libinephany 0.14.0__py3-none-any.whl → 0.15.0__py3-none-any.whl

libinephany/pydantic_models/configs/hyperparameter_configs.py

@@ -4,7 +4,7 @@
 #
 # ======================================================================================================================
 
-from typing import Any
+from typing import Any, cast
 
 from pydantic import BaseModel, ConfigDict, ValidationError, field_serializer, field_validator, model_validator
 
@@ -232,6 +232,23 @@ class BatchSizeHParamConfig(HParamConfig):
     sample_discrete_values: list[float | int] | None = None
 
 
+class GradientAccumulationHParamConfig(HParamConfig):
+    max_hparam_value: float | int = 64
+    min_hparam_value: float | int = 1
+    hparam_dtype: type[float | int] = int
+    initial_value: int = 1
+    initial_delta: float = 0.0
+    scale: float = 1.0
+
+    sampler: str = "DiscreteRangeSampler"
+    sample_initial_values: bool = False
+    sample_lower_bound: int = 1
+    sample_upper_bound: int = 64
+    sample_step: int = 1
+    sample_discrete_values: list[float | int] | None = None
+    force_limit: float | int = 64
+
+
 class EpochsHParamConfig(HParamConfig):
     max_hparam_value: float | int = 16
     min_hparam_value: float | int = 1
@@ -289,6 +306,7 @@ class HParamConfigs(BaseModel):
     sgd_momentum_config: HParamConfig = SGDMomentumHParamConfig()
 
     batch_size_config: HParamConfig = BatchSizeHParamConfig()
+    gradient_accumulation_config: GradientAccumulationHParamConfig = GradientAccumulationHParamConfig()
     epochs_config: HParamConfig = EpochsHParamConfig()
     token_config: HParamConfig = TokensHParamConfig()
     samples_config: HParamConfig = SamplesHParamConfig()
@@ -351,6 +369,9 @@ class HParamConfigs(BaseModel):
             case AgentTypes.BatchSize:
                 self.batch_size_config = hparam_config
 
+            case AgentTypes.GradientAccumulationAgent:
+                self.gradient_accumulation_config = cast(GradientAccumulationHParamConfig, hparam_config)
+
             case AgentTypes.Epochs:
                 self.epochs_config = hparam_config
 
@@ -400,6 +421,9 @@ class HParamConfigs(BaseModel):
             case AgentTypes.BatchSize:
                 return self.batch_size_config
 
+            case AgentTypes.GradientAccumulationAgent:
+                return self.gradient_accumulation_config
+
             case AgentTypes.Epochs:
                 return self.epochs_config
 

libinephany/pydantic_models/schemas/inner_task_profile.py

@@ -24,6 +24,7 @@ class InnerTaskProfile(BaseModel):
     number_of_layers: int
     observation_space_sizes: dict[str, int]
     action_space_sizes: dict[str, int]
+    number_of_parameters: int
     vram_usage: float
     idle_vram_usage: float
 
@@ -139,6 +140,40 @@ class InnerTaskProfiles(BaseModel):
 
         return sum(self.compiled_action_sizes.values())
 
+    @property
+    def max_total_observation_size(self) -> int:
+        """
+        :return: The summed observation size of all agents with the task that has the most layers.
+        """
+
+        if not self.profiles:
+            raise ValueError(
+                "No profiles to calculate max total observation size. Ensure profiles have been "
+                "added before executing the training loop"
+            )
+
+        largest_task_name = max(self.profiles, key=lambda k: self.profiles[k].number_of_layers)
+        largest_task = self.profiles[largest_task_name]
+
+        return sum(largest_task.observation_space_sizes.values())
+
+    @property
+    def max_total_action_size(self) -> int:
+        """
+        :return: The summed action size of all agents with the task that has the most layers.
+        """
+
+        if not self.profiles:
+            raise ValueError(
+                "No profiles to calculate max total action size. Ensure profiles have been "
+                "added before executing the training loop"
+            )
+
+        largest_task_name = max(self.profiles, key=lambda k: self.profiles[k].number_of_layers)
+        largest_task = self.profiles[largest_task_name]
+
+        return sum(largest_task.action_space_sizes.values())
+
     @staticmethod
     def _compile_gym_space_sizes(spaces: dict[str, dict[str, int]]) -> dict[str, int]:
         """
@@ -206,6 +241,7 @@ class InnerTaskProfiles(BaseModel):
         number_of_layers: int,
         observation_space_sizes: dict[str, int],
         action_space_sizes: dict[str, int],
+        number_of_parameters: int,
         vram_usage: float,
         idle_vram_usage: float,
     ) -> None:
@@ -226,6 +262,7 @@ class InnerTaskProfiles(BaseModel):
             number_of_layers=number_of_layers,
             observation_space_sizes=observation_space_sizes,
             action_space_sizes=action_space_sizes,
+            number_of_parameters=number_of_parameters,
             vram_usage=vram_usage,
             idle_vram_usage=idle_vram_usage,
         )

libinephany/pydantic_models/states/hyperparameter_states.py

@@ -20,6 +20,7 @@ from libinephany.utils.constants import (
     DROPOUT,
     EPOCHS,
     GRAD_NORM_CLIP,
+    GRADIENT_ACCUMULATION,
     LEARNING_RATE,
     SAMPLES,
     SGD_MOMENTUM,
@@ -60,6 +61,7 @@ class UpdateCallbacks(BaseModel):
     sgd_momentum: Callable[..., None]
 
     batch_size: Callable[..., None] | None
+    gradient_accumulation: Callable[..., None] | None
     epochs: Callable[..., None] | None
 
     def __getitem__(self, item: str) -> Callable[..., None] | None:
@@ -457,6 +459,7 @@ class ParameterGroupHParams(HyperparameterContainer):
 class GlobalHParams(HyperparameterContainer):
 
     batch_size: Hyperparameter
+    gradient_accumulation: Hyperparameter
     epochs: Hyperparameter
     tokens: Hyperparameter
     samples: Hyperparameter
@@ -550,6 +553,14 @@ class HyperparameterStates(BaseModel):
         """
         return self.global_hparams.batch_size
 
+    @computed_field  # type: ignore[misc]
+    @property
+    def gradient_accumulation(self) -> Hyperparameter:
+        """
+        :return: The gradient accumulation steps of the inner model.
+        """
+        return self.global_hparams.gradient_accumulation
+
     @computed_field  # type: ignore[misc]
     @property
     def epochs(self) -> Hyperparameter:
@@ -676,6 +687,7 @@ class HyperparameterStates(BaseModel):
 
         return {
             BATCH_SIZE: hparam_configs.batch_size_config,
+            GRADIENT_ACCUMULATION: hparam_configs.gradient_accumulation_config,
             EPOCHS: hparam_configs.epochs_config,
             TOKENS: hparam_configs.token_config,
             SAMPLES: hparam_configs.samples_config,

libinephany/utils/constants.py

@@ -21,6 +21,7 @@ ADAM_BETA_TWO = "adam_beta_two"
 ADAM_EPS = "adam_eps"
 SGD_MOMENTUM = "sgd_momentum"
 BATCH_SIZE = "batch_size"
+GRADIENT_ACCUMULATION = "gradient_accumulation"
 EPOCHS = "epochs"
 TOKENS = "tokens"
 SAMPLES = "samples"
@@ -41,6 +42,7 @@ AGENT_PREFIX_EPS = "adam-eps"
 AGENT_PREFIX_SGD_MOMENTUM = "sgd-momentum"
 
 AGENT_BATCH_SIZE = "batch-size"
+AGENT_GRADIENT_ACCUMULATION = "gradient-accumulation"
 
 AGENT_BANDIT_SUFFIX = "bandit-agent"
 
@@ -53,6 +55,7 @@ AGENT_TYPES = [
     ADAM_BETA_TWO,
     ADAM_EPS,
     SGD_MOMENTUM,
+    GRADIENT_ACCUMULATION,
 ]
 SUFFIXES = [AGENT_BANDIT_SUFFIX]
 PREFIXES = [
@@ -64,6 +67,7 @@ PREFIXES = [
     AGENT_PREFIX_BETA_TWO,
     AGENT_PREFIX_EPS,
     AGENT_PREFIX_SGD_MOMENTUM,
+    AGENT_GRADIENT_ACCUMULATION,
 ]
 PREFIXES_TO_HPARAMS = {
     AGENT_PREFIX_LR: LEARNING_RATE,
@@ -74,4 +78,5 @@ PREFIXES_TO_HPARAMS = {
     AGENT_PREFIX_BETA_TWO: ADAM_BETA_TWO,
     AGENT_PREFIX_EPS: ADAM_EPS,
     AGENT_PREFIX_SGD_MOMENTUM: SGD_MOMENTUM,
+    AGENT_GRADIENT_ACCUMULATION: GRADIENT_ACCUMULATION,
 }

libinephany/utils/enums.py

@@ -14,6 +14,7 @@ from libinephany.utils.constants import (
     DROPOUT,
     EPOCHS,
     GRAD_NORM_CLIP,
+    GRADIENT_ACCUMULATION,
     LEARNING_RATE,
     SAMPLES,
     SGD_MOMENTUM,
@@ -69,6 +70,7 @@ class AgentTypes(EnumWithIndices):
     AdamBetaTwoAgent = ADAM_BETA_TWO
     AdamEpsAgent = ADAM_EPS
     SGDMomentumAgent = SGD_MOMENTUM
+    GradientAccumulationAgent = GRADIENT_ACCUMULATION
 
     # Deprecated or Non-Agent
     BatchSize = BATCH_SIZE

libinephany/utils/samplers.py

@@ -6,6 +6,7 @@
 
 import math
 from abc import abstractmethod
+from typing import Any
 
 import numpy as np
 
@@ -20,14 +21,14 @@ from libinephany.utils import random_seeds
 
 class Sampler:
 
-    def __init__(self, lower_bound: float | int, upper_bound: float | int, **kwargs) -> None:
+    def __init__(self, lower_bound: float | int | None, upper_bound: float | int | None, **kwargs) -> None:
         """
         :param lower_bound: Lower bound of the distribution to sample from.
         :param upper_bound: Upper bound of the distribution to sample from.
         :param kwargs: Miscellaneous keyword arguments.
         """
 
-        if upper_bound < lower_bound:
+        if upper_bound is not None and lower_bound is not None and upper_bound < lower_bound:
             upper_bound, lower_bound = lower_bound, upper_bound
 
         self.lower_bound = lower_bound
@@ -37,11 +38,11 @@ class Sampler:
         self,
         number_of_samples: int = 1,
         seed: int | None = None,
-    ) -> np.ndarray | float:
+    ) -> list[Any] | np.ndarray | float:
         """
         :param number_of_samples: Number of samples to make.
         :param seed: Random seed to use for distribution sampling.
-        :return: Array of sampled values or a single sampled value.
+        :return: List/Array of sampled values or a single sampled value.
         """
 
         if seed is not None:
@@ -49,10 +50,16 @@ class Sampler:
 
         sample = self.sample(number_of_samples=number_of_samples)
 
-        return sample.item() if sample.size == 1 else sample
+        if isinstance(sample, np.ndarray) and sample.size == 1:
+            return sample.item()
+
+        elif isinstance(sample, list) and len(sample) == 1:
+            return sample[0]
+
+        return sample
 
     @abstractmethod
-    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray:
+    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray | list[Any]:
         """
         :param number_of_samples: Number of samples to make.
         :param kwargs: Miscellaneous keyword arguments.
@@ -64,7 +71,7 @@ class Sampler:
 
 class LogUniformSampler(Sampler):
 
-    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray:
+    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray | list[Any]:
         """
         :param number_of_samples: Number of samples to make.
         :param kwargs: Miscellaneous keyword arguments.
@@ -78,7 +85,7 @@ class LogUniformSampler(Sampler):
 
 class UniformSampler(Sampler):
 
-    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray:
+    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray | list[Any]:
         """
         :param number_of_samples: Number of samples to make.
         :param kwargs: Miscellaneous keyword arguments.
@@ -113,7 +120,7 @@ class DiscreteRangeSampler(Sampler):
         self.step = step
         self.sample_dtype = sample_dtype
 
-    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray:
+    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray | list[Any]:
         """
         :param number_of_samples: Number of samples to make.
         :param kwargs: Miscellaneous keyword arguments.
@@ -160,7 +167,7 @@ class RoundRobinSampler(Sampler):
 
         return magnitude_list
 
-    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray:
+    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray | list[Any]:
         """
         :param number_of_samples: Number of samples to make.
         :param kwargs: Miscellaneous keyword arguments.
@@ -187,8 +194,8 @@ class DiscreteValueSampler(Sampler):
 
     def __init__(
         self,
-        discrete_values: list[float | int],
-        sample_dtype: type[np.generic | float | int] = np.float64,
+        discrete_values: list[float | int | str],
+        sample_dtype: type[np.generic | float | int | str] = np.float64,
         **kwargs,
     ) -> None:
         """
@@ -196,12 +203,19 @@ class DiscreteValueSampler(Sampler):
         :param kwargs: Miscellaneous keyword arguments.
         """
 
-        super().__init__(lower_bound=min(discrete_values), upper_bound=max(discrete_values))
+        if sample_dtype == str and any(isinstance(value, str) for value in discrete_values):
+            lower_bound, upper_bound = None, None
+
+        else:
+            lower_bound = min(discrete_values)
+            upper_bound = max(discrete_values)
+
+        super().__init__(lower_bound=lower_bound, upper_bound=upper_bound)  # type: ignore
 
         self.discrete_values = discrete_values
         self.sample_dtype = sample_dtype
 
-    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray:
+    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray | list[Any]:
         """
         :param number_of_samples: Number of samples to make.
         :param kwargs: Miscellaneous keyword arguments.
@@ -234,7 +248,7 @@ class RoundRobinDiscreteValueSampler(Sampler):
 
         self.sampled_elements: set[float] = set()
 
-    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray:
+    def sample(self, number_of_samples: int = 1, **kwargs) -> np.ndarray | list[Any]:
         """
         :param number_of_samples: Number of samples to make.
         :param kwargs: Miscellaneous keyword arguments.
@@ -329,7 +343,7 @@ def sample_from_discrete_range(
     return sample
 
 
-def sample_from_discrete_values(discrete_values: list[float | int], number_of_samples: int = 1) -> np.ndarray:
+def sample_from_discrete_values(discrete_values: list[float | int | str], number_of_samples: int = 1) -> np.ndarray:
     """
     :param discrete_values: List of discrete values to sample from.
     :param number_of_samples: Number of samples to make.

{libinephany-0.14.0.dist-info → libinephany-0.15.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: libinephany
-Version: 0.14.0
+Version: 0.15.0
 Summary: Inephany library containing code commonly used by multiple subpackages.
 Author-email: Inephany <info@inephany.com>
 License: Apache 2.0

{libinephany-0.14.0.dist-info → libinephany-0.15.0.dist-info}/RECORD

@@ -16,32 +16,32 @@ libinephany/observations/post_processors/__init__.py,sha256=47DEQpj8HBSa-_TImW-5
 libinephany/observations/post_processors/postprocessors.py,sha256=43_e5UaDPr2KbAvqc_w3wLqnlm7bgRjqgCtyQ95-8cM,5913
 libinephany/pydantic_models/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 libinephany/pydantic_models/configs/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-libinephany/pydantic_models/configs/hyperparameter_configs.py,sha256=5BH7aPiFtSfhTKp-Z0WTbCYuGJvUTskgFgMttxpDZb0,13654
+libinephany/pydantic_models/configs/hyperparameter_configs.py,sha256=FYl8A2_9L-ohg36aZEW5kREO3tcqIyztYpW62s99tqY,14562
 libinephany/pydantic_models/configs/observer_config.py,sha256=v_ChzaVXC_rlZ7eDZPuCae1DdG7-PS3mPwC-OaWpGQo,1355
 libinephany/pydantic_models/configs/outer_model_config.py,sha256=GQ0QBSC2Xht8x8X_TEMfYM2GF_x1kErLuFrA_H6Jhs0,1209
 libinephany/pydantic_models/schemas/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 libinephany/pydantic_models/schemas/agent_info.py,sha256=me5gDxvZjP9TNK588mpUvxiiJrPDqy3Z7ZHRzryAYTs,2628
-libinephany/pydantic_models/schemas/inner_task_profile.py,sha256=DBLdXeG_igZKp9t1A9bDiWPtaQ2r0rxKTGnpDpBjGlY,10344
+libinephany/pydantic_models/schemas/inner_task_profile.py,sha256=Xu0tQmhGwV043tTamFiHekuE1RRXhhrUrGbtymjXo7g,11722
 libinephany/pydantic_models/schemas/observation_models.py,sha256=YjQmrWZ0r-_LRp92jvhSD8p1grKsMVXCXoou4q15Ue8,1849
 libinephany/pydantic_models/schemas/request_schemas.py,sha256=VED8eAUvBofxeAx9gWU8DyCZOTVD3QsHRq-TO7kyOqk,1260
 libinephany/pydantic_models/schemas/response_schemas.py,sha256=SKFuasdjX5aH_I0vT3SwnpwhyMf9cNPB1ZpDeAGgoO8,2158
 libinephany/pydantic_models/schemas/tensor_statistics.py,sha256=Z-x-Fi_Dm0pLoHI88DnJO1krY671o0zbGRzx-gXPtVY,7534
 libinephany/pydantic_models/states/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-libinephany/pydantic_models/states/hyperparameter_states.py,sha256=JWNSmQmncnQu2GTh0MeT8pvPLxttyrhsKXErucdURqQ,32223
+libinephany/pydantic_models/states/hyperparameter_states.py,sha256=fwqUmRbT5WxcnMPK8DmRXkBQOtCs9n6V24BeCyFTFL8,32688
 libinephany/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 libinephany/utils/agent_utils.py,sha256=_2w1AY5Y4mQ5hes_Rq014VhZXOtIOn-W92mZgeixv3g,2658
 libinephany/utils/asyncio_worker.py,sha256=Ew23zKIbG1zwyCudcyiObMrw4G0f3p2QXzZfM4mePqI,2751
 libinephany/utils/backend_statuses.py,sha256=ZbpBPbz0qKmeqxyGGN_ePTrQ7Wrxh7KM6W26UDbPXtQ,644
-libinephany/utils/constants.py,sha256=RH-0fZe6SL1WrrsrW4KP4k7ClQZDq8IFEdok3hEnRt4,1952
+libinephany/utils/constants.py,sha256=piawYQa51vCxxAHCH3YoWOgUhTlgqgQxKMCenkoQTsc,2170
 libinephany/utils/directory_utils.py,sha256=408unVeE_5_Hm-ZYZuxc9sdvfuU0CgYELX7EzPlPieo,1217
 libinephany/utils/dropout_utils.py,sha256=X43yCW7Dh1cC5sNnivgS5j1fn871K_RCvxCBTT0YHKg,3392
-libinephany/utils/enums.py,sha256=YH10mUhW4kjYS0cp4XUASok9vfPl0jv9ZhS3HpZD0Zg,2339
+libinephany/utils/enums.py,sha256=kEECkJO2quKAyVAqzgOzOP-d4qIENE3z_RyymSvyIB8,2420
 libinephany/utils/error_severities.py,sha256=B9oidqOVaYOe0W6P6GwjpmuDsrkyTX30v1xdiUStCFk,1427
 libinephany/utils/exceptions.py,sha256=kgwLpHOgy3kciUz_I18xnYsWRtzdonfadUtwG2uDYk8,1823
 libinephany/utils/import_utils.py,sha256=WzC6V6UIa0nCiU2MekROwG82fWBh9RuVzichtby5EvM,1495
 libinephany/utils/optim_utils.py,sha256=-PLqsyuq4ZH3spBy_olNB3yuLwvhnLrCF0384elCmXc,8777
 libinephany/utils/random_seeds.py,sha256=eF-ErrMShu8mp9V_gXrB_iUxR-Lb-OtHypEEUQAGn2Y,1565
-libinephany/utils/samplers.py,sha256=RxNib-BpFSsx0qRoWlkfRvUGJNhM-k2uhwXCYNYnThs,11546
+libinephany/utils/samplers.py,sha256=uyVGAy5cm5bCyWMOuySJmzUc_vFuieO_3zydJciwdv4,12158
 libinephany/utils/standardizers.py,sha256=pG1K_XL4OR_NjVtT6Hjbln1dk1BtQdDuSK1PQTkA17Y,8014
 libinephany/utils/torch_distributed_utils.py,sha256=ygdVz-s7hMRoBJcZkNRBlF81MYnxoRJt8S0SAwq6SC4,2467
 libinephany/utils/torch_utils.py,sha256=o5TsqrXe6Id04P6SqB_avGBRZutbu6IBB61llAHQ_PY,2696
@@ -50,8 +50,8 @@ libinephany/utils/typing.py,sha256=rGbaPO3MaUndsWiC_wHzReD_TOLYqb43i01pKN-j7Xs,6
 libinephany/web_apps/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 libinephany/web_apps/error_logger.py,sha256=gAQIaqerqP4ornXZwFF1cghjnd2mMZEt3aVrTuUCr34,16653
 libinephany/web_apps/web_app_utils.py,sha256=qiq_lasPipgN1RgRudPJc342kYci8O_4RqppxmIX8NY,4095
-libinephany-0.14.0.dist-info/licenses/LICENSE,sha256=pogfDoMBP07ehIOvWymuWIar8pg2YLUhqOHsJQU3wdc,9250
-libinephany-0.14.0.dist-info/METADATA,sha256=32VAU-sCvWbhgj89cKZLRIWP_P9aa3QtH3jK9DuPnaE,8354
-libinephany-0.14.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-libinephany-0.14.0.dist-info/top_level.txt,sha256=bYAOXQdJgIoLkO2Ui0kxe7pSYegS_e38u0dMscd7COQ,12
-libinephany-0.14.0.dist-info/RECORD,,
+libinephany-0.15.0.dist-info/licenses/LICENSE,sha256=pogfDoMBP07ehIOvWymuWIar8pg2YLUhqOHsJQU3wdc,9250
+libinephany-0.15.0.dist-info/METADATA,sha256=lU7SqV1ArMEAyuZ845Z1jAYxNUEYGfJ8Tl6Df6EwSpc,8354
+libinephany-0.15.0.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+libinephany-0.15.0.dist-info/top_level.txt,sha256=bYAOXQdJgIoLkO2Ui0kxe7pSYegS_e38u0dMscd7COQ,12
+libinephany-0.15.0.dist-info/RECORD,,