x-evolution 0.0.1__py3-none-any.whl

x_evolution/__init__.py

@@ -0,0 +1,4 @@
+
+from x_evolution.x_evolution import (
+    EvoStrategy
+)

x_evolution/x_evolution.py

@@ -0,0 +1,188 @@
+from __future__ import annotations
+from typing import Callable
+
+import torch
+from torch import tensor, is_tensor
+from torch.nn import Module
+import torch.nn.functional as F
+from torch.func import functional_call, vmap
+
+from beartype import beartype
+from beartype.door import is_bearable
+
+from accelerate import Accelerator
+
+from x_mlps_pytorch.noisable import (
+    Noisable,
+    with_seed
+)
+
+# constants
+
+MAX_SEED_VALUE = int(2 ** 32)
+
+# helper functions
+
+def exists(v):
+    return v is not None
+
+def default(v, d):
+    return v if exists(v) else d
+
+def normalize(t, eps = 1e-6):
+    return F.layer_norm(t, t.shape[-1:], eps = eps)
+
+# class
+
+class EvoStrategy(Module):
+
+    @beartype
+    def __init__(
+        self,
+        model: Module,
+        *,
+        environment: Callable[[Module], float],  # the environment is simply a function that takes in the model and returns a fitness score
+        num_generations,
+        population_size = 30,
+        learning_rate = 1e-3, # todo - optimizer
+        noise_scale = 1e-3,   # the noise scaling during rollouts with environment, todo - figure out right value and make sure it can also be customized per parameter name through a dict
+        param_names_to_optimize: list[str] | None = None,
+        fitness_to_weighted_factor: Callable[[Tensor], Tensor] = normalize,
+        cpu = False,
+        accelerate_kwargs: dict = dict(),
+    ):
+        super().__init__()
+
+        self.accelerate = Accelerator(cpu = cpu, **accelerate_kwargs)
+
+        self.model = model
+        self.noisable_model = Noisable(model)
+
+        self.environment = environment
+
+        param_names = set(dict(model.named_parameters()).keys())
+
+        # default to all parameters to optimize with evo strategy
+
+        param_names_to_optimize = default(param_names_to_optimize, param_names)
+
+        # validate
+
+        assert all([name in param_names for name in param_names_to_optimize])
+        assert len(param_names_to_optimize) > 0, 'nothing to optimize'
+
+        # sort param names and store
+
+        param_names_list = list(param_names_to_optimize)
+        param_names_list.sort()
+
+        self.param_names_to_optimize = param_names_list
+
+        # hyperparameters
+
+        self.population_size = population_size
+        self.num_params = len(param_names_list) # just convenience for generating all the seeds for all the randn for the proposed memory efficient way
+
+        self.num_generations = num_generations
+
+        # the function that transforms a tensor of fitness floats to the weight for the weighted average of the noise for rolling out 1x1 ES
+
+        self.fitness_to_weighted_factor = fitness_to_weighted_factor
+
+        self.noise_scale = noise_scale
+        self.learning_rate = learning_rate
+
+        self.register_buffer('_dummy', tensor(0), persistent = False)
+
+    @property
+    def device(self):
+        return self._dummy.device
+
+    def print(self, *args, **kwargs):
+        return self.accelerate.print(*args, **kwargs)
+
+    @torch.inference_mode()
+    def evolve_(
+        self,
+        fitnesses: list[float] | Tensor,
+        seeds_for_population: list[int] | Tensor
+    ):
+        model = self.noisable_model
+
+        if isinstance(fitnesses, list):
+            fitnesses = tensor(fitnesses)
+
+        if isinstance(seeds_for_population, list):
+            seeds_for_population = tensor(seeds_for_population)
+
+        fitnesses = fitnesses.to(self.device)
+        seeds_for_population.to(self.device)
+
+        # they use a simple z-score for the fitnesses, need to figure out the natural ES connection
+
+        noise_weights = self.fitness_to_weighted_factor(fitnesses)
+
+        noise_weights *= self.learning_rate # some learning rate that subsumes another constant
+
+        # update one seed at a time for enabling evolutionary strategy for large models
+
+        for individual_seed, noise_weight in zip(seeds_for_population.tolist(), noise_weights.tolist()):
+
+            individual_param_seeds = with_seed(individual_seed)(torch.randint)(0, MAX_SEED_VALUE, (self.num_params,))
+
+            noise_config = dict(zip(self.param_names_to_optimize, individual_param_seeds.tolist()))
+
+            # set the noise weight
+
+            noise_config = {param_name: (seed, noise_weight) for param_name, seed in noise_config.items()}
+
+            # now update
+
+            model.add_noise_(noise_config)
+
+    @torch.inference_mode()
+    def forward(
+        self
+    ):
+
+        model = self.noisable_model
+
+        for index in range(self.num_generations):
+            generation = index + 1
+
+            fitnesses = []
+
+            # predetermine the seeds for each population
+            # each seed is then used as a seed for all the parameters
+
+            seeds_for_population = torch.randint(0, MAX_SEED_VALUE, (self.population_size,))
+
+            # now loop through the entire population of noise
+
+            for individual_seed in seeds_for_population.tolist():
+
+                individual_param_seeds = with_seed(individual_seed)(torch.randint)(0, MAX_SEED_VALUE, (self.num_params,))
+
+                noise_config = dict(zip(self.param_names_to_optimize, individual_param_seeds.tolist()))
+                noise_config = {param_name: (seed, self.noise_scale) for param_name, seed in noise_config.items()}
+
+                with model.temp_add_noise_(noise_config):
+                    fitness = self.environment(model)
+
+                    if is_tensor(fitness):
+                        assert fitness.numel() == 1
+                        fitness = fitness.item()
+
+                fitnesses.append(fitness)
+
+            # normalize the fitness and weighted sum of all the noise is the update
+
+            fitnesses = tensor(fitnesses).float()
+
+            self.evolve_(fitnesses, seeds_for_population)
+
+            # log
+
+            self.print(f'[{generation}] average fitness: {fitnesses.mean():.3f} | fitness std: {fitnesses.std():.3f}')
+
+        self.print('evolution complete')

x_evolution-0.0.1.dist-info/METADATA

@@ -0,0 +1,104 @@
+Metadata-Version: 2.4
+Name: x-evolution
+Version: 0.0.1
+Summary: x-evolution
+Project-URL: Homepage, https://pypi.org/project/x-evolution/
+Project-URL: Repository, https://github.com/lucidrains/x-evolution
+Author-email: Phil Wang <lucidrains@gmail.com>
+License: MIT License
+        
+        Copyright (c) 2025 Phil Wang
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+License-File: LICENSE
+Keywords: artificial intelligence,evolution,evolutionary algorithms
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.9
+Requires-Dist: accelerate
+Requires-Dist: beartype
+Requires-Dist: einops>=0.8.0
+Requires-Dist: torch>=2.4
+Requires-Dist: x-mlps-pytorch>=0.1.8
+Requires-Dist: x-transformers>=2.11.23
+Provides-Extra: examples
+Provides-Extra: test
+Requires-Dist: pytest; extra == 'test'
+Description-Content-Type: text/markdown
+
+## x-evolution (wip)
+
+Implementation of various evolutionary algorithms, starting with evolutionary strategies
+
+## Install
+
+```bash
+$ pip install x-evolution
+```
+
+## Usage
+
+```python
+import torch
+from x_evolution import EvoStrategy
+
+# model
+
+from torch import nn
+model = torch.nn.Sequential(
+    nn.Linear(8, 16),
+    nn.ReLU(),
+    nn.Linear(16, 4)
+)
+
+# evolution wrapper
+
+evo_strat = EvoStrategy(
+    model,
+    environment = lambda model: torch.randint(0, 100, ()), # environment is just a function that takes in the individual model (with unique noise) and outputs the fitness - you can select for whatever you want here, does not have to be differentiable.
+    population_size = 30,
+    num_generations = 100,
+    learning_rate = 1e-3,
+    noise_scale = 1e-3
+)
+
+# do evolution with your desired fitness function for so many generations
+
+evo_strat()
+
+# then save your evolved model, maybe for alternating with gradient based training
+
+torch.save(model.state_dict(), './evolved.pt')
+```
+
+## Citations
+
+```bibtex
+@article{Qiu2025EvolutionSA,
+    title   = {Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning},
+    author  = {Xin Qiu and Yulu Gan and Conor F. Hayes and Qiyao Liang and Elliot Meyerson and Babak Hodjat and Risto Miikkulainen},
+    journal = {ArXiv},
+    year    = {2025},
+    volume  = {abs/2509.24372},
+    url     = {https://api.semanticscholar.org/CorpusID:281674745}
+}
+```

x_evolution-0.0.1.dist-info/RECORD

@@ -0,0 +1,6 @@
+x_evolution/__init__.py,sha256=XcwXJgIMPnCWGfGws3-vKgoR_7IfVslJBtiMvmEeSg0,57
+x_evolution/x_evolution.py,sha256=Ww7RLTRDG2x_p1WuLVHWZle50GnleibQVihd3DEGow4,5941
+x_evolution-0.0.1.dist-info/METADATA,sha256=3uYkZheDcENk76UMz94nK_KScz3xjfekEcJs-J8oW_U,3586
+x_evolution-0.0.1.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+x_evolution-0.0.1.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+x_evolution-0.0.1.dist-info/RECORD,,

x_evolution-0.0.1.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: hatchling 1.27.0
+Root-Is-Purelib: true
+Tag: py3-none-any

x_evolution-0.0.1.dist-info/licenses/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 Phil Wang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.