torchrl 0.11.0__cp314-cp314t-macosx_11_0_arm64.whl

sota-implementations/ddpg/ddpg.py

@@ -0,0 +1,231 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+"""DDPG Example.
+
+This is a simple self-contained example of a DDPG training script.
+
+It supports state environments like MuJoCo.
+
+The helper functions are coded in the utils.py associated with this script.
+"""
+from __future__ import annotations
+
+import warnings
+
+import hydra
+import numpy as np
+import torch
+import torch.cuda
+import tqdm
+from tensordict import TensorDict
+from tensordict.nn import CudaGraphModule
+from torchrl._utils import get_available_device, timeit
+from torchrl.envs.utils import ExplorationType, set_exploration_type
+from torchrl.objectives import group_optimizers
+from torchrl.record.loggers import generate_exp_name, get_logger
+from utils import (
+    dump_video,
+    log_metrics,
+    make_collector,
+    make_ddpg_agent,
+    make_environment,
+    make_loss_module,
+    make_optimizer,
+    make_replay_buffer,
+)
+
+
+@hydra.main(version_base="1.1", config_path="", config_name="config")
+def main(cfg: DictConfig):  # noqa: F821
+    device = (
+        torch.device(cfg.optim.device) if cfg.optim.device else get_available_device()
+    )
+    collector_device = (
+        torch.device(cfg.collector.device)
+        if cfg.collector.device
+        else get_available_device()
+    )
+
+    # Create logger
+    exp_name = generate_exp_name("DDPG", cfg.logger.exp_name)
+    logger = None
+    if cfg.logger.backend:
+        logger = get_logger(
+            logger_type=cfg.logger.backend,
+            logger_name="ddpg_logging",
+            experiment_name=exp_name,
+            wandb_kwargs={
+                "mode": cfg.logger.mode,
+                "config": dict(cfg),
+                "project": cfg.logger.project_name,
+                "group": cfg.logger.group_name,
+            },
+        )
+
+    # Set seeds
+    torch.manual_seed(cfg.env.seed)
+    np.random.seed(cfg.env.seed)
+
+    # Create environments
+    train_env, eval_env = make_environment(cfg, logger=logger)
+
+    # Create agent
+    model, exploration_policy = make_ddpg_agent(cfg, train_env, eval_env, device)
+
+    # Create DDPG loss
+    loss_module, target_net_updater = make_loss_module(cfg, model)
+
+    compile_mode = None
+    if cfg.compile.compile:
+        if cfg.compile.compile_mode not in (None, ""):
+            compile_mode = cfg.compile.compile_mode
+        elif cfg.compile.cudagraphs:
+            compile_mode = "default"
+        else:
+            compile_mode = "reduce-overhead"
+
+    # Create off-policy collector
+    collector = make_collector(
+        cfg,
+        train_env,
+        exploration_policy,
+        compile=cfg.compile.compile,
+        compile_mode=compile_mode,
+        cudagraph=cfg.compile.cudagraphs,
+        device=collector_device,
+    )
+
+    # Create replay buffer
+    replay_buffer = make_replay_buffer(
+        batch_size=cfg.optim.batch_size,
+        prb=cfg.replay_buffer.prb,
+        buffer_size=cfg.replay_buffer.size,
+        scratch_dir=cfg.replay_buffer.scratch_dir,
+        device="cpu",
+    )
+
+    # Create optimizers
+    optimizer_actor, optimizer_critic = make_optimizer(cfg, loss_module)
+    optimizer = group_optimizers(optimizer_actor, optimizer_critic)
+
+    def update(sampled_tensordict):
+        optimizer.zero_grad(set_to_none=True)
+
+        td_loss: TensorDict = loss_module(sampled_tensordict)
+        td_loss.sum(reduce=True).backward()
+        optimizer.step()
+
+        # Update qnet_target params
+        target_net_updater.step()
+        return td_loss.detach()
+
+    if cfg.compile.compile:
+        update = torch.compile(update, mode=compile_mode)
+    if cfg.compile.cudagraphs:
+        warnings.warn(
+            "CudaGraphModule is experimental and may lead to silently wrong results. Use with caution.",
+            category=UserWarning,
+        )
+        update = CudaGraphModule(update, warmup=50)
+
+    # Main loop
+    collected_frames = 0
+    pbar = tqdm.tqdm(total=cfg.collector.total_frames)
+
+    init_random_frames = cfg.collector.init_random_frames
+    num_updates = int(cfg.collector.frames_per_batch * cfg.optim.utd_ratio)
+    prb = cfg.replay_buffer.prb
+    frames_per_batch = cfg.collector.frames_per_batch
+    eval_iter = cfg.logger.eval_iter
+    eval_rollout_steps = cfg.env.max_episode_steps
+
+    c_iter = iter(collector)
+    total_iter = len(collector)
+    for _ in range(total_iter):
+        timeit.printevery(1000, total_iter, erase=True)
+        with timeit("collecting"):
+            tensordict = next(c_iter)
+        # Update exploration policy
+        exploration_policy[1].step(tensordict.numel())
+
+        # Update weights of the inference policy
+        collector.update_policy_weights_()
+
+        current_frames = tensordict.numel()
+        pbar.update(current_frames)
+
+        # Add to replay buffer
+        with timeit("rb - extend"):
+            tensordict = tensordict.reshape(-1)
+            replay_buffer.extend(tensordict)
+
+        collected_frames += current_frames
+
+        # Optimization steps
+        if collected_frames >= init_random_frames:
+            tds = []
+            for _ in range(num_updates):
+                # Sample from replay buffer
+                with timeit("rb - sample"):
+                    sampled_tensordict = replay_buffer.sample().to(device)
+                with timeit("update"):
+                    torch.compiler.cudagraph_mark_step_begin()
+                    td_loss = update(sampled_tensordict)
+                tds.append(td_loss.clone())
+
+                # Update priority
+                if prb:
+                    replay_buffer.update_priority(sampled_tensordict)
+            tds = torch.stack(tds)
+
+        episode_end = (
+            tensordict["next", "done"]
+            if tensordict["next", "done"].any()
+            else tensordict["next", "truncated"]
+        )
+        episode_rewards = tensordict["next", "episode_reward"][episode_end]
+
+        # Logging
+        metrics_to_log = {}
+        if len(episode_rewards) > 0:
+            episode_length = tensordict["next", "step_count"][episode_end]
+            metrics_to_log["train/reward"] = episode_rewards.mean().item()
+            metrics_to_log["train/episode_length"] = episode_length.sum().item() / len(
+                episode_length
+            )
+
+        if collected_frames >= init_random_frames:
+            tds = TensorDict(train=tds).flatten_keys("/").mean()
+            metrics_to_log.update(tds.to_dict())
+
+        # Evaluation
+        if abs(collected_frames % eval_iter) < frames_per_batch:
+            with set_exploration_type(
+                ExplorationType.DETERMINISTIC
+            ), torch.no_grad(), timeit("eval"):
+                eval_rollout = eval_env.rollout(
+                    eval_rollout_steps,
+                    exploration_policy,
+                    auto_cast_to_device=True,
+                    break_when_any_done=True,
+                )
+                eval_env.apply(dump_video)
+                eval_reward = eval_rollout["next", "reward"].sum(-2).mean().item()
+                metrics_to_log["eval/reward"] = eval_reward
+
+        if logger is not None:
+            metrics_to_log.update(timeit.todict(prefix="time"))
+            metrics_to_log["time/speed"] = pbar.format_dict["rate"]
+            log_metrics(logger, metrics_to_log, collected_frames)
+
+    collector.shutdown()
+    if not eval_env.is_closed:
+        eval_env.close()
+    if not train_env.is_closed:
+        train_env.close()
+
+
+if __name__ == "__main__":
+    main()

sota-implementations/ddpg/utils.py

@@ -0,0 +1,325 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+from __future__ import annotations
+
+import functools
+
+import torch
+
+from tensordict.nn import TensorDictModule, TensorDictSequential
+
+from torch import nn, optim
+from torchrl.collectors import SyncDataCollector
+from torchrl.data import TensorDictPrioritizedReplayBuffer, TensorDictReplayBuffer
+from torchrl.data.replay_buffers.storages import LazyMemmapStorage
+from torchrl.envs import (
+    CatTensors,
+    Compose,
+    DMControlEnv,
+    DoubleToFloat,
+    EnvCreator,
+    InitTracker,
+    ParallelEnv,
+    RewardSum,
+    StepCounter,
+    TransformedEnv,
+)
+from torchrl.envs.libs.gym import GymEnv, set_gym_backend
+from torchrl.envs.utils import ExplorationType, set_exploration_type
+from torchrl.modules import (
+    AdditiveGaussianModule,
+    MLP,
+    OrnsteinUhlenbeckProcessModule,
+    TanhModule,
+    ValueOperator,
+)
+
+from torchrl.objectives import SoftUpdate
+from torchrl.objectives.ddpg import DDPGLoss
+from torchrl.record import VideoRecorder
+
+
+# ====================================================================
+# Environment utils
+# -----------------
+
+
+def env_maker(cfg, device="cpu", from_pixels=False):
+    lib = cfg.env.library
+    if lib in ("gym", "gymnasium"):
+        with set_gym_backend(lib):
+            return GymEnv(
+                cfg.env.name, device=device, from_pixels=from_pixels, pixels_only=False
+            )
+    elif lib == "dm_control":
+        env = DMControlEnv(
+            cfg.env.name, cfg.env.task, from_pixels=from_pixels, pixels_only=False
+        )
+        return TransformedEnv(
+            env, CatTensors(in_keys=env.observation_spec.keys(), out_key="observation")
+        )
+    else:
+        raise NotImplementedError(f"Unknown lib {lib}.")
+
+
+def apply_env_transforms(env, max_episode_steps=1000):
+    transformed_env = TransformedEnv(
+        env,
+        Compose(
+            InitTracker(),
+            StepCounter(max_episode_steps),
+            DoubleToFloat(),
+            RewardSum(),
+        ),
+    )
+    return transformed_env
+
+
+def make_environment(cfg, logger):
+    """Make environments for training and evaluation."""
+    maker = functools.partial(env_maker, cfg, from_pixels=False)
+    parallel_env = ParallelEnv(
+        cfg.collector.env_per_collector,
+        EnvCreator(maker),
+        serial_for_single=True,
+    )
+    parallel_env.set_seed(cfg.env.seed)
+
+    train_env = apply_env_transforms(
+        parallel_env, max_episode_steps=cfg.env.max_episode_steps
+    )
+
+    maker = functools.partial(env_maker, cfg, from_pixels=cfg.logger.video)
+    eval_env = TransformedEnv(
+        ParallelEnv(
+            cfg.logger.num_eval_envs,
+            EnvCreator(maker),
+            serial_for_single=True,
+        ),
+        train_env.transform.clone(),
+    )
+    eval_env.set_seed(0)
+    if cfg.logger.video:
+        eval_env = eval_env.append_transform(
+            VideoRecorder(logger, tag="rendered", in_keys=["pixels"])
+        )
+    return train_env, eval_env
+
+
+# ====================================================================
+# Collector and replay buffer
+# ---------------------------
+
+
+def make_collector(
+    cfg,
+    train_env,
+    actor_model_explore,
+    compile=False,
+    compile_mode=None,
+    cudagraph=False,
+    device: torch.device | None = None,
+):
+    """Make collector."""
+    collector = SyncDataCollector(
+        train_env,
+        actor_model_explore,
+        frames_per_batch=cfg.collector.frames_per_batch,
+        init_random_frames=cfg.collector.init_random_frames,
+        reset_at_each_iter=cfg.collector.reset_at_each_iter,
+        total_frames=cfg.collector.total_frames,
+        device=device,
+        compile_policy={"mode": compile_mode, "fullgraph": True} if compile else False,
+        cudagraph_policy=cudagraph,
+    )
+    collector.set_seed(cfg.env.seed)
+    return collector
+
+
+def make_replay_buffer(
+    batch_size,
+    prb=False,
+    buffer_size=1000000,
+    scratch_dir=None,
+    device="cpu",
+    prefetch=3,
+):
+    if prb:
+        replay_buffer = TensorDictPrioritizedReplayBuffer(
+            alpha=0.7,
+            beta=0.5,
+            pin_memory=False,
+            prefetch=prefetch,
+            storage=LazyMemmapStorage(
+                buffer_size,
+                scratch_dir=scratch_dir,
+                device=device,
+            ),
+            batch_size=batch_size,
+        )
+    else:
+        replay_buffer = TensorDictReplayBuffer(
+            pin_memory=False,
+            prefetch=prefetch,
+            storage=LazyMemmapStorage(
+                buffer_size,
+                scratch_dir=scratch_dir,
+                device=device,
+            ),
+            batch_size=batch_size,
+        )
+    return replay_buffer
+
+
+# ====================================================================
+# Model
+# -----
+
+
+def make_ddpg_agent(cfg, train_env, eval_env, device):
+    """Make DDPG agent."""
+    # Define Actor Network
+    in_keys = ["observation"]
+    action_spec = train_env.action_spec_unbatched
+    actor_net_kwargs = {
+        "num_cells": cfg.network.hidden_sizes,
+        "out_features": action_spec.shape[-1],
+        "activation_class": get_activation(cfg),
+    }
+
+    actor_net = MLP(**actor_net_kwargs)
+
+    in_keys_actor = in_keys
+    actor_module = TensorDictModule(
+        actor_net,
+        in_keys=in_keys_actor,
+        out_keys=["param"],
+    )
+    actor = TensorDictSequential(
+        actor_module,
+        TanhModule(
+            in_keys=["param"],
+            out_keys=["action"],
+        ),
+    )
+
+    # Define Critic Network
+    qvalue_net_kwargs = {
+        "num_cells": cfg.network.hidden_sizes,
+        "out_features": 1,
+        "activation_class": get_activation(cfg),
+    }
+
+    qvalue_net = MLP(
+        **qvalue_net_kwargs,
+    )
+
+    qvalue = ValueOperator(
+        in_keys=["action"] + in_keys,
+        module=qvalue_net,
+    )
+
+    model = nn.ModuleList([actor, qvalue]).to(device)
+
+    # init nets
+    with torch.no_grad(), set_exploration_type(ExplorationType.RANDOM):
+        td = eval_env.reset()
+        td = td.to(device)
+        for net in model:
+            net(td)
+    del td
+    eval_env.close()
+
+    # Exploration wrappers:
+    if cfg.network.noise_type == "ou":
+        actor_model_explore = TensorDictSequential(
+            model[0],
+            OrnsteinUhlenbeckProcessModule(
+                spec=action_spec,
+                annealing_num_steps=1_000_000,
+                device=device,
+                safe=False,
+            ),
+        )
+    elif cfg.network.noise_type == "gaussian":
+        actor_model_explore = TensorDictSequential(
+            model[0],
+            AdditiveGaussianModule(
+                spec=action_spec,
+                sigma_end=1.0,
+                sigma_init=1.0,
+                mean=0.0,
+                std=0.1,
+                device=device,
+                safe=False,
+            ),
+        )
+    else:
+        raise NotImplementedError
+
+    return model, actor_model_explore
+
+
+# ====================================================================
+# DDPG Loss
+# ---------
+
+
+def make_loss_module(cfg, model):
+    """Make loss module and target network updater."""
+    # Create DDPG loss
+    loss_module = DDPGLoss(
+        actor_network=model[0],
+        value_network=model[1],
+        loss_function=cfg.optim.loss_function,
+        delay_actor=True,
+        delay_value=True,
+    )
+    loss_module.make_value_estimator(gamma=cfg.optim.gamma)
+
+    # Define Target Network Updater
+    target_net_updater = SoftUpdate(loss_module, eps=cfg.optim.target_update_polyak)
+    return loss_module, target_net_updater
+
+
+def make_optimizer(cfg, loss_module):
+    critic_params = list(loss_module.value_network_params.flatten_keys().values())
+    actor_params = list(loss_module.actor_network_params.flatten_keys().values())
+
+    optimizer_actor = optim.Adam(
+        actor_params, lr=cfg.optim.lr, weight_decay=cfg.optim.weight_decay
+    )
+    optimizer_critic = optim.Adam(
+        critic_params,
+        lr=cfg.optim.lr,
+        weight_decay=cfg.optim.weight_decay,
+    )
+    return optimizer_actor, optimizer_critic
+
+
+# ====================================================================
+# General utils
+# ---------
+
+
+def log_metrics(logger, metrics, step):
+    for metric_name, metric_value in metrics.items():
+        logger.log_scalar(metric_name, metric_value, step)
+
+
+def get_activation(cfg):
+    if cfg.network.activation == "relu":
+        return nn.ReLU
+    elif cfg.network.activation == "tanh":
+        return nn.Tanh
+    elif cfg.network.activation == "leaky_relu":
+        return nn.LeakyReLU
+    else:
+        raise NotImplementedError
+
+
+def dump_video(module):
+    if isinstance(module, VideoRecorder):
+        module.dump()

sota-implementations/decision_transformer/dt.py

@@ -0,0 +1,163 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+"""Decision Transformer Example.
+This is a self-contained example of an offline Decision Transformer training script.
+The helper functions are coded in the utils.py associated with this script.
+"""
+
+from __future__ import annotations
+
+import warnings
+
+import hydra
+import numpy as np
+import torch
+import tqdm
+from tensordict import TensorDict
+from tensordict.nn import CudaGraphModule
+from torchrl._utils import get_available_device, logger as torchrl_logger, timeit
+from torchrl.envs.libs.gym import set_gym_backend
+from torchrl.envs.utils import ExplorationType, set_exploration_type
+from torchrl.modules.tensordict_module import DecisionTransformerInferenceWrapper
+from torchrl.record import VideoRecorder
+from utils import (
+    dump_video,
+    log_metrics,
+    make_dt_loss,
+    make_dt_model,
+    make_dt_optimizer,
+    make_env,
+    make_logger,
+    make_offline_replay_buffer,
+)
+
+
+@hydra.main(config_path="", config_name="dt_config", version_base="1.1")
+def main(cfg: DictConfig):  # noqa: F821
+    set_gym_backend(cfg.env.backend).set()
+
+    model_device = (
+        torch.device(cfg.optim.device) if cfg.optim.device else get_available_device()
+    )
+
+    # Set seeds
+    torch.manual_seed(cfg.env.seed)
+    np.random.seed(cfg.env.seed)
+
+    # Create logger
+    logger = make_logger(cfg)
+
+    # Create offline replay buffer
+    offline_buffer, obs_loc, obs_std = make_offline_replay_buffer(
+        cfg.replay_buffer, cfg.env.reward_scaling
+    )
+
+    # Create test environment
+    test_env = make_env(
+        cfg.env, obs_loc, obs_std, from_pixels=cfg.logger.video, device=model_device
+    )
+    if cfg.logger.video:
+        test_env = test_env.append_transform(
+            VideoRecorder(logger, tag="rendered", in_keys=["pixels"])
+        )
+
+    # Create policy model
+    actor = make_dt_model(cfg, device=model_device)
+
+    # Create loss
+    loss_module = make_dt_loss(cfg.loss, actor, device=model_device)
+
+    # Create optimizer
+    transformer_optim, scheduler = make_dt_optimizer(
+        cfg.optim, loss_module, model_device
+    )
+
+    # Create inference policy
+    inference_policy = DecisionTransformerInferenceWrapper(
+        policy=actor,
+        inference_context=cfg.env.inference_context,
+        device=model_device,
+    )
+    inference_policy.set_tensor_keys(
+        observation="observation_cat",
+        action="action_cat",
+        return_to_go="return_to_go_cat",
+    )
+
+    pretrain_gradient_steps = cfg.optim.pretrain_gradient_steps
+    clip_grad = cfg.optim.clip_grad
+
+    def update(data: TensorDict) -> TensorDict:
+        transformer_optim.zero_grad(set_to_none=True)
+        # Compute loss
+        loss_vals = loss_module(data)
+        transformer_loss = loss_vals["loss"]
+
+        transformer_loss.backward()
+        torch.nn.utils.clip_grad_norm_(actor.parameters(), clip_grad)
+        transformer_optim.step()
+
+        return loss_vals
+
+    if cfg.compile.compile:
+        compile_mode = cfg.compile.compile_mode
+        if compile_mode in ("", None):
+            if cfg.compile.cudagraphs:
+                compile_mode = "default"
+            else:
+                compile_mode = "reduce-overhead"
+        update = torch.compile(update, mode=compile_mode, dynamic=True)
+    if cfg.compile.cudagraphs:
+        warnings.warn(
+            "CudaGraphModule is experimental and may lead to silently wrong results. Use with caution.",
+            category=UserWarning,
+        )
+        update = CudaGraphModule(update, warmup=50)
+
+    eval_steps = cfg.logger.eval_steps
+    pretrain_log_interval = cfg.logger.pretrain_log_interval
+    reward_scaling = cfg.env.reward_scaling
+
+    torchrl_logger.info(" ***Pretraining*** ")
+    # Pretraining
+    pbar = tqdm.tqdm(range(pretrain_gradient_steps))
+    for i in pbar:
+        timeit.printevery(1000, pretrain_gradient_steps, erase=True)
+        # Sample data
+        with timeit("rb - sample"):
+            data = offline_buffer.sample().to(model_device)
+        with timeit("update"):
+            loss_vals = update(data)
+        scheduler.step()
+        # Log metrics
+        metrics_to_log = {"train/loss": loss_vals["loss"]}
+
+        # Evaluation
+        with set_exploration_type(
+            ExplorationType.DETERMINISTIC
+        ), torch.no_grad(), timeit("eval"):
+            if i % pretrain_log_interval == 0:
+                eval_td = test_env.rollout(
+                    max_steps=eval_steps,
+                    policy=inference_policy,
+                    auto_cast_to_device=True,
+                )
+                test_env.apply(dump_video)
+            metrics_to_log["eval/reward"] = (
+                eval_td["next", "reward"].sum(1).mean().item() / reward_scaling
+            )
+
+        if logger is not None:
+            metrics_to_log.update(timeit.todict(prefix="time"))
+            metrics_to_log["time/speed"] = pbar.format_dict["rate"]
+            log_metrics(logger, metrics_to_log, i)
+
+    pbar.close()
+    if not test_env.is_closed:
+        test_env.close()
+
+
+if __name__ == "__main__":
+    main()