No module named 'libero' when trying to run main.py

[sanikanandpure]

Hi,

I'm trying to run the default script in main.py using LIBERO.

import os

os.environ["TOKENIZERS_PARALLELISM"] = "false"
import sys
import json
import multiprocessing
import pprint
import time
from pathlib import Path

import hydra
import numpy as np
import wandb
import yaml
import torch
from easydict import EasyDict
from hydra.utils import to_absolute_path
from omegaconf import OmegaConf

from libero.libero import get_libero_path
from libero.libero.benchmark import get_benchmark
from libero.lifelong.algos import get_algo_class, get_algo_list
from libero.lifelong.models import get_policy_list
from libero.lifelong.datasets import GroupedTaskDataset, SequenceVLDataset, get_dataset
from libero.lifelong.metric import evaluate_loss, evaluate_success
from libero.lifelong.utils import (
    NpEncoder,
    compute_flops,
    control_seed,
    safe_device,
    torch_load_model,
    create_experiment_dir,
    get_task_embs,
)

@hydra.main(config_path="../configs", config_name="config", version_base=None)
def main(hydra_cfg):
    # preprocessing
    yaml_config = OmegaConf.to_yaml(hydra_cfg)
    cfg = EasyDict(yaml.safe_load(yaml_config))

    # print configs to terminal
    pp = pprint.PrettyPrinter(indent=2)
    pp.pprint(cfg)

    pp.pprint("Available algorithms:")
    pp.pprint(get_algo_list())

    pp.pprint("Available policies:")
    pp.pprint(get_policy_list())

    # control seed
    control_seed(cfg.seed)

    # prepare lifelong learning
    cfg.folder = cfg.folder or get_libero_path("datasets")
    cfg.bddl_folder = cfg.bddl_folder or get_libero_path("bddl_files")
    cfg.init_states_folder = cfg.init_states_folder or get_libero_path("init_states")

    benchmark = get_benchmark(cfg.benchmark_name)(cfg.data.task_order_index)
    n_manip_tasks = benchmark.n_tasks

    # prepare datasets from the benchmark
    manip_datasets = []
    descriptions = []
    shape_meta = None

    for i in range(n_manip_tasks):
        # currently we assume tasks from same benchmark have the same shape_meta
        try:
            task_i_dataset, shape_meta = get_dataset(
                dataset_path=os.path.join(
                    cfg.folder, benchmark.get_task_demonstration(i)
                ),
                obs_modality=cfg.data.obs.modality,
                initialize_obs_utils=(i == 0),
                seq_len=cfg.data.seq_len,
            )
        except Exception as e:
            print(
                f"[error] failed to load task {i} name {benchmark.get_task_names()[i]}"
            )
            print(f"[error] {e}")
        print(os.path.join(cfg.folder, benchmark.get_task_demonstration(i)))
        # add language to the vision dataset, hence we call vl_dataset
        task_description = benchmark.get_task(i).language
        descriptions.append(task_description)
        manip_datasets.append(task_i_dataset)

    task_embs = get_task_embs(cfg, descriptions)
    benchmark.set_task_embs(task_embs)

    gsz = cfg.data.task_group_size
    if gsz == 1:  # each manipulation task is its own lifelong learning task
        datasets = [
            SequenceVLDataset(ds, emb) for (ds, emb) in zip(manip_datasets, task_embs)
        ]
        n_demos = [data.n_demos for data in datasets]
        n_sequences = [data.total_num_sequences for data in datasets]
    else:  # group gsz manipulation tasks into a lifelong task, currently not used
        assert (
            n_manip_tasks % gsz == 0
        ), f"[error] task_group_size does not divide n_tasks"
        datasets = []
        n_demos = []
        n_sequences = []
        for i in range(0, n_manip_tasks, gsz):
            dataset = GroupedTaskDataset(
                manip_datasets[i : i + gsz], task_embs[i : i + gsz]
            )
            datasets.append(dataset)
            n_demos.extend([x.n_demos for x in dataset.sequence_datasets])
            n_sequences.extend(
                [x.total_num_sequences for x in dataset.sequence_datasets]
            )

    n_tasks = n_manip_tasks // gsz  # number of lifelong learning tasks
    print("\n=================== Lifelong Benchmark Information  ===================")
    print(f" Name: {benchmark.name}")
    print(f" # Tasks: {n_manip_tasks // gsz}")
    for i in range(n_tasks):
        print(f"    - Task {i+1}:")
        for j in range(gsz):
            print(f"        {benchmark.get_task(i*gsz+j).language}")
    print(" # demonstrations: " + " ".join(f"({x})" for x in n_demos))
    print(" # sequences: " + " ".join(f"({x})" for x in n_sequences))
    print("=======================================================================\n")

    # prepare experiment and update the config
    create_experiment_dir(cfg)
    cfg.shape_meta = shape_meta

    if cfg.use_wandb:
        wandb.init(project="libero", config=cfg)
        wandb.run.name = cfg.experiment_name

    result_summary = {
        "L_conf_mat": np.zeros((n_manip_tasks, n_manip_tasks)),  # loss confusion matrix
        "S_conf_mat": np.zeros((n_manip_tasks, n_manip_tasks)),  # success confusion matrix
        "L_fwd": np.zeros((n_manip_tasks,)),  # loss AUC, how fast the agent learns
        "S_fwd": np.zeros((n_manip_tasks,)),  # success AUC, how fast the agent succeeds
    }

    if cfg.eval.save_sim_states:
        # for saving the evaluate simulation states, so we can replay them later
        for k in range(n_manip_tasks):
            for p in range(k + 1):  # for testing task p when the agent learns to task k
                result_summary[f"k{k}_p{p}"] = [[] for _ in range(cfg.eval.n_eval)]
            for e in range(
                cfg.train.n_epochs + 1
            ):  # for testing task k at the e-th epoch when the agent learns on task k
                if e % cfg.eval.eval_every == 0:
                    result_summary[f"k{k}_e{e//cfg.eval.eval_every}"] = [
                        [] for _ in range(cfg.eval.n_eval)
                    ]

    # define lifelong algorithm
    algo = safe_device(get_algo_class(cfg.lifelong.algo)(n_tasks, cfg), cfg.device)
    if cfg.pretrain_model_path != "":  # load a pretrained model if there is any
        try:
            algo.policy.load_state_dict(torch_load_model(cfg.pretrain_model_path)[0])
        except:
            print(
                f"[error] cannot load pretrained model from {cfg.pretrain_model_path}"
            )
            sys.exit(0)

    print(f"[info] start lifelong learning with algo {cfg.lifelong.algo}")
    GFLOPs, MParams = compute_flops(algo, datasets[0], cfg)
    print(f"[info] policy has {GFLOPs:.1f} GFLOPs and {MParams:.1f} MParams\n")

    # save the experiment config file, so we can resume or replay later
    with open(os.path.join(cfg.experiment_dir, "config.json"), "w") as f:
        json.dump(cfg, f, cls=NpEncoder, indent=4)

    if cfg.lifelong.algo == "Multitask":

        algo.train()
        s_fwd, l_fwd = algo.learn_all_tasks(datasets, benchmark, result_summary)
        result_summary["L_fwd"][-1] = l_fwd
        result_summary["S_fwd"][-1] = s_fwd

        # evalute on all seen tasks at the end if eval.eval is true
        if cfg.eval.eval:
            L = evaluate_loss(cfg, algo, benchmark, datasets)
            S = evaluate_success(
                cfg=cfg,
                algo=algo,
                benchmark=benchmark,
                task_ids=list(range(n_manip_tasks)),
                result_summary=result_summary if cfg.eval.save_sim_states else None,
            )

            result_summary["L_conf_mat"][-1] = L
            result_summary["S_conf_mat"][-1] = S

            if cfg.use_wandb:
                wandb.run.summary["success_confusion_matrix"] = result_summary[
                    "S_conf_mat"
                ]
                wandb.run.summary["loss_confusion_matrix"] = result_summary[
                    "L_conf_mat"
                ]
                wandb.run.summary["fwd_transfer_success"] = result_summary["S_fwd"]
                wandb.run.summary["fwd_transfer_loss"] = result_summary["L_fwd"]
                wandb.run.summary.update()

            print(("[All task loss ] " + " %4.2f |" * n_tasks) % tuple(L))
            print(("[All task succ.] " + " %4.2f |" * n_tasks) % tuple(S))

            torch.save(result_summary, os.path.join(cfg.experiment_dir, f"result.pt"))
    else:
        for i in range(n_tasks):
            print(f"[info] start training on task {i}")
            algo.train()

            t0 = time.time()
            s_fwd, l_fwd = algo.learn_one_task(
                datasets[i], i, benchmark, result_summary
            )
            result_summary["S_fwd"][i] = s_fwd
            result_summary["L_fwd"][i] = l_fwd
            t1 = time.time()

            # evalute on all seen tasks at the end of learning each task
            if cfg.eval.eval:
                L = evaluate_loss(cfg, algo, benchmark, datasets[: i + 1])
                t2 = time.time()
                S = evaluate_success(
                    cfg=cfg,
                    algo=algo,
                    benchmark=benchmark,
                    task_ids=list(range((i + 1) * gsz)),
                    result_summary=result_summary if cfg.eval.save_sim_states else None,
                )
                t3 = time.time()
                result_summary["L_conf_mat"][i][: i + 1] = L
                result_summary["S_conf_mat"][i][: i + 1] = S

                if cfg.use_wandb:
                    wandb.run.summary["success_confusion_matrix"] = result_summary[
                        "S_conf_mat"
                    ]
                    wandb.run.summary["loss_confusion_matrix"] = result_summary[
                        "L_conf_mat"
                    ]
                    wandb.run.summary["fwd_transfer_success"] = result_summary["S_fwd"]
                    wandb.run.summary["fwd_transfer_loss"] = result_summary["L_fwd"]
                    wandb.run.summary.update()

                print(
                    f"[info] train time (min) {(t1-t0)/60:.1f} "
                    + f"eval loss time {(t2-t1)/60:.1f} "
                    + f"eval success time {(t3-t2)/60:.1f}"
                )
                print(("[Task %2d loss ] " + " %4.2f |" * (i + 1)) % (i, *L))
                print(("[Task %2d succ.] " + " %4.2f |" * (i + 1)) % (i, *S))
                torch.save(
                    result_summary, os.path.join(cfg.experiment_dir, f"result.pt")
                )

    print("[info] finished learning\n")
    if cfg.use_wandb:
        wandb.finish()

if __name__ == "__main__":
    # Set the multiprocessing start method to 'spawn'
    if multiprocessing.get_start_method(allow_none=True) != "spawn":  
        multiprocessing.set_start_method("spawn", force=True)
    main()
`

The above is the script I'm trying to run, and I am getting the following error on line 22 (from libero.libero import get_libero_path); ModuleNotFoundError: No module named 'libero'

I have tried running pip install robosuite followed by pip install -e . to no avail. I also tried uninstalling/reinstalling. Could not find any solutions in the documentation, so was wondering if I could get some input.

[sanikanandpure]

Resolved this; the imports were not referring to the correct directory.