torchrl 0.11.0__cp314-cp314-manylinux_2_28_aarch64.whl

sota-implementations/discrete_sac/discrete_sac.py

@@ -0,0 +1,243 @@
+# 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.
+"""Discrete SAC Example.
+
+This is a simple self-contained example of a discrete SAC training script.
+
+It supports gym state environments like CartPole.
+
+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.nn import CudaGraphModule
+from torchrl._utils import 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_environment,
+    make_loss_module,
+    make_optimizer,
+    make_replay_buffer,
+    make_sac_agent,
+)
+
+
+@hydra.main(version_base="1.1", config_path="", config_name="config")
+def main(cfg: DictConfig):  # noqa: F821
+    device = cfg.network.device
+    if device in ("", None):
+        if torch.cuda.is_available():
+            device = "cuda:0"
+        else:
+            device = "cpu"
+    device = torch.device(device)
+
+    # Create logger
+    exp_name = generate_exp_name("DiscreteSAC", cfg.logger.exp_name)
+    logger = None
+    if cfg.logger.backend:
+        logger = get_logger(
+            logger_type=cfg.logger.backend,
+            logger_name="DiscreteSAC_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 = make_sac_agent(cfg, train_env, eval_env, device)
+
+    # Create TD3 loss
+    loss_module, target_net_updater = make_loss_module(cfg, model)
+
+    # 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, optimizer_alpha = make_optimizer(
+        cfg, loss_module
+    )
+    optimizer = group_optimizers(optimizer_actor, optimizer_critic, optimizer_alpha)
+    del optimizer_actor, optimizer_critic, optimizer_alpha
+
+    def update(sampled_tensordict):
+        optimizer.zero_grad(set_to_none=True)
+
+        # Compute loss
+        loss_out = loss_module(sampled_tensordict)
+
+        actor_loss, q_loss, alpha_loss = (
+            loss_out["loss_actor"],
+            loss_out["loss_qvalue"],
+            loss_out["loss_alpha"],
+        )
+
+        # Update critic
+        (q_loss + actor_loss + alpha_loss).backward()
+        optimizer.step()
+
+        # Update target params
+        target_net_updater.step()
+
+        return loss_out.detach()
+
+    compile_mode = None
+    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)
+    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)
+
+    # Create off-policy collector
+    collector = make_collector(
+        cfg,
+        train_env,
+        model[0],
+        compile=compile_mode is not None,
+        compile_mode=compile_mode,
+        cudagraphs=cfg.compile.cudagraphs,
+    )
+
+    # 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
+    eval_rollout_steps = cfg.env.max_episode_steps
+    eval_iter = cfg.logger.eval_iter
+    frames_per_batch = cfg.collector.frames_per_batch
+
+    c_iter = iter(collector)
+    total_iter = len(collector)
+    for i in range(total_iter):
+        timeit.printevery(1000, total_iter, erase=True)
+        with timeit("collecting"):
+            collected_data = next(c_iter)
+
+        # Update weights of the inference policy
+        collector.update_policy_weights_()
+        current_frames = collected_data.numel()
+
+        pbar.update(current_frames)
+
+        collected_data = collected_data.reshape(-1)
+        with timeit("rb - extend"):
+            # Add to replay buffer
+            replay_buffer.extend(collected_data)
+        collected_frames += current_frames
+
+        # Optimization steps
+        if collected_frames >= init_random_frames:
+            tds = []
+            for _ in range(num_updates):
+                with timeit("rb - sample"):
+                    # Sample from replay buffer
+                    sampled_tensordict = replay_buffer.sample()
+
+                with timeit("update"):
+                    torch.compiler.cudagraph_mark_step_begin()
+                    sampled_tensordict = sampled_tensordict.to(device)
+                    loss_out = update(sampled_tensordict).clone()
+
+                tds.append(loss_out)
+
+                # Update priority
+                if prb:
+                    replay_buffer.update_priority(sampled_tensordict)
+            tds = torch.stack(tds).mean()
+
+        # Logging
+        episode_end = (
+            collected_data["next", "done"]
+            if collected_data["next", "done"].any()
+            else collected_data["next", "truncated"]
+        )
+        episode_rewards = collected_data["next", "episode_reward"][episode_end]
+
+        metrics_to_log = {}
+        if len(episode_rewards) > 0:
+            episode_length = collected_data["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:
+            metrics_to_log["train/q_loss"] = tds["loss_qvalue"]
+            metrics_to_log["train/a_loss"] = tds["loss_actor"]
+            metrics_to_log["train/alpha_loss"] = tds["loss_alpha"]
+
+        # Evaluation
+        prev_test_frame = ((i - 1) * frames_per_batch) // eval_iter
+        cur_test_frame = (i * frames_per_batch) // eval_iter
+        final = current_frames >= collector.total_frames
+        if (i >= 1 and (prev_test_frame < cur_test_frame)) or final:
+            with set_exploration_type(
+                ExplorationType.DETERMINISTIC
+            ), torch.no_grad(), timeit("eval"):
+                eval_rollout = eval_env.rollout(
+                    eval_rollout_steps,
+                    model[0],
+                    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/discrete_sac/utils.py

@@ -0,0 +1,324 @@
+# 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 tempfile
+from contextlib import nullcontext
+
+import torch
+from tensordict.nn import InteractionType, TensorDictModule
+
+from torch import nn, optim
+from torchrl.collectors import SyncDataCollector
+from torchrl.data import (
+    Composite,
+    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 MLP, SafeModule
+from torchrl.modules.distributions import OneHotCategorical
+
+from torchrl.modules.tensordict_module.actors import ProbabilisticActor
+from torchrl.objectives import SoftUpdate
+from torchrl.objectives.sac import DiscreteSACLoss
+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):
+    transformed_env = TransformedEnv(
+        env,
+        Compose(
+            StepCounter(max_steps=max_episode_steps),
+            InitTracker(),
+            DoubleToFloat(),
+            RewardSum(),
+        ),
+    )
+    return transformed_env
+
+
+def make_environment(cfg, logger=None):
+    """Make environments for training and evaluation."""
+    maker = functools.partial(env_maker, cfg)
+    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.collector.env_per_collector,
+            EnvCreator(maker),
+            serial_for_single=True,
+        ),
+        train_env.transform.clone(),
+    )
+    if cfg.logger.video:
+        eval_env = eval_env.insert_transform(
+            0, 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,
+    cudagraphs=False,
+):
+    """Make collector."""
+    device = cfg.collector.device
+    if device in ("", None):
+        if torch.cuda.is_available():
+            device = "cuda:0"
+        else:
+            device = "cpu"
+    device = torch.device(device)
+    collector = SyncDataCollector(
+        train_env,
+        actor_model_explore,
+        init_random_frames=cfg.collector.init_random_frames,
+        frames_per_batch=cfg.collector.frames_per_batch,
+        total_frames=cfg.collector.total_frames,
+        reset_at_each_iter=cfg.collector.reset_at_each_iter,
+        device=device,
+        storing_device="cpu",
+        compile_policy=False if not compile else {"mode": compile_mode},
+        cudagraph_policy={"warmup": 10} if cudagraphs else False,
+    )
+    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,
+):
+    with (
+        tempfile.TemporaryDirectory()
+        if scratch_dir is None
+        else nullcontext(scratch_dir)
+    ) as scratch_dir:
+        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_sac_agent(cfg, train_env, eval_env, device):
+    """Make discrete SAC agent."""
+    # Define Actor Network
+    in_keys = ["observation"]
+    action_spec = train_env.action_spec
+    if train_env.batch_size:
+        action_spec = action_spec[(0,) * len(train_env.batch_size)]
+    # Define Actor Network
+    in_keys = ["observation"]
+
+    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)
+
+    actor_module = SafeModule(
+        module=actor_net,
+        in_keys=in_keys,
+        out_keys=["logits"],
+    )
+    actor = ProbabilisticActor(
+        spec=Composite(action=eval_env.action_spec),
+        module=actor_module,
+        in_keys=["logits"],
+        out_keys=["action"],
+        distribution_class=OneHotCategorical,
+        distribution_kwargs={},
+        default_interaction_type=InteractionType.RANDOM,
+        return_log_prob=False,
+    )
+
+    # Define Critic Network
+    qvalue_net_kwargs = {
+        "num_cells": cfg.network.hidden_sizes,
+        "out_features": action_spec.shape[-1],
+        "activation_class": get_activation(cfg),
+    }
+    qvalue_net = MLP(
+        **qvalue_net_kwargs,
+    )
+
+    qvalue = TensorDictModule(
+        in_keys=in_keys,
+        out_keys=["action_value"],
+        module=qvalue_net,
+    )
+
+    model = torch.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()
+
+    return model
+
+
+# ====================================================================
+# Discrete SAC Loss
+# ---------
+
+
+def make_loss_module(cfg, model):
+    """Make loss module and target network updater."""
+    # Create discrete SAC loss
+    loss_module = DiscreteSACLoss(
+        actor_network=model[0],
+        qvalue_network=model[1],
+        num_actions=model[0].spec["action"].space.n,
+        num_qvalue_nets=2,
+        loss_function=cfg.optim.loss_function,
+        target_entropy_weight=cfg.optim.target_entropy_weight,
+        delay_qvalue=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.qvalue_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,
+    )
+    optimizer_alpha = optim.Adam(
+        [loss_module.log_alpha],
+        lr=3.0e-4,
+    )
+    return optimizer_actor, optimizer_critic, optimizer_alpha
+
+
+# ====================================================================
+# 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/dqn/README.md

@@ -0,0 +1,30 @@
+## Reproducing Deep Q-Learning (DQN) Algorithm Results
+
+This repository contains scripts that enable training agents using the Deep Q-Learning (DQN) Algorithm on CartPole and Atari environments. For Atari, We follow the original paper [Playing Atari with Deep Reinforcement Learning](https://arxiv.org/abs/1312.5602) by Mnih et al. (2013).
+
+
+## Examples Structure
+
+Please note that each example is independent of each other for the sake of simplicity. Each example contains the following files:
+
+1. **Main Script:** The definition of algorithm components and the training loop can be found in the main script  (e.g. dqn_atari.py).
+
+2. **Utils File:** A utility file is provided to contain various helper functions, generally to create the environment and the models (e.g. utils_atari.py).
+
+3. **Configuration File:** This file includes default hyperparameters specified in the original paper. Users can modify these hyperparameters to customize their experiments  (e.g. config_atari.yaml).
+
+
+## Running the Examples
+
+You can execute the DQN algorithm on the CartPole environment by running the following command:
+
+```bash
+python dqn_cartpole.py
+
+``` 
+
+You can execute the DQN algorithm on Atari environments by running the following command:
+
+```bash
+python dqn_atari.py
+```