetprovider_exporter_single_tensor_different_shapes():
"""
When a featurizer returns arrays of different shape for each
system, one can either choose between:
A. Pad them to the same dimension with PadFeaturizer,
and apply the Concatenated aggregator.
B. Keep them separate and export them with `.dict_of_arrays()`,
This creates a dictionary of arrays, where each key is
autogenerated like `X_s{int}`, where `s` is the
system index.
"""
from kinoml.core.ligands import RDKitLigand
from kinoml.features.ligand import OneHotSMILESFeaturizer
conditions = AssayConditions()
smiles = ("CCCCC", "CCCCCCCC")
systems = [System([RDKitLigand.from_smiles(smi)]) for smi in smiles]
measurements = [
BaseMeasurement(50, conditions=conditions, system=systems[0]),
BaseMeasurement(30, conditions=conditions, system=systems[1]),
]
dataset = DatasetProvider(measurements=measurements)
featurizer = OneHotSMILESFeaturizer()
featurizer.featurize(dataset.systems)
for system, smi in zip(systems, smiles):
assert system.featurizations["last"].shape == (53, len(smi))
arrays = dataset.to_dict_of_arrays()
X_keys = [k for k in arrays.keys() if k.startswith("X")]
assert sorted(X_keys) == ["X_s0_", "X_s1_"]
for X_key, smi in zip(X_keys, smiles):
assert arrays[X_key].shape == (53, len(smi))