Merge pull request #212 from starquakee/evoxtorch-dev-fcc

Brax now supports Single-layer vmap (Hpo Problem)

src/evox/operators/crossover/differential_evolution.py

@@ -25,11 +25,11 @@ def DE_differential_sum(
 
     select_len = num_diff_vectors.unsqueeze(1) * 2 + 1
     rand_indices = torch.randint(0, pop_size, (pop_size, diff_padding_num), device=device)
-    rand_indices = torch.where(rand_indices == index.unsqueeze(1), torch.tensor(pop_size - 1, device=device), rand_indices)
+    rand_indices = torch.where(rand_indices == index.unsqueeze(1), pop_size - 1, rand_indices)
 
     pop_permute = population[rand_indices]
     mask = torch.arange(diff_padding_num, device=device).unsqueeze(0) < select_len
-    pop_permute_padding = torch.where(mask.unsqueeze(2), pop_permute, torch.zeros_like(pop_permute))
+    pop_permute_padding = torch.where(mask.unsqueeze(2), pop_permute, 0)
 
     diff_vectors = pop_permute_padding[:, 1:]
     difference_sum = diff_vectors[:, 0::2].sum(dim=1) - diff_vectors[:, 1::2].sum(dim=1)

src/evox/problems/neuroevolution/brax.py

@@ -11,7 +11,7 @@
 from brax import envs
 from brax.io import html, image
 
-from ...core import Problem, jit_class
+from ...core import Problem, _vmap_fix, jit_class, vmap_impl
 from .utils import get_vmap_model_state_forward
 
 
@@ -70,7 +70,8 @@ def __init__(
         The initial key is obtained from `torch.random.get_rng_state()`.
 
         ## Warning
-        This problem does NOT support HPO wrapper (`problems.hpo_wrapper.HPOProblemWrapper`), i.e., the workflow containing this problem CANNOT be vmapped.
+        This problem does NOT support HPO wrapper (`problems.hpo_wrapper.HPOProblemWrapper`) out-of-box, i.e., the workflow containing this problem CANNOT be vmapped.
+        *However*, by setting `pop_size` to the multiplication of inner population size and outer population size, you can still use this problem in a HPO workflow.
 
         ## Examples
         >>> from evox import problems
@@ -137,13 +138,7 @@ def __init__(
         self.rotate_key = rotate_key
         self.reduce_fn = reduce_fn
 
-    def evaluate(self, pop_params: Dict[str, nn.Parameter]) -> torch.Tensor:
-        """Evaluate the final rewards of a population (batch) of model parameters.
-
-        :param pop_params: A dictionary of parameters where each key is a parameter name and each value is a tensor of shape (batch_size, *param_shape) representing the batched parameters of batched models.
-
-        :return: A tensor of shape (batch_size,) containing the reward of each sample in the population.
-        """
+    def _normal_evaluate(self, pop_params: Dict[str, nn.Parameter]) -> torch.Tensor:
         # Unpack parameters and buffers
         state_params = {self._param_to_state_key_map[key]: value for key, value in pop_params.items()}
         model_state = dict(self._vmap_model_buffers)
@@ -157,6 +152,25 @@ def evaluate(self, pop_params: Dict[str, nn.Parameter]) -> torch.Tensor:
         # Return
         return rewards
 
+    def evaluate(self, pop_params: Dict[str, nn.Parameter]) -> torch.Tensor:
+        """Evaluate the final rewards of a population (batch) of model parameters.
+
+        :param pop_params: A dictionary of parameters where each key is a parameter name and each value is a tensor of shape (batch_size, *param_shape) representing the batched parameters of batched models.
+
+        :return: A tensor of shape (batch_size,) containing the reward of each sample in the population.
+        """
+        return self._normal_evaluate(pop_params)
+
+    @vmap_impl(evaluate)
+    def _vmap_evaluate(self, pop_params: Dict[str, nn.Parameter]) -> torch.Tensor:
+        _, vmap_dim, vmap_size = _vmap_fix.unwrap_batch_tensor(list(pop_params.values())[0])
+        assert vmap_dim == (0,)
+        vmap_size = vmap_size[0]
+        pop_params = {k: _vmap_fix.unwrap_batch_tensor(v)[0].view(vmap_size * v.size(0), *v.size()[1:]) for k, v in pop_params.items()}
+        flat_rewards = self._normal_evaluate(pop_params)
+        rewards = flat_rewards.view(vmap_size, flat_rewards.size(0) // vmap_size, *flat_rewards.size()[1:])
+        return _vmap_fix.wrap_batch_tensor(rewards, vmap_dim)
+
     def _model_forward(
         self, model_state: Dict[str, torch.Tensor], obs: torch.Tensor, record_trajectory: bool = False
     ) -> Tuple[Dict[str, torch.Tensor], torch.Tensor]: