expected to be in range of [-1, 0], but got 1

[WillInvest]

import argparse
import datetime
import os
import sys
import pprint
import numpy as np
import torch
# Add the parent directory to the system path
sys.path.append('..')

from tianshou.data import Collector, ReplayBuffer, VectorReplayBuffer, PrioritizedVectorReplayBuffer, Batch
from tianshou.env.venvs import DummyVectorEnv, SubprocVectorEnv
from tianshou.exploration import GaussianNoise
from tianshou.policy import DDPGPolicy
from tianshou.policy.base import BasePolicy
from tianshou.trainer import OffpolicyTrainer
from tianshou.utils.net.common import Net
from tianshou.utils.net.continuous import Actor, Critic
from tianshou.highlevel.logger import LoggerFactoryDefault

from env.amm_env import ArbitrageEnv
from env.market import GBMPriceSimulator
from env.new_amm import AMM

def get_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser()
    parser.add_argument("--task", type=str, default="AMM")
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument("--alpha", type=float, default=0.5)
    parser.add_argument("--beta", type=float, default=0.4)
    parser.add_argument("--buffer-size", type=int, default=1e6)
    parser.add_argument("--hidden-sizes", type=int, nargs="*", default=[256, 256])
    parser.add_argument("--actor-lr", type=float, default=1e-5)
    parser.add_argument("--critic-lr", type=float, default=1e-5)
    parser.add_argument("--gamma", type=float, default=0.0)
    parser.add_argument("--tau", type=float, default=0.0005)
    parser.add_argument("--exploration-noise", type=float, default=0.01)
    parser.add_argument("--start-timesteps", type=int, default=1)
    parser.add_argument("--epoch", type=int, default=200)
    parser.add_argument("--step-per-epoch", type=int, default=5000)
    parser.add_argument("--step-per-collect", type=int, default=10)
    parser.add_argument("--update-per-step", type=int, default=1)
    parser.add_argument("--n-step", type=int, default=3)
    parser.add_argument("--batch-size", type=int, default=64)
    parser.add_argument("--training-num", type=int, default=10)
    parser.add_argument("--test-num", type=int, default=10)
    parser.add_argument("--logdir", type=str, default="log")
    parser.add_argument("--render", type=float, default=0.0)
    parser.add_argument(
        "--device",
        type=str,
        default="cuda" if torch.cuda.is_available() else "cpu",
    )
    parser.add_argument("--resume-path", type=str, default=None)
    parser.add_argument("--resume-id", type=str, default=None)
    parser.add_argument(
        "--logger",
        type=str,
        default="tensorboard",
        choices=["tensorboard", "wandb"],
    )
    parser.add_argument("--wandb-project", type=str, default="mujoco.benchmark")
    parser.add_argument(
        "--watch",
        default=False,
        action="store_true",
        help="watch the play of pre-trained policy only",
    )
    parser.add_argument("--USING_USD", type=bool, default=True)
    parser.add_argument("--mkt_start", type=float, default=1.0)
    parser.add_argument("--fee_rate", type=float, default=0.02)

    return parser.parse_args()

def test_ddpg(args: argparse.Namespace = get_args()) -> None:
    market = GBMPriceSimulator(start_price=args.mkt_start, deterministic=False)
    fee_rate = args.fee_rate
    amm = AMM(initial_a=10000, initial_b=10000, fee=fee_rate)
    env = ArbitrageEnv(market, amm, USD=args.USING_USD)
    eval_market = GBMPriceSimulator(start_price=args.mkt_start, deterministic=True)
    # test_env = ArbitrageEnv(eval_market, amm, USD=args.USING_USD)
    train_env = SubprocVectorEnv([lambda: ArbitrageEnv(market, amm, USD=args.USING_USD) for _ in range(args.training_num)])
    test_env = SubprocVectorEnv([lambda: ArbitrageEnv(eval_market, amm, USD=args.USING_USD) for _ in range(args.test_num)])

    args.state_shape = env.observation_space.shape or env.observation_space.n
    args.action_shape = env.action_space.shape or env.action_space.n
    args.max_action = env.action_space.high[0]
    args.exploration_noise = args.exploration_noise * args.max_action
    args.USING_USD = False
    print("Observations shape:", args.state_shape)
    print("Actions shape:", args.action_shape)
    print("Action range:", np.min(env.action_space.low), np.max(env.action_space.high))
    print(f"max_action: {args.max_action}")

    # seed
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    # model
    net_a = Net(state_shape=args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device)
    actor = Actor(net_a, args.action_shape, max_action=args.max_action, device=args.device).to(
        args.device,
    )
    actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
    net_c = Net(
        state_shape=args.state_shape,
        action_shape=args.action_shape,
        hidden_sizes=args.hidden_sizes,
        concat=True,
        device=args.device,
    )
    critic = Critic(net_c, device=args.device).to(args.device)
    critic_optim = torch.optim.Adam(critic.parameters(), lr=args.critic_lr)
    policy: DDPGPolicy = DDPGPolicy(
        actor=actor,
        actor_optim=actor_optim,
        critic=critic,
        critic_optim=critic_optim,
        tau=args.tau,
        gamma=args.gamma,
        exploration_noise=GaussianNoise(sigma=args.exploration_noise),
        estimation_step=args.n_step,
        action_space=env.action_space,
    )
    # load a previous policy
    if args.resume_path:
        policy.load_state_dict(torch.load(args.resume_path, map_location=args.device))
        print("Loaded agent from: ", args.resume_path)

    # collector
    buffer = PrioritizedVectorReplayBuffer(
        args.buffer_size,
        buffer_num=len(train_env),
        ignore_obs_next=True,
        save_only_last_obs=True,
        alpha=args.alpha,
        beta=args.beta,
    )
    train_collector = Collector(policy, train_env, buffer, exploration_noise=True)
    test_collector = Collector(policy, test_env, exploration_noise=True)
    train_collector.reset()
    train_collector.collect(n_step=args.start_timesteps, random=True)

    # log
    now = datetime.datetime.now().strftime("%y%m%d-%H%M%S")
    args.algo_name = "ddpg"
    log_name = os.path.join(args.task, args.algo_name, str(args.seed), now)
    log_path = os.path.join(args.logdir, log_name)

    # logger
    logger_factory = LoggerFactoryDefault()
    if args.logger == "wandb":
        logger_factory.logger_type = "wandb"
        logger_factory.wandb_project = args.wandb_project
    else:
        logger_factory.logger_type = "tensorboard"

    logger = logger_factory.create_logger(
        log_dir=log_path,
        experiment_name=log_name,
        run_id=args.resume_id,
        config_dict=vars(args),
    )

    def save_best_fn(policy: BasePolicy) -> None:
        torch.save(policy.state_dict(), os.path.join(log_path, "policy.pth"))

    def save_dist_fn(policy: BasePolicy) -> None:
        torch.save(policy.state_dict(), os.path.join(log_path, "best_dist_policy.pth"))

    if not args.watch:
        # trainer
        result = OffpolicyTrainer(
            policy=policy,
            train_collector=train_collector,
            test_collector=test_collector,
            max_epoch=args.epoch,
            step_per_epoch=args.step_per_epoch,
            step_per_collect=args.step_per_collect,
            episode_per_test=args.test_num,
            batch_size=args.batch_size,
            save_best_fn=save_best_fn,
            save_dist_fn=save_dist_fn,
            logger=logger,
            update_per_step=args.update_per_step,
            test_in_train=False,
            verbose=True,
        ).run()
        pprint.pprint(result)

#     # Let's watch its performance!
    test_env.seed(args.seed)
    test_collector.reset()
    collector_stats = test_collector.collect(n_episode=args.test_num, render=args.render)
    print(collector_stats)

if __name__ == "__main__":
    test_ddpg()

Result:

Observations shape: (2,)
Actions shape: (1,)
Action range: -0.99999 0.99999
max_action: 0.9999899864196777
/home/shiftpub/miniconda/envs/amm_tianshou/lib/python3.11/site-packages/tianshou/policy/modelfree/ddpg.py:93: UserWarning: action_scaling and action_bound_method are only intended to dealwith unbounded model action space, but find actor model boundaction space with max_action=0.9999899864196777.Consider using unbounded=True option of the actor model,or set action_scaling to False and action_bound_method to None.
  warnings.warn(
/home/shiftpub/miniconda/envs/amm_tianshou/lib/python3.11/site-packages/tianshou/data/collector.py:331: UserWarning: n_step=1 is not a multiple of (self.env_num=10), which may cause extra transitions being collected into the buffer.
...
obs: torch.Size([6, 2])
obs: torch.Size([6, 256])
Epoch #1:   0%|                                                                                                                                                              | 0/5000 [00:00<?, ?it/s]obs: torch.Size([10, 2])
obs: torch.Size([10, 256])
obs: torch.Size([64])
Epoch #1:   0%|       
...
  File "/home/shiftpub/miniconda/envs/amm_tianshou/lib/python3.11/site-packages/tianshou/utils/net/common.py", line 143, in forward
    obs = obs.flatten(1)
          ^^^^^^^^^^^^^^
IndexError: Dimension out of range (expected to be in range of [-1, 0], but got 1)

[WillInvest]

I noticed that if I change batch size to 32, then the weird "obs: torch.Size([64])" becomes "obs: torch.Size([32])"

so the error is somehow connected to the batch size

[WillInvest]

buffer = PrioritizedVectorReplayBuffer(
    args.buffer_size,
    buffer_num=len(train_env),
    # ignore_obs_next=True,
    # save_only_last_obs=True,
    alpha=args.alpha,
    beta=args.beta,
)

when I comment out those two lines, problem solved.

Can someone help me explain what happens here? Really appreciate.

[dantp-ai]

@WillInvest I can look into it, but can you format your posts above to make the code more readable?

[MischaPanch]

buffer = PrioritizedVectorReplayBuffer(
    args.buffer_size,
    buffer_num=len(train_env),
    # ignore_obs_next=True,
    # save_only_last_obs=True,
    alpha=args.alpha,
    beta=args.beta,
)

when I comment out those two lines, problem solved.

Can someone help me explain what happens here? Really appreciate.

The buffer is currently very complicated, I was working on simplifying it and extending documentation in the last months. Will be merged to master soon.

ignore_obs_next should essentially never be set to True, it means that obs rolled forwards will be used instead of obs_next, which means making errors at the boundaries of episodes and at interrupted collections. It's only an option if one really really needs to save RAM, like maybe with atari envs, where observations are large.

save_only_last_obs should also not be touched.