locoformer 0.0.6__tar.gz → 0.0.17__tar.gz

{locoformer-0.0.6 → locoformer-0.0.17}/.gitignore

@@ -1,3 +1,5 @@
+recordings/
+replay/
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

{locoformer-0.0.6 → locoformer-0.0.17}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: locoformer
-Version: 0.0.6
+Version: 0.0.17
 Summary: LocoFormer
 Project-URL: Homepage, https://pypi.org/project/locoformer/
 Project-URL: Repository, https://github.com/lucidrains/locoformer
@@ -35,6 +35,7 @@ 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: rotary-embedding-torch
@@ -53,7 +54,7 @@ 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). 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) 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.
 
 ## Sponsors
 

{locoformer-0.0.6 → locoformer-0.0.17}/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). 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) 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.
 
 ## Sponsors
 

{locoformer-0.0.6 → locoformer-0.0.17}/locoformer/locoformer.py

@@ -1,11 +1,24 @@
 from __future__ import annotations
 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, 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
@@ -15,6 +28,8 @@ from rotary_embedding_torch import RotaryEmbedding
 
 from assoc_scan import AssocScan
 
+# constants
+
 LinearNoBias = partial(Linear, bias = False)
 
 # helper functions
@@ -31,20 +46,30 @@ 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 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 combine_kv_cache(cache1, cache2):
-    combined_cache = []
+def pad_at_dim(
+    t,
+    pad: tuple[int, int],
+    dim = -1,
+    value = 0.
+):
+    if pad == (0, 0):
+        return t
 
-    for layer_cache1, layer_cache2 in zip(cache1, cache2):
-        next_cache = cat((layer_cache1, layer_cache2), dim = -2)
-        combined_cache.append(next_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)
 
-    return combined_cache
+def calc_entropy(logits):
+    prob = logits.softmax(dim = -1)
+    return -(prob * log(prob)).sum(dim = -1)
 
 # generalized advantage estimate
 
@@ -52,7 +77,7 @@ 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 +88,9 @@ 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 +100,7 @@ def calc_gae(
 
     returns = gae + values
 
-    return returns
+    return gae, returns
 
 # transformer-xl mask w/ flex attn
 
@@ -114,8 +142,8 @@ 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 +174,217 @@ 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 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 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) -> Dataset:
+        self.flush()
+
+        return ReplayDataset(self.folder)
+
+    def dataloader(self, batch_size, **kwargs) -> DataLoader:
+        self.flush()
+
+        return DataLoader(self.dataset(), batch_size = batch_size, collate_fn = collate_var_time, **kwargs)
+
 # transformer-xl with ppo
 
 class Attention(Module):
@@ -204,7 +443,6 @@ class Attention(Module):
         return_kv_cache = False,
     ):
         seq_len = tokens.shape[-2]
-        assert seq_len <= self.window_size
 
         device = tokens.device
 
@@ -365,7 +603,14 @@ class Locoformer(Module):
         embedder: Module,
         unembedder: Module,
         transformer: dict | TransformerXL,
-        value_network: Module | None = None
+        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,
+        value_loss_weight = 0.5,
+        calc_gae_kwargs: dict = dict()
     ):
         super().__init__()
 
@@ -382,15 +627,160 @@ class Locoformer(Module):
         self.fixed_window_size = transformer.fixed_window_size
         self.window_size = transformer.window_size
 
+        # 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
+
+        # loss related
+
+        self.register_buffer('zero', tensor(0.), persistent = False)
+
     @property
     def device(self):
         return next(self.parameters()).device
 
+    def actor_parameters(self):
+        return self.unembedder.parameters()
+
+    def critic_parameters(self):
+        if not exists(self.value_network):
+            return []
+
+        return self.value_network.parameters()
+
+    def ppo(
+        self,
+        state,
+        action,
+        old_action_log_prob,
+        reward,
+        old_value,
+        mask,
+        actor_optim: Optimizer | None = None,
+        critic_optim: Optimizer | None = None
+    ):
+        window_size = self.window_size
+        total_learnable_tokens = mask.sum().item()
+
+        windowed_tensors = [
+            t.split(window_size, dim = 1) for t in
+            (
+                state,
+                action,
+                old_action_log_prob,
+                reward,
+                old_value,
+                mask
+            )
+        ]
+
+        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
+        ) in zip(*windowed_tensors):
+
+            (action_logits, value), cache = self.forward(state, cache = cache, detach_cache = True, return_values = 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()
+
+            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)
+
+            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 = F.mse_loss(returns, value, reduction = 'none')
+
+            value_clip = self.ppo_value_clip
+            clipped_value = old_value + (value - old_value).clamp(-value_clip, value_clip)
+            clipped_value_loss = F.mse_loss(returns, clipped_value, 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 wrap_env_functions(self, env):
+
+        def wrapped_reset(*args, **kwargs):
+            state, _ =  env.reset(*args, **kwargs)
+
+            if isinstance(state, ndarray):
+                state = from_numpy(state)
+
+            return state, _
+
+        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
+
+            return tree_map(transform_output, out)
+
+        return wrapped_reset, wrapped_step
+
     def get_stateful_forward(
         self,
         initial_states: Tensor | None = None,
         inference_mode = False,
         has_batch_dim = False,
+        has_time_dim = False,
         **kwargs
     ):
         window_size = self.window_size
@@ -400,11 +790,14 @@ class Locoformer(Module):
         def stateful_forward(state: Tensor, **override_kwargs):
             nonlocal cache
 
-            # handle no batch, for easier time rolling out against envs
+            # handle no batch or time, for easier time rolling out against envs
 
             if not has_batch_dim:
                 state = rearrange(state, '... -> 1 ...')
 
+            if not has_time_dim:
+                state = rearrange(state, '... d -> ... 1 d')
+
             # forwards
 
             out, cache = self.forward(state, cache = cache, **{**kwargs, **override_kwargs})
@@ -416,7 +809,10 @@ class Locoformer(Module):
             if self.fixed_window_size or divisible_by(cache_len, window_size * 2):
                 cache = cache[..., -window_size:, :]
 
-            # maybe remove batch
+            # maybe remove batch or time
+
+            if not has_time_dim:
+                out = tree_map_tensor(out, lambda t: rearrange(t, '... 1 d -> ... d'))
 
             if not has_batch_dim:
                 out = tree_map_tensor(out, lambda t: rearrange(t, '1 ... -> ...'))
@@ -450,6 +846,8 @@ class Locoformer(Module):
         return_values = False
     ):
 
+        state = state.to(self.device)
+
         tokens = self.embedder(state)
 
         embed, kv_cache = self.transformer(tokens, cache = cache, return_kv_cache = True)
@@ -463,7 +861,7 @@ class Locoformer(Module):
         # maybe detach cache
 
         if detach_cache:
-            kv_cache = detach_all(kv_cache)
+            kv_cache = kv_cache.detach()
 
         # handle returning of values
 

{locoformer-0.0.6 → locoformer-0.0.17}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "locoformer"
-version = "0.0.6"
+version = "0.0.17"
 description = "LocoFormer"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }
@@ -27,6 +27,7 @@ classifiers=[
 
 dependencies = [
     "assoc-scan",
+    "beartype",
     "einx>=0.3.0",
     "einops>=0.8.0",
     "rotary-embedding-torch",

locoformer-0.0.17/tests/test_locoformer.py

@@ -0,0 +1,86 @@
+import pytest
+param = pytest.mark.parametrize
+
+import torch
+from x_mlps_pytorch import MLP
+
+from einops import rearrange
+
+def test_locoformer():
+    from locoformer.locoformer import Locoformer
+    from torch import nn
+    
+    model = Locoformer(
+        embedder = nn.Embedding(256, 128),
+        unembedder = nn.Linear(128, 256, bias = False),
+        value_network = MLP(128, 32, 1),
+        transformer = dict(
+            dim = 128,
+            depth = 1,
+            window_size = 512
+        )
+    )
+
+    seq = torch.randint(0, 256, (3, 512))
+
+    (logits, values), cache = model(seq, return_values = True)
+    (logits, values), cache = model(seq, return_values = True, cache = cache)
+    (logits, values), cache = model(seq, return_values = True, cache = cache)
+
+    assert logits.shape == (3, 512, 256)
+
+    stateful_forward = model.get_stateful_forward(has_batch_dim = True, has_time_dim = True, return_values = True, inference_mode = True)
+
+    for state in seq.unbind(dim = -1):
+        state = rearrange(state, 'b -> b 1')
+
+        logits, values = stateful_forward(state)
+        assert logits.shape == (3, 1, 256)
+
+def test_replay():
+    from locoformer.locoformer import ReplayBuffer
+
+    replay_buffer = ReplayBuffer(
+        './replay_data',
+        max_episodes = 10_000,
+        max_timesteps = 501,
+        fields = dict(
+            state = ('float', (8,)),
+            action = 'int',
+            action_log_prob = 'float',
+            reward = 'float',
+            value = 'float',
+            done = 'bool'
+        )
+    )
+
+    lens = [3, 5, 4]
+
+    for episode_len in lens:
+        with replay_buffer.one_episode():
+            for _ in range(episode_len):
+                state = torch.randn((8,))
+                action = torch.randint(0, 4, ())
+                log_prob = torch.randn(())
+                reward = torch.randn(())
+                value = torch.randn(())
+                done = torch.randint(0, 2, ()).bool()
+
+                replay_buffer.store(
+                    state = state,
+                    action = action,
+                    action_log_prob = log_prob,
+                    reward = reward,
+                    value = value,
+                    done = done
+                )
+
+    dataset = replay_buffer.dataset()
+
+    assert len(dataset) == 3
+
+    assert torch.is_tensor(dataset[0]['state'])
+
+    dataloader = replay_buffer.dataloader(batch_size = 3)
+
+    assert next(iter(dataloader))['state'].shape[0] == 3

{locoformer-0.0.6 → locoformer-0.0.17}/train.py

@@ -169,7 +169,7 @@ for i in range(NUM_BATCHES):
         prime = prime.to(model.device)
         out = prime
 
-        stateful_forward, logits = model.get_stateful_forward(has_batch_dim = False, initial_states = prime, inference_mode = True)
+        stateful_forward, logits = model.get_stateful_forward(has_batch_dim = False, has_time_dim = True, initial_states = prime, inference_mode = True)
 
         # sample
 

locoformer-0.0.17/train_gym.py

@@ -0,0 +1,262 @@
+# /// script
+# dependencies = [
+#     "accelerate",
+#     "fire",
+#     "gymnasium[box2d]>=1.0.0",
+#     "locoformer>=0.0.12",
+#     "moviepy",
+#     "tqdm"
+# ]
+# ///
+
+from fire import Fire
+from shutil import rmtree
+from tqdm import tqdm
+from collections import deque
+from types import SimpleNamespace
+
+from accelerate import Accelerator
+
+import gymnasium as gym
+
+import torch
+from torch import from_numpy, randint, tensor, stack, arange
+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
+from x_mlps_pytorch import MLP
+
+# helper functions
+
+def exists(v):
+    return v is not None
+
+def divisible_by(num, den):
+    return (num % den) == 0
+
+def log(t, eps = 1e-20):
+    return t.clamp(min = eps).log()
+
+def gumbel_noise(t):
+    return -log(-log(torch.rand_like(t)))
+
+def gumbel_sample(logits, temperature = 1., eps = 1e-6):
+    noise = gumbel_noise(logits)
+    return ((logits / max(temperature, eps)) + noise).argmax(dim = -1)
+
+# learn
+
+def learn(
+    model,
+    actor_optim,
+    critic_optim,
+    accelerator,
+    replay,
+    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)
+
+    for _ in range(epochs):
+        for data in dl:
+
+            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,
+                mask = data.learnable,
+                actor_optim = actor_optim,
+                critic_optim = critic_optim
+            )
+
+            accelerator.print(f'actor: {actor_loss.item():.3f} | critic: {critic_loss.item():.3f}')
+
+# main function
+
+def main(
+    env_name = 'LunarLander-v3',
+    num_episodes = 50_000,
+    max_timesteps = 500,
+    num_episodes_before_learn = 32,
+    clear_video = True,
+    video_folder = 'recordings',
+    record_every_episode = 250,
+    learning_rate = 8e-4,
+    discount_factor = 0.99,
+    betas = (0.9, 0.99),
+    gae_lam = 0.95,
+    ppo_eps_clip = 0.2,
+    ppo_entropy_weight = .01,
+    batch_size = 16,
+    epochs = 2
+):
+
+    # accelerate
+
+    accelerator = Accelerator()
+    device = accelerator.device
+
+    # environment
+
+    env = gym.make(env_name, render_mode = 'rgb_array')
+
+    if clear_video:
+        rmtree(video_folder, ignore_errors = True)
+
+    env = gym.wrappers.RecordVideo(
+        env = env,
+        video_folder = video_folder,
+        name_prefix = 'lunar-video',
+        episode_trigger = lambda eps: divisible_by(eps, record_every_episode),
+        disable_logger = True
+    )
+
+    dim_state = env.observation_space.shape[0]
+    num_actions = env.action_space.n
+
+    # memory
+
+    replay = ReplayBuffer(
+        'replay',
+        num_episodes,
+        max_timesteps + 1, # one extra node for bootstrap node - not relevant for locoformer, but for completeness
+        fields = dict(
+            state = ('float', (dim_state,)),
+            action = 'int',
+            action_log_prob = 'float',
+            reward = 'float',
+            value = 'float',
+            done = 'bool',
+            learnable = 'bool'
+        )
+    )
+
+    # networks
+
+    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,
+            heads = 4,
+            depth = 4,
+            window_size = 16
+        ),
+        discount_factor = discount_factor,
+        gae_lam = gae_lam,
+        ppo_eps_clip = ppo_eps_clip,
+        ppo_entropy_weight = ppo_entropy_weight,
+        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
+
+    env_reset, env_step = locoformer.wrap_env_functions(env)
+
+    # loop
+
+    for episodes_index in tqdm(range(num_episodes)):
+
+        state, *_ = env_reset()
+
+        timestep = 0
+
+        stateful_forward = locoformer.get_stateful_forward(has_batch_dim = False, has_time_dim = False, inference_mode = True)
+
+        with replay.one_episode():
+            while True:
+
+                # predict next action
+
+                action_logits, value = stateful_forward(state, return_values = True)
+
+                action = gumbel_sample(action_logits)
+
+                # pass to environment
+
+                next_state, reward, truncated, terminated, *_ = env_step(action)
+
+                # append to memory
+
+                done = truncated or terminated
+
+                # get log prob of action
+
+                action_log_prob = action_logits.gather(-1, rearrange(action, '-> 1'))
+                action_log_prob = rearrange(action_log_prob, '1 ->')
+
+                memory = replay.store(
+                    state = state,
+                    action = action,
+                    action_log_prob = action_log_prob,
+                    reward = reward,
+                    value = value,
+                    done = done,
+                    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
+
+                if not terminated:
+                    _, next_value = stateful_forward(next_state, return_values = True)
+
+                    memory._replace(value = next_value, learnable = False)
+
+                    replay.store(**memory._asdict())
+
+                # increment counters
+
+                timestep += 1
+
+                # break if done or exceed max timestep
+
+                if done or timestep >= max_timesteps:
+                    break
+
+                state = next_state
+
+            # learn if hit the number of learn timesteps
+
+            if divisible_by(episodes_index + 1, num_episodes_before_learn):
+
+                learn(
+                    locoformer,
+                    optim_actor,
+                    optim_critic,
+                    accelerator,
+                    replay,
+                    batch_size,
+                    epochs,
+                )
+# main
+
+if __name__ == '__main__':
+    Fire(main)

locoformer-0.0.6/tests/test_locoformer.py

@@ -1,38 +0,0 @@
-import pytest
-param = pytest.mark.parametrize
-
-import torch
-from x_mlps_pytorch import MLP
-
-from einops import rearrange
-
-def test_locoformer():
-    from locoformer.locoformer import Locoformer
-    from torch import nn
-    
-    model = Locoformer(
-        embedder = nn.Embedding(256, 128),
-        unembedder = nn.Linear(128, 256, bias = False),
-        value_network = MLP(128, 32, 1),
-        transformer = dict(
-            dim = 128,
-            depth = 1,
-            window_size = 256
-        )
-    )
-
-    seq = torch.randint(0, 256, (3, 512))
-
-    (logits, values), cache = model(seq, return_values = True)
-    (logits, values), cache = model(seq, return_values = True, cache = cache)
-    (logits, values), cache = model(seq, return_values = True, cache = cache)
-
-    assert logits.shape == (3, 512, 256)
-
-    stateful_forward = model.get_stateful_forward(256, has_batch_dim = True, return_values = True, inference_mode = True)
-
-    for state in seq.unbind(dim = -1):
-        state = rearrange(state, 'b -> b 1')
-
-        logits, values = stateful_forward(state)
-        assert logits.shape == (3, 1, 256)