evolutionary-policy-optimization 0.0.53__tar.gz → 0.0.54__tar.gz

{evolutionary_policy_optimization-0.0.53 → evolutionary_policy_optimization-0.0.54}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evolutionary-policy-optimization
-Version: 0.0.53
+Version: 0.0.54
 Summary: EPO - Pytorch
 Project-URL: Homepage, https://pypi.org/project/evolutionary-policy-optimization/
 Project-URL: Repository, https://github.com/lucidrains/evolutionary-policy-optimization

evolutionary_policy_optimization-0.0.54/evolutionary_policy_optimization/distributed.py

@@ -0,0 +1,82 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+from torch.autograd import Function
+
+import torch.distributed as dist
+
+import einx
+from einops import rearrange
+
+def exists(val):
+    return val is not None
+
+def default(val, d):
+    return val if exists(val) else d
+
+def divisible_by(num, den):
+    return (num % den) == 0
+
+def pad_dim_to(t, length, dim = 0):
+    pad_length = length - t.shape[dim]
+    zero_pairs = (-dim - 1) if dim < 0 else (t.ndim - dim - 1)
+    return F.pad(t, (*((0, 0) * zero_pairs), 0, pad_length))
+
+def is_distributed():
+    return dist.is_initialized() and dist.get_world_size() > 1
+
+def maybe_sync_seed(device, max_size = int(1e6)):
+    rand_int = torch.randint(0, max_size, (), device = device)
+
+    if is_distributed():
+        dist.all_reduce(rand_int)
+
+    return rand_int.item()
+
+def maybe_barrier():
+    if not is_distributed():
+        return
+
+    dist.barrier()
+
+def all_gather_same_dim(t):
+    t = t.contiguous()
+    world_size = dist.get_world_size()
+    gathered_tensors = [torch.empty_like(t, device = t.device, dtype = t.dtype) for i in range(world_size)]
+    dist.all_gather(gathered_tensors, t)
+    return gathered_tensors
+
+def gather_sizes(t, *, dim):
+    size = torch.tensor(t.shape[dim], device = t.device, dtype = torch.long)
+    sizes = all_gather_same_dim(size)
+    return torch.stack(sizes)
+
+def has_only_one_value(t):
+    return (t == t[0]).all()
+
+def all_gather_variable_dim(t, dim = 0, sizes = None):
+    device, rank, world_size = t.device, dist.get_rank(), dist.get_world_size()
+
+    if not exists(sizes):
+        sizes = gather_sizes(t, dim = dim)
+
+    if has_only_one_value(sizes):
+        gathered_tensors = all_gather_same_dim(t)
+        gathered_tensors = torch.cat(gathered_tensors, dim = dim)
+        return gathered_tensors, sizes
+
+    max_size = sizes.amax().item()
+
+    padded_t = pad_dim_to(t, max_size, dim = dim)
+    gathered_tensors = all_gather_same_dim(padded_t)
+
+    gathered_tensors = torch.cat(gathered_tensors, dim = dim)
+    seq = torch.arange(max_size, device = device)
+
+    mask = einx.less('j i -> (i j)', seq, sizes)
+    seq = torch.arange(mask.shape[-1], device = device)
+    indices = seq[mask]
+
+    gathered_tensors = gathered_tensors.index_select(dim, indices)
+
+    return gathered_tensors, sizes

{evolutionary_policy_optimization-0.0.53 → evolutionary_policy_optimization-0.0.54}/evolutionary_policy_optimization/epo.py

@@ -8,14 +8,22 @@ from random import randrange
 import torch
 from torch import nn, cat, stack, is_tensor, tensor, Tensor
 import torch.nn.functional as F
+import torch.distributed as dist
 from torch.nn import Linear, Module, ModuleList
 from torch.utils.data import TensorDataset, DataLoader
-from torch.utils._pytree import tree_map, tree_flatten, tree_unflatten
+from torch.utils._pytree import tree_map
 
 import einx
 from einops import rearrange, repeat, einsum, pack
 from einops.layers.torch import Rearrange
 
+from evolutionary_policy_optimization.distributed import (
+    is_distributed,
+    maybe_sync_seed,
+    all_gather_variable_dim,
+    maybe_barrier
+)
+
 from assoc_scan import AssocScan
 
 from adam_atan2_pytorch import AdoptAtan2
@@ -822,6 +830,15 @@ class Agent(Module):
 
         fitness_scores = self.get_fitness_scores(cumulative_rewards, memories)
 
+        # stack memories
+
+        memories = map(stack, zip(*memories))
+
+        maybe_barrier()
+
+        if is_distributed():
+            memories = map(partial(all_gather_variable_dim, dim = 0), memories)
+
         (
             episode_ids,
             states,
@@ -831,7 +848,7 @@ class Agent(Module):
             rewards,
             values,
             dones
-        ) = map(stack, zip(*memories))
+        ) = memories
 
         advantages = self.calc_gae(
             rewards[:-1],
@@ -1035,6 +1052,32 @@ class EPO(Module):
         self.episodes_per_latent = episodes_per_latent
         self.max_episode_length = max_episode_length
 
+        self.register_buffer('dummy', tensor(0))
+
+    @property
+    def device(self):
+        return self.dummy.device
+
+    def latents_for_machine(self):
+        num_latents = self.num_latents
+
+        if not is_distributed():
+            return list(range(self.num_latents))
+
+        world_size, rank = dist.get_world_size(), dist.get_rank()
+        assert num_latents >= world_size, 'number of latents must be greater than world size for now'
+        assert rank < world_size
+
+        pad_id = -1
+        num_latents_rounded_up = ceil(num_latents / world_size) * world_size
+        latent_ids = torch.arange(num_latents_rounded_up)
+        latent_ids[latent_ids >= num_latents] = pad_id
+
+        latent_ids = rearrange(latent_ids, '(world latents) -> world latents', world = world_size)
+        out = latent_ids[rank]
+
+        return out[out != pad_id].tolist()
+
     @torch.no_grad()
     def forward(
         self,
@@ -1050,19 +1093,23 @@ class EPO(Module):
 
         cumulative_rewards = torch.zeros((self.num_latents))
 
+        latent_ids = self.latents_for_machine()
+
         for episode_id in tqdm(range(self.episodes_per_latent), desc = 'episode'):
 
+            maybe_barrier()
+
             # maybe fix seed for environment across all latents
 
             env_reset_kwargs = dict()
 
             if fix_seed_across_latents:
-                seed = randrange(int(1e6))
+                seed = maybe_sync_seed(device = self.device)
                 env_reset_kwargs = dict(seed = seed)
 
             # for each latent (on a single machine for now)
 
-            for latent_id in tqdm(range(self.num_latents), desc = 'latent'):
+            for latent_id in tqdm(latent_ids, desc = 'latent'):
                 time = 0
 
                 # initial state

{evolutionary_policy_optimization-0.0.53 → evolutionary_policy_optimization-0.0.54}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "evolutionary-policy-optimization"
-version = "0.0.53"
+version = "0.0.54"
 description = "EPO - Pytorch"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }