locoformer 0.0.7__py3-none-any.whl → 0.0.29__py3-none-any.whl

locoformer/locoformer.py

@@ -1,11 +1,25 @@
 from __future__ import annotations
+from typing import Callable
 from functools import partial
 
+from pathlib import Path
+from contextlib import contextmanager
+from collections import namedtuple
+
+import numpy as np
+from numpy import ndarray
+from numpy.lib.format import open_memmap
+
+from beartype import beartype
+from beartype.door import is_bearable
+
 import torch
-from torch import nn, cat, stack, arange, Tensor, is_tensor
+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.data import Dataset, DataLoader
+from torch.optim import Optimizer
 
 import einx
 from einops import rearrange, einsum
@@ -13,10 +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):
@@ -31,20 +51,36 @@ def first(arr):
 def divisible_by(num, den):
     return (num % den) == 0
 
+# tensor helpers
+
+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)
 
-def detach_all(x):
-    return tree_map_tensor(x, lambda t: t.detach())
+def pad_at_dim(
+    t,
+    pad: tuple[int, int],
+    dim = -1,
+    value = 0.
+):
+    if pad == (0, 0):
+        return t
 
-def combine_kv_cache(cache1, cache2):
-    combined_cache = []
+    dims_from_right = (- dim - 1) if dim < 0 else (t.ndim - dim - 1)
+    zeros = ((0, 0) * dims_from_right)
+    return F.pad(t, (*zeros, *pad), value = value)
 
-    for layer_cache1, layer_cache2 in zip(cache1, cache2):
-        next_cache = cat((layer_cache1, layer_cache2), dim = -2)
-        combined_cache.append(next_cache)
+def normalize(t, eps = 1e-5):
+    return (t - t.mean()) / t.std().clamp_min(eps)
 
-    return combined_cache
+def calc_entropy(logits):
+    prob = logits.softmax(dim = -1)
+    return -(prob * log(prob)).sum(dim = -1)
 
 # generalized advantage estimate
 
@@ -52,7 +88,7 @@ def combine_kv_cache(cache1, cache2):
 def calc_gae(
     rewards,
     values,
-    masks,
+    masks = None,
     gamma = 0.99,
     lam = 0.95,
     use_accelerated = None
@@ -63,6 +99,9 @@ def calc_gae(
     values = F.pad(values, (0, 1), value = 0.)
     values, values_next = values[..., :-1], values[..., 1:]
 
+    if not exists(masks):
+        masks = torch.ones_like(values)
+
     delta = rewards + gamma * values_next * masks - values
     gates = gamma * lam * masks
 
@@ -72,7 +111,7 @@ def calc_gae(
 
     returns = gae + values
 
-    return returns
+    return gae, returns
 
 # transformer-xl mask w/ flex attn
 
@@ -114,8 +153,8 @@ def create_xl_mask(
         # handle intra-episodic attention if needed
 
         if exists(episode_ids):
-            q_episode = episodes[b, q + offset]
-            k_episode = episodes[b, k]
+            q_episode = episode_ids[b, q + offset]
+            k_episode = episode_ids[b, k]
 
             intra_episode_mask = q_episode == k_episode
             mask = mask & intra_episode_mask
@@ -146,6 +185,284 @@ def create_sliding_mask(
     create_kwargs = dict(device = device) if exists(device) else dict()
     return create_block_mask(sliding_mask, B = None, H = None, Q_LEN = seq_len, KV_LEN = kv_seq_len, _compile = True, **create_kwargs)
 
+# data
+
+def collate_var_time(data):
+
+    datum = first(data)
+    keys = datum.keys()
+
+    all_tensors = zip(*[datum.values() for datum in data])
+
+    collated_values = []
+
+    for key, tensors in zip(keys, all_tensors):
+
+        # the episode lens have zero dimension - think of a cleaner way to handle this later
+
+        if key != '_lens':
+
+            times = [t.shape[0] for t in tensors]
+            max_time = max(times)
+            tensors = [pad_at_dim(t, (0, max_time - t.shape[0]), dim = 0) for t in tensors]
+
+        collated_values.append(stack(tensors))
+
+    return dict(zip(keys, collated_values))
+
+class ReplayDataset(Dataset):
+    def __init__(
+        self,
+        folder: str | Path,
+        fields: tuple[str, ...] | None = None
+    ):
+        if isinstance(folder, str):
+            folder = Path(folder)
+
+        episode_lens = folder / 'episode_lens.npy'
+        self.episode_lens = open_memmap(str(episode_lens), mode = 'r')
+
+        # get indices of non-zero lengthed episodes
+
+        nonzero_episodes = self.episode_lens > 0
+        self.indices = np.arange(self.episode_lens.shape[-1])[nonzero_episodes]
+
+        # get all data files
+
+        filepaths = [*folder.glob('*.data.npy')]
+        assert len(filepaths) > 0
+
+        fieldname_to_filepath = {path.name.split('.')[0]: path for path in filepaths}
+
+        fieldnames_from_files = set(fieldname_to_filepath.keys())
+
+        fields = default(fields, fieldnames_from_files)
+
+        self.memmaps = dict()
+
+        for field in fields:
+            assert field in fieldnames_from_files, f'invalid field {field} - must be one of {fieldnames_from_files}'
+
+            path = fieldname_to_filepath[field]
+
+            self.memmaps[field] = open_memmap(str(path), mode = 'r')
+
+    def __len__(self):
+        return len(self.indices)
+
+    def __getitem__(self, idx):
+        episode_index = self.indices[idx]
+
+        episode_len = self.episode_lens[episode_index]
+
+        data = {field: from_numpy(memmap[episode_index, :episode_len].copy()) for field, memmap in self.memmaps.items()}
+
+        data['_lens'] = tensor(episode_len)
+
+        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
+    def __init__(
+        self,
+        folder: str | Path,
+        max_episodes: int,
+        max_timesteps: int,
+        fields: dict[
+            str,
+            str | tuple[str, int | tuple[int, ...]]
+        ]
+    ):
+
+        # folder for data
+
+        if not isinstance(folder, Path):
+            folder = Path(folder)
+            folder.mkdir(exist_ok = True)
+
+        self.folder = folder
+        assert folder.is_dir()
+
+        # keeping track of episode length
+
+        episode_lens = folder / 'episode_lens.npy'
+
+        self.episode_index = 0
+        self.timestep_index = 0
+
+        self.max_episodes = max_episodes
+        self.max_timesteps= max_timesteps
+
+        self.episode_lens = open_memmap(str(episode_lens), mode = 'w+', dtype = np.int32, shape = (max_episodes,))
+
+        # create the memmap for individual data tracks
+
+        self.shapes = dict()
+        self.dtypes = dict()
+        self.memmaps = dict()
+        self.fieldnames = set(fields.keys())
+
+        for field_name, field_info in fields.items():
+
+            # some flexibility
+
+            field_info = (field_info, ()) if isinstance(field_info, str) else field_info
+
+            dtype_str, shape = field_info
+            assert dtype_str in {'int', 'float', 'bool'}
+
+            dtype = dict(int = np.int32, float = np.float32, bool = np.bool_)[dtype_str]
+
+            # memmap file
+
+            filepath = folder / f'{field_name}.data.npy'
+            memmap = open_memmap(str(filepath), mode = 'w+', dtype = dtype, shape = (max_episodes, max_timesteps, *shape))
+
+            self.memmaps[field_name] = memmap
+            self.shapes[field_name] = shape
+            self.dtypes[field_name] = dtype
+
+        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
+        self.timestep_index = 0
+
+    def advance_episode(self):
+        self.episode_index = (self.episode_index + 1) % self.max_episodes
+        self.timestep_index = 0
+
+    def flush(self):
+        self.episode_lens[self.episode_index] = self.timestep_index
+
+        for memmap in self.memmaps.values():
+            memmap.flush()
+
+        self.episode_lens.flush()
+
+    @contextmanager
+    def one_episode(self):
+
+        yield
+
+        self.flush()
+        self.advance_episode()
+
+    @beartype
+    def store_datapoint(
+        self,
+        episode_index: int,
+        timestep_index: int,
+        name: str,
+        datapoint: Tensor | ndarray
+    ):
+        assert 0 <= episode_index < self.max_episodes
+        assert 0 <= timestep_index < self.max_timesteps
+
+        if is_tensor(datapoint):
+            datapoint = datapoint.detach().cpu().numpy()
+
+        assert name in self.fieldnames, f'invalid field name {name} - must be one of {self.fieldnames}'
+
+        assert datapoint.shape == self.shapes[name], f'invalid shape {datapoint.shape} - shape must be {self.shapes[name]}'
+
+        self.memmaps[name][self.episode_index, self.timestep_index] = datapoint
+
+    def store(
+        self,
+        **data
+    ):
+        assert is_bearable(data, dict[str, Tensor | ndarray])
+
+        assert not self.timestep_index >= self.max_timesteps, 'you exceeded the `max_timesteps` set on the replay buffer'
+
+        for name, datapoint in data.items():
+
+            self.store_datapoint(self.episode_index, self.timestep_index, name, datapoint)
+
+        self.timestep_index += 1
+
+        return self.memory_namedtuple(**data)
+
+    def dataset(
+        self,
+        episode_mapping: Tensor | list[list[int]] | None = None,
+    ) -> Dataset:
+        self.flush()
+
+        dataset = ReplayDataset(self.folder)
+
+        if not exists(episode_mapping):
+            return dataset
+
+        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(episode_mapping), batch_size = batch_size, collate_fn = collate_var_time, **kwargs)
+
 # transformer-xl with ppo
 
 class Attention(Module):
@@ -204,7 +521,6 @@ class Attention(Module):
         return_kv_cache = False,
     ):
         seq_len = tokens.shape[-2]
-        assert seq_len <= self.window_size
 
         device = tokens.device
 
@@ -365,7 +681,21 @@ 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(),
+        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__()
 
@@ -377,11 +707,58 @@ 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
+        self.gae_lam = gae_lam
+        self.ppo_eps_clip = ppo_eps_clip
+        self.ppo_entropy_weight = ppo_entropy_weight
+        self.ppo_value_clip = ppo_value_clip
+        self.value_loss_weight = value_loss_weight
+
+        self.calc_gae_kwargs = calc_gae_kwargs
+
+        # maybe use spo
+
+        self.use_spo = use_spo
+
+        # maybe recurrent kv cache (todo: find and cite this paper from ages ago)
+
+        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)
+
     @property
     def device(self):
         return next(self.parameters()).device
@@ -390,7 +767,165 @@ class Locoformer(Module):
         return self.unembedder.parameters()
 
     def critic_parameters(self):
-        return self.value_network.parameters()
+        if not exists(self.to_value_pred):
+            return []
+
+        return self.to_value_pred.parameters()
+
+    def ppo(
+        self,
+        state,
+        action,
+        old_action_log_prob,
+        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
+            (
+                state,
+                action,
+                old_action_log_prob,
+                reward,
+                old_value,
+                mask,
+                advantage,
+                returns
+            )
+        ]
+
+        mean_actor_loss = self.zero.clone()
+        mean_critic_loss = self.zero.clone()
+
+        # learn across windows
+
+        cache = None
+
+        for (
+            state,
+            action,
+            old_action_log_prob,
+            reward,
+            old_value,
+            mask,
+            advantage,
+            returns
+        ) in zip(*windowed_tensors):
+
+            (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')
+            log_prob = action_logits.gather(-1, action)
+            log_prob = rearrange(log_prob, 'b t 1 -> b t')
+
+            # update actor, classic clipped surrogate loss
+
+            eps_clip = self.ppo_eps_clip
+            ratio = (log_prob - old_action_log_prob).exp()
+
+            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
+
+            windowed_actor_loss = actor_loss[mask].sum() / total_learnable_tokens
+            windowed_actor_loss.backward(retain_graph = True)
+
+            # update critic
+
+            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 = self.hl_gauss_loss(clipped_value, returns, reduction = 'none')
+
+            critic_loss = torch.maximum(value_loss, clipped_value_loss) * self.value_loss_weight
+
+            windowed_critic_loss = critic_loss[mask].sum() / total_learnable_tokens
+            windowed_critic_loss.backward(retain_graph = True)
+
+            # accumulate
+
+            mean_actor_loss.add_(windowed_actor_loss)
+            mean_critic_loss.add_(windowed_critic_loss)
+
+        # optimizer update
+
+        if exists(actor_optim):
+            actor_optim.step()
+            actor_optim.zero_grad()
+
+        if exists(critic_optim):
+            critic_optim.step()
+            critic_optim.zero_grad()
+
+        # return losses for logging
+
+        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 transform_output(el):
+            if isinstance(el, ndarray):
+                return from_numpy(el)
+            elif isinstance(el, (int, bool, float)):
+                return tensor(el)
+            else:
+                return el
+
+        def wrapped_reset(*args, **kwargs):
+            env_reset_out =  env.reset(*args, **kwargs)
+
+            return tree_map(transform_output, env_reset_out)
+
+        def wrapped_step(action, *args, **kwargs):
+
+            if is_tensor(action):
+                action = action.item()
+
+            env_step_out = env.step(action, *args, **kwargs)
+
+            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
 
     def get_stateful_forward(
         self,
@@ -398,6 +933,7 @@ class Locoformer(Module):
         inference_mode = False,
         has_batch_dim = False,
         has_time_dim = False,
+        state_time_dim = 1,
         **kwargs
     ):
         window_size = self.window_size
@@ -413,23 +949,16 @@ 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 ... -> ...'))
@@ -458,14 +987,35 @@ 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
 
@@ -476,21 +1026,34 @@ class Locoformer(Module):
         # maybe detach cache
 
         if detach_cache:
-            kv_cache = detach_all(kv_cache)
+            kv_cache = kv_cache.detach()
 
         # 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-0.0.7.dist-info → locoformer-0.0.29.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: locoformer
-Version: 0.0.7
+Version: 0.0.29
 Summary: LocoFormer
 Project-URL: Homepage, https://pypi.org/project/locoformer/
 Project-URL: Repository, https://github.com/lucidrains/locoformer
@@ -35,8 +35,10 @@ Classifier: Programming Language :: Python :: 3.9
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.9
 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
@@ -53,7 +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-0.0.29.dist-info/RECORD

@@ -0,0 +1,6 @@
+locoformer/__init__.py,sha256=XctsMGEZSR4mVl75fhds_1BtS5qGFiiItTDV7CmCt_I,45
+locoformer/locoformer.py,sha256=Tr_1btuoTZ0huXeDcAeuHxTPaVeCUEGc5iLvMYGDLck,29982
+locoformer-0.0.29.dist-info/METADATA,sha256=5Fi3EOsgpBvpzAFVZQyrlink-HcHE8EgFl10Y5l8mqM,3256
+locoformer-0.0.29.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+locoformer-0.0.29.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+locoformer-0.0.29.dist-info/RECORD,,

locoformer-0.0.7.dist-info/RECORD

@@ -1,6 +0,0 @@
-locoformer/__init__.py,sha256=XctsMGEZSR4mVl75fhds_1BtS5qGFiiItTDV7CmCt_I,45
-locoformer/locoformer.py,sha256=lJQs0CKr9iztF8tie1FRUVEItCt-IZbIILQqKcgK2sI,13142
-locoformer-0.0.7.dist-info/METADATA,sha256=PZ_phKV3t4Bha0GnUB5HPmE9w8A5fvNevsuN532Ls3s,3193
-locoformer-0.0.7.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-locoformer-0.0.7.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-locoformer-0.0.7.dist-info/RECORD,,