locoformer 0.0.17__tar.gz → 0.0.30__tar.gz

{locoformer-0.0.17 → locoformer-0.0.30}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: locoformer
-Version: 0.0.17
+Version: 0.0.30
 Summary: LocoFormer
 Project-URL: Homepage, https://pypi.org/project/locoformer/
 Project-URL: Repository, https://github.com/lucidrains/locoformer
@@ -38,6 +38,7 @@ Requires-Dist: assoc-scan
 Requires-Dist: beartype
 Requires-Dist: einops>=0.8.0
 Requires-Dist: einx>=0.3.0
+Requires-Dist: hl-gauss-pytorch>=0.2.0
 Requires-Dist: rotary-embedding-torch
 Requires-Dist: torch>=2.4
 Requires-Dist: x-mlps-pytorch
@@ -54,7 +55,7 @@ Description-Content-Type: text/markdown
 
 [LocoFormer - Generalist Locomotion via Long-Context Adaptation](https://generalist-locomotion.github.io/)
 
-The gist is they trained a simple Transformer-XL in simulation on robots with many different bodies (cross-embodiment) with extreme domain randomization. When transferring to the real-world, they noticed the robot now gains the ability to adapt to insults. The XL memories span across multiple trials, which allowed the robot to learn in-context adaptation.
+The gist is they trained a simple Transformer-XL in simulation on robots with many different bodies (cross-embodiment) and extreme domain randomization. When transferring to the real-world, they noticed the robot now gains the ability to adapt to insults. The XL memories span across multiple trials, which allowed the robot to learn in-context adaptation.
 
 ## Sponsors
 

{locoformer-0.0.17 → locoformer-0.0.30}/README.md

@@ -4,7 +4,7 @@
 
 [LocoFormer - Generalist Locomotion via Long-Context Adaptation](https://generalist-locomotion.github.io/)
 
-The gist is they trained a simple Transformer-XL in simulation on robots with many different bodies (cross-embodiment) with extreme domain randomization. When transferring to the real-world, they noticed the robot now gains the ability to adapt to insults. The XL memories span across multiple trials, which allowed the robot to learn in-context adaptation.
+The gist is they trained a simple Transformer-XL in simulation on robots with many different bodies (cross-embodiment) and extreme domain randomization. When transferring to the real-world, they noticed the robot now gains the ability to adapt to insults. The XL memories span across multiple trials, which allowed the robot to learn in-context adaptation.
 
 ## Sponsors
 

{locoformer-0.0.17 → locoformer-0.0.30}/locoformer/locoformer.py

@@ -1,4 +1,5 @@
 from __future__ import annotations
+from typing import Callable
 from functools import partial
 
 from pathlib import Path
@@ -16,7 +17,7 @@ import torch
 from torch import nn, cat, stack, arange, Tensor, tensor, is_tensor, from_numpy
 import torch.nn.functional as F
 from torch.nn import Module, ModuleList, Linear, RMSNorm, Identity, Sequential
-from torch.utils._pytree import tree_map
+from torch.utils._pytree import tree_map, tree_flatten, tree_unflatten
 from torch.utils.data import Dataset, DataLoader
 from torch.optim import Optimizer
 
@@ -26,12 +27,16 @@ from einops.layers.torch import Rearrange
 
 from rotary_embedding_torch import RotaryEmbedding
 
+from hl_gauss_pytorch import HLGaussLoss
+
 from assoc_scan import AssocScan
 
 # constants
 
 LinearNoBias = partial(Linear, bias = False)
 
+Cache = namedtuple('Cache', ('curr_timestep', 'kv_cache')) # (int, Tensor)
+
 # helper functions
 
 def exists(v):
@@ -43,6 +48,9 @@ def default(v, d):
 def first(arr):
     return arr[0]
 
+def xnor(x, y):
+    return not (x ^ y)
+
 def divisible_by(num, den):
     return (num % den) == 0
 
@@ -51,6 +59,9 @@ def divisible_by(num, den):
 def log(t, eps = 1e-20):
     return t.clamp_min(eps).log()
 
+def is_empty(t):
+    return t.numel() == 0
+
 def tree_map_tensor(x, fn):
     return tree_map(lambda t: t if not is_tensor(t) else fn(t), x)
 
@@ -67,6 +78,9 @@ def pad_at_dim(
     zeros = ((0, 0) * dims_from_right)
     return F.pad(t, (*zeros, *pad), value = value)
 
+def normalize(t, eps = 1e-5):
+    return (t - t.mean()) / t.std().clamp_min(eps)
+
 def calc_entropy(logits):
     prob = logits.softmax(dim = -1)
     return -(prob * log(prob)).sum(dim = -1)
@@ -250,6 +264,57 @@ class ReplayDataset(Dataset):
 
         return data
 
+class RemappedReplayDataset(Dataset):
+    def __init__(
+        self,
+        dataset: ReplayDataset,
+        episode_mapping: Tensor | list[list[int]],
+        shuffle_episodes = False
+    ):
+        assert len(dataset) > 0
+        self.dataset = dataset
+
+        if is_tensor(episode_mapping):
+            assert episode_mapping.dtype in (torch.int, torch.long) and episode_mapping.ndim == 2
+            episode_mapping = episode_mapping.tolist()
+
+        self.episode_mapping = episode_mapping
+        self.shuffle_episodes = shuffle_episodes
+
+    def __len__(self):
+        return len(self.episode_mapping)
+
+    def __getitem__(self, idx):
+
+        episode_indices = self.episode_mapping[idx]
+
+        episode_indices = tensor(episode_indices)
+        episode_indices = episode_indices[(episode_indices >= 0) & (episode_indices < len(self.dataset))]
+
+        assert not is_empty(episode_indices)
+
+        if self.shuffle_episodes and episode_indices.numel() > 1:
+            num_episodes = len(episode_indices)
+            episode_indices = episode_indices[torch.randperm(num_episodes)]
+
+        episode_data = [self.dataset[i] for i in episode_indices.tolist()]
+
+        episode_lens = stack([data.pop('_lens') for data in episode_data])
+
+        keys = first(episode_data).keys()
+
+        values = [list(data.values()) for data in episode_data]
+
+        values = [cat(field_values) for field_values in zip(*values)] # concat across time
+
+        multi_episode_data = dict(zip(keys, values))
+
+        multi_episode_data['_lens'] = episode_lens.sum()
+
+        multi_episode_data['_episode_indices'] = cat([torch.full((episode_len,), episode_index) for episode_len, episode_index in zip(episode_lens, episode_indices)])
+
+        return multi_episode_data
+
 class ReplayBuffer:
 
     @beartype
@@ -314,6 +379,9 @@ class ReplayBuffer:
 
         self.memory_namedtuple = namedtuple('Memory', list(fields.keys()))
 
+    def __len__(self):
+        return (self.episode_lens > 0).sum().item()
+
     def reset_(self):
         self.episode_lens[:] = 0
         self.episode_index = 0
@@ -375,15 +443,91 @@ class ReplayBuffer:
 
         return self.memory_namedtuple(**data)
 
-    def dataset(self) -> Dataset:
+    def dataset(
+        self,
+        episode_mapping: Tensor | list[list[int]] | None = None,
+    ) -> Dataset:
         self.flush()
 
-        return ReplayDataset(self.folder)
+        dataset = ReplayDataset(self.folder)
+
+        if not exists(episode_mapping):
+            return dataset
 
-    def dataloader(self, batch_size, **kwargs) -> DataLoader:
+        return RemappedReplayDataset(dataset, episode_mapping)
+
+    def dataloader(
+        self,
+        batch_size,
+        episode_mapping: Tensor | list[list[int]] | None = None,
+        **kwargs
+    ) -> DataLoader:
         self.flush()
 
-        return DataLoader(self.dataset(), batch_size = batch_size, collate_fn = collate_var_time, **kwargs)
+        return DataLoader(self.dataset(episode_mapping), batch_size = batch_size, collate_fn = collate_var_time, **kwargs)
+
+# normalization + conditioning (needed for the commands to the robot)
+
+class MaybeAdaRMSNormWrapper(Module):
+    def __init__(
+        self,
+        fn: Module,
+        dim,
+        dim_cond = None
+    ):
+        super().__init__()
+        condition = exists(dim_cond)
+
+        self.fn = fn
+        self.norm = nn.RMSNorm(dim, elementwise_affine = not condition)
+
+        self.accept_condition = condition
+
+        if condition:
+            self.to_gamma = LinearNoBias(dim_cond, dim)
+            self.to_ada_norm_zero = nn.Linear(dim_cond, dim)
+
+            nn.init.zeros_(self.to_gamma.weight, 0.)
+            nn.init.zeros_(self.to_ada_norm_zero.weight, 0.)
+            nn.init.constant_(self.to_ada_norm_zero.bias, -5.)
+
+    def forward(
+        self,
+        x,
+        cond = None,
+        **kwargs
+    ):
+
+        need_cond = self.accept_condition
+        assert xnor(exists(cond), need_cond)
+
+        prenormed = self.norm(x)
+
+        if need_cond:
+            if cond.ndim == 2:
+                cond = rearrange(cond, 'b d -> b 1 d')
+
+            scale_in = self.to_gamma(cond)
+            prenormed = prenormed * (scale_in + 1.)
+
+        all_fn_out = self.fn(prenormed, **kwargs)
+
+        if not need_cond:
+            return all_fn_out
+
+        # function may return multiple args
+
+        (out, *rest), tree_spec = tree_flatten(all_fn_out)
+
+        if need_cond:
+            scale_out = self.to_ada_norm_zero(cond).sigmoid()
+            out = out * scale_out
+
+        # restore
+
+        all_fn_out = tree_unflatten((out, *rest), tree_spec)
+
+        return all_fn_out
 
 # transformer-xl with ppo
 
@@ -394,15 +538,12 @@ class Attention(Module):
         window_size,
         dim_head = 64,
         heads = 8,
-        pre_rmsnorm = True,
         fixed_window_size = False,
         accept_value_residual = False
     ):
         super().__init__()
         self.scale = dim_head ** -0.5
 
-        self.norm = RMSNorm(dim) if pre_rmsnorm else Identity()
-
         self.split_heads = Rearrange('b n (h d) -> b h n d', h = heads)
         self.merge_heads = Rearrange('b h n d -> b n (h d)')
 
@@ -446,8 +587,6 @@ class Attention(Module):
 
         device = tokens.device
 
-        tokens = self.norm(tokens)
-
         q, k, v = (self.to_q(tokens), *self.to_kv(tokens).chunk(2, dim = -1))
 
         q, k, v = map(self.split_heads, (q, k, v))
@@ -536,19 +675,24 @@ class TransformerXL(Module):
         dim_head = 64,
         heads = 8,
         expansion_factor = 4.,
+        dim_cond = None,
         final_norm = True,
         fixed_window_size = False,
     ):
         super().__init__()
 
+        condition = exists(dim_cond)
+
+        norm_fn = partial(MaybeAdaRMSNormWrapper, dim = dim, dim_cond = dim_cond)
+
         layers = ModuleList([])
 
         for i in range(depth):
             is_first = i == 0
 
-            attn = Attention(dim = dim, dim_head = dim_head, heads = heads, fixed_window_size = fixed_window_size, window_size = window_size, accept_value_residual = not is_first)
+            attn = norm_fn(Attention(dim = dim, dim_head = dim_head, heads = heads, fixed_window_size = fixed_window_size, window_size = window_size, accept_value_residual = not is_first))
 
-            ff = FeedForward(dim = dim, expansion_factor = expansion_factor)
+            ff = norm_fn(FeedForward(dim = dim, expansion_factor = expansion_factor))
 
             layers.append(ModuleList([
                 attn, ff
@@ -603,14 +747,21 @@ class Locoformer(Module):
         embedder: Module,
         unembedder: Module,
         transformer: dict | TransformerXL,
-        value_network: Module | None = None,
         discount_factor = 0.999,
         gae_lam = 0.95,
         ppo_eps_clip = 0.2,
         ppo_entropy_weight = 0.01,
         ppo_value_clip = 0.4,
+        dim_value_input = None, # needs to be set for value network to be available
+        value_network: Module = nn.Identity(),
+        reward_range: tuple[float, float] | None = None,
+        reward_shaping_fns: list[Callable[[Tensor], float | Tensor]] | None = None,
+        num_reward_bins = 32,
+        hl_gauss_loss_kwargs = dict(),
         value_loss_weight = 0.5,
-        calc_gae_kwargs: dict = dict()
+        calc_gae_kwargs: dict = dict(),
+        recurrent_kv_cache = True,
+        use_spo = False # simple policy optimization https://arxiv.org/abs/2401.16025 - Levine's group (PI) verified it is more stable than PPO
     ):
         super().__init__()
 
@@ -622,11 +773,30 @@ class Locoformer(Module):
         self.embedder = embedder
         self.unembedder = unembedder
 
-        self.value_network = value_network
-
         self.fixed_window_size = transformer.fixed_window_size
         self.window_size = transformer.window_size
 
+        # determine value network, using HL Gauss Layer
+
+        self.to_value_pred = None
+
+        if exists(dim_value_input):
+            assert exists(reward_range)
+
+            self.to_value_pred = nn.Sequential(
+                value_network,
+                LinearNoBias(dim_value_input, num_reward_bins)
+            )
+
+            reward_min, reward_max = reward_range
+
+            self.hl_gauss_loss = HLGaussLoss(
+                min_value = reward_min,
+                max_value = reward_max,
+                num_bins = num_reward_bins,
+                **hl_gauss_loss_kwargs
+            )
+
         # ppo related
 
         self.discount_factor = discount_factor
@@ -638,6 +808,19 @@ class Locoformer(Module):
 
         self.calc_gae_kwargs = calc_gae_kwargs
 
+        # maybe use spo
+
+        self.use_spo = use_spo
+
+        # maybe recurrent kv cache, from Ding et al. https://arxiv.org/abs/2012.15688
+
+        self.recurrent_kv_cache = recurrent_kv_cache
+
+        # reward shaping function
+
+        self.has_reward_shaping = exists(reward_shaping_fns)
+        self.reward_shaping_fns = reward_shaping_fns
+
         # loss related
 
         self.register_buffer('zero', tensor(0.), persistent = False)
@@ -650,10 +833,10 @@ class Locoformer(Module):
         return self.unembedder.parameters()
 
     def critic_parameters(self):
-        if not exists(self.value_network):
+        if not exists(self.to_value_pred):
             return []
 
-        return self.value_network.parameters()
+        return self.to_value_pred.parameters()
 
     def ppo(
         self,
@@ -663,12 +846,20 @@ class Locoformer(Module):
         reward,
         old_value,
         mask,
+        episode_lens,
         actor_optim: Optimizer | None = None,
         critic_optim: Optimizer | None = None
     ):
         window_size = self.window_size
         total_learnable_tokens = mask.sum().item()
 
+        seq_len = state.shape[1]
+        gae_mask = einx.less('j, i -> i j', arange(seq_len, device = self.device), episode_lens)
+
+        advantage, returns = calc_gae(reward, old_value, masks = gae_mask, lam = self.gae_lam, gamma = self.discount_factor, **self.calc_gae_kwargs)
+
+        advantage = normalize(advantage)
+
         windowed_tensors = [
             t.split(window_size, dim = 1) for t in
             (
@@ -677,7 +868,9 @@ class Locoformer(Module):
                 old_action_log_prob,
                 reward,
                 old_value,
-                mask
+                mask,
+                advantage,
+                returns
             )
         ]
 
@@ -694,10 +887,12 @@ class Locoformer(Module):
             old_action_log_prob,
             reward,
             old_value,
-            mask
+            mask,
+            advantage,
+            returns
         ) in zip(*windowed_tensors):
 
-            (action_logits, value), cache = self.forward(state, cache = cache, detach_cache = True, return_values = True)
+            (action_logits, value_logits), cache = self.forward(state, cache = cache, detach_cache = True, return_values = True, return_raw_value_logits = True)
             entropy = calc_entropy(action_logits)
 
             action = rearrange(action, 'b t -> b t 1')
@@ -709,9 +904,10 @@ class Locoformer(Module):
             eps_clip = self.ppo_eps_clip
             ratio = (log_prob - old_action_log_prob).exp()
 
-            advantage, returns = calc_gae(reward, old_value, lam = self.gae_lam, gamma = self.discount_factor, **self.calc_gae_kwargs)
-
-            actor_loss = -torch.min(ratio * advantage, ratio.clamp(1. - eps_clip, 1. + eps_clip) * advantage)
+            if self.use_spo:
+                actor_loss = -(ratio * advantage - (advantage.abs() * (ratio - 1.).square()) / (2 * eps_clip))
+            else:
+                actor_loss = -torch.min(ratio * advantage, ratio.clamp(1. - eps_clip, 1. + eps_clip) * advantage)
 
             actor_loss = actor_loss - self.ppo_entropy_weight * entropy
 
@@ -720,11 +916,13 @@ class Locoformer(Module):
 
             # update critic
 
-            value_loss = F.mse_loss(returns, value, reduction = 'none')
+            value_loss = self.hl_gauss_loss(value_logits, returns, reduction = 'none')
 
             value_clip = self.ppo_value_clip
+            value = self.hl_gauss_loss(value_logits)
+
             clipped_value = old_value + (value - old_value).clamp(-value_clip, value_clip)
-            clipped_value_loss = F.mse_loss(returns, clipped_value, reduction = 'none')
+            clipped_value_loss = self.hl_gauss_loss(clipped_value, returns, reduction = 'none')
 
             critic_loss = torch.maximum(value_loss, clipped_value_loss) * self.value_loss_weight
 
@@ -750,28 +948,48 @@ class Locoformer(Module):
 
         return mean_actor_loss.detach(), mean_critic_loss.detach()
 
+    def state_to_rewards(
+        self,
+        state
+    ) -> Tensor:
+
+        assert self.has_reward_shaping
+
+        rewards = [fn(state) for fn in self.reward_shaping_fns]
+
+        rewards = [tensor(reward) if not is_tensor(reward) else reward for reward in rewards]
+        return stack(rewards)
+
     def wrap_env_functions(self, env):
 
-        def wrapped_reset(*args, **kwargs):
-            state, _ =  env.reset(*args, **kwargs)
+        def transform_output(el):
+            if isinstance(el, ndarray):
+                return from_numpy(el)
+            elif isinstance(el, (int, bool, float)):
+                return tensor(el)
+            else:
+                return el
 
-            if isinstance(state, ndarray):
-                state = from_numpy(state)
+        def wrapped_reset(*args, **kwargs):
+            env_reset_out =  env.reset(*args, **kwargs)
 
-            return state, _
+            return tree_map(transform_output, env_reset_out)
 
         def wrapped_step(action, *args, **kwargs):
-            out = env.step(action.item(), *args, **kwargs)
 
-            def transform_output(el):
-                if isinstance(el, ndarray):
-                    return from_numpy(el)
-                elif isinstance(el, (int, bool, float)):
-                    return tensor(el)
-                else:
-                    return el
+            if is_tensor(action):
+                action = action.item()
+
+            env_step_out = env.step(action, *args, **kwargs)
 
-            return tree_map(transform_output, out)
+            env_step_out_torch = tree_map(transform_output, env_step_out)
+
+            if not self.has_reward_shaping:
+                return env_step_out_torch
+
+            shaped_rewards = self.state_to_rewards(env_step_out_torch)
+
+            return env_step_out_torch, shaped_rewards
 
         return wrapped_reset, wrapped_step
 
@@ -781,6 +999,7 @@ class Locoformer(Module):
         inference_mode = False,
         has_batch_dim = False,
         has_time_dim = False,
+        state_time_dim = 1,
         **kwargs
     ):
         window_size = self.window_size
@@ -796,23 +1015,16 @@ class Locoformer(Module):
                 state = rearrange(state, '... -> 1 ...')
 
             if not has_time_dim:
-                state = rearrange(state, '... d -> ... 1 d')
+                state = state.unsqueeze(state_time_dim)
 
             # forwards
 
             out, cache = self.forward(state, cache = cache, **{**kwargs, **override_kwargs})
 
-            # handle cache
-
-            cache_len = cache.shape[-2]
-
-            if self.fixed_window_size or divisible_by(cache_len, window_size * 2):
-                cache = cache[..., -window_size:, :]
-
             # maybe remove batch or time
 
             if not has_time_dim:
-                out = tree_map_tensor(out, lambda t: rearrange(t, '... 1 d -> ... d'))
+                out = tree_map_tensor(out, lambda t: t.squeeze(state_time_dim))
 
             if not has_batch_dim:
                 out = tree_map_tensor(out, lambda t: rearrange(t, '1 ... -> ...'))
@@ -841,16 +1053,35 @@ class Locoformer(Module):
     def forward(
         self,
         state: Tensor,
-        cache: Tensor | None = None,
+        cache: Cache | None = None,
         detach_cache = False,
-        return_values = False
+        return_values = False,
+        return_raw_value_logits = False
     ):
 
         state = state.to(self.device)
 
         tokens = self.embedder(state)
 
-        embed, kv_cache = self.transformer(tokens, cache = cache, return_kv_cache = True)
+        # time
+
+        time = tokens.shape[-2]
+
+        # destruct the cache for the current timestep and the cache
+
+        prev_kv_cache = None
+        timestep_start = 0
+
+        if exists(cache):
+            timestep_start, prev_kv_cache = cache
+
+        # an assert - make sure during training or inference, forward never gets anything that crosses the window segment boundary, to open up some possibilities with extending memory
+
+        assert ((timestep_start % self.window_size) + time) <= self.window_size
+
+        # attention
+
+        embed, kv_cache = self.transformer(tokens, cache = prev_kv_cache, return_kv_cache = True)
 
         # unembed to actions - in language models this would be the next state
 
@@ -866,16 +1097,29 @@ class Locoformer(Module):
         # handle returning of values
 
         if return_values:
-            assert exists(self.value_network)
+            assert exists(self.to_value_pred)
 
-            values = self.value_network(embed)
+            values = self.to_value_pred(embed)
 
-            if values.ndim == 3:
-                assert values.shape[-1] == 1
-                values = rearrange(values, '... 1 -> ...')
+            if not return_raw_value_logits:
+                values = self.hl_gauss_loss(values) # converts the value logits to scalar values
 
             out = (out, values)
 
         # output and cache
 
-        return out, kv_cache
+        next_timestep = time + timestep_start
+
+        # handle curtailing kv cache at the right intervals
+
+        window_size = self.window_size
+
+        if self.fixed_window_size or divisible_by(next_timestep, window_size * 2):
+            kv_cache = kv_cache[..., -window_size:, :]
+
+        # maybe recurrent cache - shift the kv cache from one layer above to the one below, for extending on receptive field of past
+
+        if self.recurrent_kv_cache and divisible_by(next_timestep, window_size):
+            kv_cache = torch.roll(kv_cache, shifts = -1, dims = 0)
+
+        return out, (next_timestep, kv_cache)

{locoformer-0.0.17 → locoformer-0.0.30}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "locoformer"
-version = "0.0.17"
+version = "0.0.30"
 description = "LocoFormer"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }
@@ -30,6 +30,7 @@ dependencies = [
     "beartype",
     "einx>=0.3.0",
     "einops>=0.8.0",
+    "hl-gauss-pytorch>=0.2.0",
     "rotary-embedding-torch",
     "torch>=2.4",
     "x-mlps-pytorch",

{locoformer-0.0.17 → locoformer-0.0.30}/tests/test_locoformer.py

@@ -2,18 +2,25 @@ import pytest
 param = pytest.mark.parametrize
 
 import torch
+from torch import nn
 from x_mlps_pytorch import MLP
 
 from einops import rearrange
 
-def test_locoformer():
-    from locoformer.locoformer import Locoformer
-    from torch import nn
+from locoformer.locoformer import Locoformer
+
+@param('recurrent_kv_cache', (False, True))
+def test_locoformer(
+    recurrent_kv_cache
+):
     
     model = Locoformer(
         embedder = nn.Embedding(256, 128),
         unembedder = nn.Linear(128, 256, bias = False),
-        value_network = MLP(128, 32, 1),
+        value_network = MLP(128, 64, 32),
+        dim_value_input = 32,
+        reward_range = (-100., 100.),
+        recurrent_kv_cache = recurrent_kv_cache,
         transformer = dict(
             dim = 128,
             depth = 1,
@@ -83,4 +90,58 @@ def test_replay():
 
     dataloader = replay_buffer.dataloader(batch_size = 3)
 
-    assert next(iter(dataloader))['state'].shape[0] == 3
+    assert next(iter(dataloader))['state'].shape[0] == 3
+
+    # we will now consider consecutive pairs of episodes as 2 trials to be used for in-context adaptation
+    # but realistically there will be a function that converts a given ReplayBuffer -> Int[batch, episode_indices]
+
+    from torch import stack, arange
+
+    episode_indices = arange(len(replay_buffer))
+    remapped_episodes = stack((episode_indices[:-1], episode_indices[1:]))
+
+    dataloader = replay_buffer.dataloader(
+        batch_size = 1,
+        episode_mapping = remapped_episodes
+    )
+
+    assert next(iter(dataloader))['_lens'][0] == (3 + 5) # first and second episodes are concatted together timewise
+
+def test_reward_shaping():
+
+    model = Locoformer(
+        embedder = nn.Embedding(256, 128),
+        unembedder = nn.Linear(128, 256, bias = False),
+        value_network = MLP(128, 64, 32),
+        dim_value_input = 32,
+        reward_range = (-100., 100.),
+        reward_shaping_fns = [
+            lambda state: (state[3] - 2.5).pow(2).mean(),
+            lambda state: state[4:6].norm(dim = -1)
+        ],
+        transformer = dict(
+            dim = 128,
+            depth = 1,
+            window_size = 512
+        )
+    )
+
+    import numpy as np
+
+    class MockEnv:
+        def reset(self):
+            return np.random.normal(size = (10,))
+
+        def step(self, *args, **kwargs):
+            return np.random.normal(size = (10,))
+
+
+    env = MockEnv()
+
+    reset_fn, step_fn = model.wrap_env_functions(env)
+
+    reset_fn()
+
+    _, rewards = step_fn(3)
+
+    assert len(rewards) == 2

{locoformer-0.0.17 → locoformer-0.0.30}/train.py

@@ -160,7 +160,7 @@ for i in range(NUM_BATCHES):
         optim.step()
         optim.zero_grad()
 
-    if divisible_by(i + 1, GENERATE_EVERY):
+    if divisible_by(i, GENERATE_EVERY):
         model.eval()
 
         val_seq = next(val_loader_iter)

{locoformer-0.0.17 → locoformer-0.0.30}/train_gym.py

@@ -25,7 +25,6 @@ import torch.nn.functional as F
 from torch.utils.data import TensorDataset, DataLoader
 from torch.optim import Adam
 
-import einx
 from einops import rearrange
 
 from locoformer.locoformer import Locoformer, ReplayBuffer
@@ -60,8 +59,6 @@ def learn(
     batch_size = 16,
     epochs = 2,
 ):
-    device = accelerator.device
-
     dl = replay.dataloader(batch_size = batch_size, shuffle = True)
     model, dl, actor_optim, critic_optim = accelerator.prepare(model, dl, actor_optim, critic_optim)
 
@@ -70,18 +67,14 @@ def learn(
 
             data = SimpleNamespace(**data)
 
-            seq_len = data.state.shape[1]
-
-            value_mask = einx.less('j, i -> i j', arange(seq_len, device = device), data._lens)
-            value = torch.where(value_mask, data.value, 0.)
-
             actor_loss, critic_loss = model.ppo(
                 state = data.state,
                 action = data.action,
                 old_action_log_prob = data.action_log_prob,
                 reward = data.reward,
-                old_value = value,
+                old_value = data.value,
                 mask = data.learnable,
+                episode_lens = data._lens,
                 actor_optim = actor_optim,
                 critic_optim = critic_optim
             )
@@ -94,7 +87,7 @@ def main(
     env_name = 'LunarLander-v3',
     num_episodes = 50_000,
     max_timesteps = 500,
-    num_episodes_before_learn = 32,
+    num_episodes_before_learn = 64,
     clear_video = True,
     video_folder = 'recordings',
     record_every_episode = 250,
@@ -105,7 +98,8 @@ def main(
     ppo_eps_clip = 0.2,
     ppo_entropy_weight = .01,
     batch_size = 16,
-    epochs = 2
+    epochs = 3,
+    reward_range = (-100., 100.)
 ):
 
     # accelerate
@@ -153,7 +147,6 @@ def main(
     locoformer = Locoformer(
         embedder = MLP(dim_state, 64, bias = False),
         unembedder = MLP(64, num_actions, bias = False),
-        value_network = MLP(64, 1, bias = False),
         transformer = dict(
             dim = 64,
             dim_head = 32,
@@ -165,16 +158,20 @@ def main(
         gae_lam = gae_lam,
         ppo_eps_clip = ppo_eps_clip,
         ppo_entropy_weight = ppo_entropy_weight,
+        use_spo = True,
+        value_network = MLP(64, 64),
+        dim_value_input = 64,
+        reward_range = reward_range,
+        hl_gauss_loss_kwargs = dict(),
+        recurrent_kv_cache = True,
         calc_gae_kwargs = dict(
             use_accelerated = False
-        )
+        ),
     ).to(device)
 
     optim_actor = Adam([*locoformer.transformer.parameters(), *locoformer.actor_parameters()], lr = learning_rate, betas = betas)
     optim_critic = Adam([*locoformer.transformer.parameters(), *locoformer.critic_parameters()], lr = learning_rate, betas = betas)
 
-    timesteps_learn = 0
-
     # able to wrap the env for all values to torch tensors and back
     # all environments should follow usual MDP interface, domain randomization should be given at instantiation
 
@@ -205,7 +202,8 @@ def main(
 
                 # append to memory
 
-                done = truncated or terminated
+                exceeds_max_timesteps = timestep == (max_timesteps - 1)
+                done = truncated or terminated or tensor(exceeds_max_timesteps)
 
                 # get log prob of action
 
@@ -222,23 +220,24 @@ def main(
                     learnable = tensor(True)
                 )
 
-                # handle bootstrap value, which is a non-learnable timestep added with the next value for GAE
-                # only if terminated signal not detected
+                # increment counters
 
-                if not terminated:
-                    _, next_value = stateful_forward(next_state, return_values = True)
+                timestep += 1
 
-                    memory._replace(value = next_value, learnable = False)
+                # break if done or exceed max timestep
 
-                    replay.store(**memory._asdict())
+                if done:
 
-                # increment counters
+                    # handle bootstrap value, which is a non-learnable timestep added with the next value for GAE
+                    # only if terminated signal not detected
 
-                timestep += 1
+                    if not terminated:
+                        _, next_value = stateful_forward(next_state, return_values = True)
 
-                # break if done or exceed max timestep
+                        memory._replace(value = next_value, learnable = False)
+
+                        replay.store(**memory._asdict())
 
-                if done or timestep >= max_timesteps:
                     break
 
                 state = next_state