Composite Sampler

src/confopt/enums.py

@@ -21,6 +21,7 @@ class SamplerType(Enum):
     GDAS = "gdas"
     SNAS = "snas"
     REINMAX = "reinmax"
+    COMPOSITE = "composite"
 
     def __str__(self) -> str:
         return self.value

src/confopt/oneshot/archsampler/__init__.py

@@ -1,4 +1,5 @@
 from .base_sampler import BaseSampler
+from .composite_sampler import CompositeSampler
 from .darts.sampler import DARTSSampler
 from .drnas.sampler import DRNASSampler
 from .gdas.sampler import GDASSampler
@@ -12,4 +13,5 @@
     "GDASSampler",
     "SNASSampler",
     "ReinMaxSampler",
+    "CompositeSampler",
 ]

src/confopt/oneshot/archsampler/composite_sampler.py

@@ -0,0 +1,41 @@
+from __future__ import annotations
+
+import torch
+
+from confopt.oneshot.archsampler import BaseSampler
+from confopt.oneshot.base import OneShotComponent
+
+
+class CompositeSampler(OneShotComponent):
+    def __init__(
+        self,
+        arch_samplers: list[BaseSampler],
+        arch_parameters: list[torch.Tensor],
+    ) -> None:
+        super().__init__()
+        self.arch_samplers = arch_samplers
+        self.arch_parameters = arch_parameters
+
+        # get sample frequency from the samplers
+        self.sample_frequency = arch_samplers[0].sample_frequency
+        for sampler in arch_samplers:
+            assert (
+                self.sample_frequency == sampler.sample_frequency
+            ), "All samplers must have the same sample frequency"
+
+    def sample(self, alpha: torch.Tensor) -> torch.Tensor:
+        sampled_alphas = alpha
+        for sampler in self.arch_samplers:
+            sampled_alphas = sampler.sample(sampled_alphas)
+
+        return sampled_alphas
+
+    def new_epoch(self) -> None:
+        super().new_epoch()
+        for sampler in self.arch_samplers:
+            sampler.new_epoch()
+
+    def new_step(self) -> None:
+        super().new_step()
+        for sampler in self.arch_samplers:
+            sampler.new_step()

src/confopt/profile/__init__.py

@@ -1,5 +1,6 @@
 from .base import BaseProfile
 from .profiles import (
+    CompositeProfile,
     DARTSProfile,
     DiscreteProfile,
     DRNASProfile,
@@ -16,4 +17,5 @@
     "SNASProfile",
     "DiscreteProfile",
     "ReinMaxProfile",
+    "CompositeProfile",
 ]

src/confopt/profile/profiles.py

@@ -3,6 +3,8 @@
 from abc import ABC
 from typing import Any
 
+from typing_extensions import override
+
 from confopt.enums import SamplerType, SearchSpaceType
 from confopt.searchspace.darts.core.genotypes import DARTSGenotype
 from confopt.utils import get_num_classes
@@ -402,6 +404,147 @@ def configure_sampler(self, **kwargs) -> None:  # type: ignore
         super().configure_sampler(**kwargs)
 
 
+class CompositeProfile(BaseProfile, ABC):
+    SAMPLER_TYPE = SamplerType.COMPOSITE
+
+    def __init__(
+        self,
+        searchspace_type: str | SearchSpaceType,
+        samplers: list[str | SamplerType],
+        epochs: int,
+        # GDAS configs
+        tau_min: float = 0.1,
+        tau_max: float = 10,
+        # SNAS configs
+        temp_init: float = 1.0,
+        temp_min: float = 0.03,
+        temp_annealing: bool = True,
+        **kwargs: Any,
+    ) -> None:
+        self.samplers = []
+        for sampler in samplers:
+            if isinstance(sampler, str):
+                sampler = SamplerType(sampler)
+            self.samplers.append(sampler)
+
+        self.tau_min = tau_min
+        self.tau_max = tau_max
+
+        self.temp_init = temp_init
+        self.temp_min = temp_min
+        self.temp_annealing = temp_annealing
+
+        super().__init__(  # type: ignore
+            self.SAMPLER_TYPE,
+            searchspace_type,
+            epochs,
+            **kwargs,
+        )
+
+    def _initialize_sampler_config(self) -> None:
+        """Initializes the sampler configuration for Composite samplers.
+
+        The sampler configuration includes the sample frequency and the architecture
+        combine function.
+
+        Args:
+            None
+
+        Returns:
+            None
+        """
+        self.sampler_config: dict[int, dict] = {}  # type: ignore
+        for i, sampler in enumerate(self.samplers):
+            if sampler in [SamplerType.DARTS, SamplerType.DRNAS]:
+                config = {
+                    "sampler_type": sampler,
+                    "sample_frequency": self.sampler_sample_frequency,
+                    "arch_combine_fn": self.sampler_arch_combine_fn,
+                }
+            elif sampler in [SamplerType.GDAS, SamplerType.REINMAX]:
+                config = {
+                    "sampler_type": sampler,
+                    "sample_frequency": self.sampler_sample_frequency,
+                    "arch_combine_fn": self.sampler_arch_combine_fn,
+                    "tau_min": self.tau_min,
+                    "tau_max": self.tau_max,
+                }
+            elif sampler == SamplerType.SNAS:
+                config = {
+                    "sampler_type": sampler,
+                    "sample_frequency": self.sampler_sample_frequency,
+                    "arch_combine_fn": self.sampler_arch_combine_fn,
+                    "temp_init": self.temp_init,
+                    "temp_min": self.temp_min,
+                    "temp_annealing": self.temp_annealing,
+                    "total_epochs": self.epochs,
+                }
+            else:
+                raise AttributeError(f"Illegal sampler type {sampler} provided!")
+
+            self.sampler_config[i] = config
+
+    @override
+    def configure_sampler(  # type: ignore[override]
+        self, sampler_config_map: dict[int, dict]
+    ) -> None:
+        """Configures the sampler settings based on the provided configurations.
+
+        Args:
+            sampler_config_map (dict[int, dict]): A dictionary where each key is an \
+                integer representing the order of the sampler (zero-indexed), and \
+                each value is a dictionary containing the configuration parameters \
+                for that sampler.
+
+            The inner configuration dictionary can contain different sets of keys
+            depending on the type of sampler being configured. The available keys
+            include:
+
+                Generic Configurations:
+                    - sample_frequency (str): The rate at which samples should be taken.
+                    - arch_combine_fn (str): Function to combine architectures.
+                      For FairDARTS, set this to 'sigmoid'. Default is 'default'.
+
+                GDAS-specific Configurations:
+                    - tau_min (float): Minimum temperature for sampling.
+                    - tau_max (float): Maximum temperature for sampling.
+
+                SNAS-specific Configurations:
+                    - temp_init (float): Initial temperature for sampling.
+                    - temp_min (float): Minimum temperature for sampling.
+                    - temp_annealing (bool): Whether to apply temperature annealing.
+                    - total_epochs (int): Total number of training epochs.
+
+        The specific keys required in the dictionary depend on the type of
+        sampler being used.
+        Please make sure that the sample frequency of all the configurations are same.
+        Each configuration is validated, and an error is raised if unknown keys are
+        provided.
+
+        Raises:
+            ValueError: If an unrecognized configuration key is detected.
+
+        Returns:
+            None
+        """
+        assert self.sampler_config is not None
+
+        for idx in sampler_config_map:
+            for config_key in sampler_config_map[idx]:
+                assert idx in self.sampler_config
+                exists = False
+                sampler_type = self.sampler_config[idx]["sampler_type"]
+                if config_key in self.sampler_config[idx]:
+                    exists = True
+                    self.sampler_config[idx][config_key] = sampler_config_map[idx][
+                        config_key
+                    ]
+                assert exists, (
+                    f"{config_key} is not a valid configuration for {sampler_type}",
+                    "sampler inside composite sampler",
+                )
+
+
 class DiscreteProfile:
     def __init__(
         self,

src/confopt/train/experiment.py

@@ -40,6 +40,7 @@
 )
 from confopt.oneshot.archsampler import (
     BaseSampler,
+    CompositeSampler,
     DARTSSampler,
     DRNASSampler,
     GDASSampler,
@@ -374,17 +375,36 @@ def _set_sampler(
         config: dict,
     ) -> None:
         arch_params = self.search_space.arch_parameters
-        self.sampler: BaseSampler | None = None
-        if sampler == SamplerType.DARTS:
-            self.sampler = DARTSSampler(**config, arch_parameters=arch_params)
-        elif sampler == SamplerType.DRNAS:
-            self.sampler = DRNASSampler(**config, arch_parameters=arch_params)
-        elif sampler == SamplerType.GDAS:
-            self.sampler = GDASSampler(**config, arch_parameters=arch_params)
-        elif sampler == SamplerType.SNAS:
-            self.sampler = SNASSampler(**config, arch_parameters=arch_params)
-        elif sampler == SamplerType.REINMAX:
-            self.sampler = ReinMaxSampler(**config, arch_parameters=arch_params)
+        self.sampler: BaseSampler | CompositeSampler | None = None
+
+        def _get_sampler_class(sampler: SamplerType) -> Callable:
+            if sampler == SamplerType.DARTS:
+                return DARTSSampler
+            if sampler == SamplerType.DRNAS:
+                return DRNASSampler
+            if sampler == SamplerType.GDAS:
+                return GDASSampler
+            if sampler == SamplerType.SNAS:
+                return SNASSampler
+            if sampler == SamplerType.REINMAX:
+                return ReinMaxSampler
+
+            raise ValueError(f"Illegal sampler {sampler} provided")
+
+        if sampler == SamplerType.COMPOSITE:
+            sub_samplers: list[BaseSampler] = []
+            for _, sampler_config in config.items():
+                sampler_type = sampler_config["sampler_type"]
+                del sampler_config["sampler_type"]
+                sampler_component = _get_sampler_class(sampler_type)(
+                    **sampler_config, arch_parameters=arch_params
+                )
+                sub_samplers.append(sampler_component)
+            self.sampler = CompositeSampler(sub_samplers, arch_parameters=arch_params)
+        else:
+            self.sampler = _get_sampler_class(sampler)(
+                **config, arch_parameters=arch_params
+            )
 
     def _set_perturbator(
         self,

src/confopt/train/search_space_handler.py

@@ -2,7 +2,12 @@
 
 import torch
 
-from confopt.oneshot.archsampler import BaseSampler, DARTSSampler, GDASSampler
+from confopt.oneshot.archsampler import (
+    BaseSampler,
+    CompositeSampler,
+    DARTSSampler,
+    GDASSampler,
+)
 from confopt.oneshot.dropout import Dropout
 from confopt.oneshot.lora_toggler import LoRAToggler
 from confopt.oneshot.partial_connector import PartialConnector
@@ -23,7 +28,7 @@
 class SearchSpaceHandler:
     def __init__(
         self,
-        sampler: BaseSampler,
+        sampler: BaseSampler | CompositeSampler,
         edge_normalization: bool = False,
         partial_connector: PartialConnector | None = None,
         perturbation: BasePerturbator | None = None,

tests/test_sampler.py

@@ -8,6 +8,7 @@
     DRNASSampler,
     GDASSampler,
     ReinMaxSampler,
+    CompositeSampler,
 )
 
 
@@ -111,5 +112,59 @@ def test_post_sample_fn_sigmoid(self) -> None:
         _test_arch_combine_fn_default(sampler, alphas)
 
 
+class TestCompositeSampler(unittest.TestCase):
+    def test_post_sample_fn_default(self) -> None:
+        alphas = [torch.randn(14, 8), torch.randn(14, 8)]
+        inner_samplers = [
+            DRNASSampler(alphas, sample_frequency="epoch", arch_combine_fn="default"),
+            GDASSampler(alphas, sample_frequency="epoch", arch_combine_fn="default"),
+            DARTSSampler(alphas, sample_frequency="epoch", arch_combine_fn="default"),
+        ]
+        sampler = CompositeSampler(inner_samplers, alphas)
+        _test_arch_combine_fn_default(sampler, alphas)
+
+    def test_post_sample_fn_sigmoid(self) -> None:
+        alphas = [torch.randn(14, 8), torch.randn(14, 8)]
+        inner_samplers = [
+            DRNASSampler(alphas, sample_frequency="epoch", arch_combine_fn="sigmoid"),
+            GDASSampler(alphas, sample_frequency="epoch", arch_combine_fn="sigmoid"),
+            DARTSSampler(alphas, sample_frequency="epoch", arch_combine_fn="sigmoid"),
+        ]
+        sampler = CompositeSampler(inner_samplers, alphas)
+        _test_arch_combine_fn_sigmoid(sampler, alphas)
+
+    def test_post_sample_fn_mixed(self) -> None:
+        alphas = [torch.randn(14, 8), torch.randn(14, 8)]
+
+        # Case 1
+        inner_samplers = [
+            DRNASSampler(alphas, sample_frequency="epoch", arch_combine_fn="default"),
+            GDASSampler(alphas, sample_frequency="epoch", arch_combine_fn="sigmoid"),
+            DARTSSampler(alphas, sample_frequency="epoch", arch_combine_fn="default"),
+        ]
+        sampler = CompositeSampler(inner_samplers, alphas)
+        # The last sampler's combine function matters,
+        # and if its GDAS/Reinmax, it would get ignored
+        _test_arch_combine_fn_default(sampler, alphas)
+
+        # Case 2
+        inner_samplers = [
+            DRNASSampler(alphas, sample_frequency="epoch", arch_combine_fn="default"),
+            DARTSSampler(alphas, sample_frequency="epoch", arch_combine_fn="sigmoid"),
+            GDASSampler(alphas, sample_frequency="epoch", arch_combine_fn="sigmoid"),
+        ]
+        sampler = CompositeSampler(inner_samplers, alphas)
+        _test_arch_combine_fn_default(sampler, alphas)
+
+        # Case 3
+        inner_samplers = [
+            DRNASSampler(alphas, sample_frequency="epoch", arch_combine_fn="default"),
+            GDASSampler(alphas, sample_frequency="epoch", arch_combine_fn="sigmoid"),
+            DARTSSampler(alphas, sample_frequency="epoch", arch_combine_fn="sigmoid"),
+        ]
+        sampler = CompositeSampler(inner_samplers, alphas)
+        _test_arch_combine_fn_sigmoid(sampler, alphas)
+
+
 if __name__ == "__main__":
     unittest.main()