Implementation:Facebookresearch Habitat lab BaseILTrainer

Knowledge Sources	Facebookresearch_Habitat_lab
Domains	Embodied_AI, Imitation_Learning
Last Updated	2026-02-15 00:00 GMT

Overview

BaseILTrainer is an abstract base class for imitation learning (IL) trainers that provides common directory management, checkpoint saving, and evaluation scaffolding for IL-based training algorithms in Habitat.

Description

BaseILTrainer extends BaseTrainer to provide a foundation for all imitation learning trainers. Upon initialization it creates directories for logging, checkpoints, and evaluation results based on the provided configuration. It exposes a configurable flush_secs property for controlling TensorBoard flush intervals and defines abstract methods train, _eval_checkpoint, and load_checkpoint that subclasses must implement. The save_checkpoint method serializes a model state dictionary to the configured checkpoint folder using PyTorch's save mechanism.

Usage

Subclass BaseILTrainer when building a new imitation learning training algorithm. Override the train, _eval_checkpoint, and load_checkpoint methods with the specific training loop, evaluation logic, and checkpoint loading behavior for your IL approach.

Code Reference

Source Location

• Repository: Facebookresearch_Habitat_lab
• File: habitat-baselines/habitat_baselines/common/base_il_trainer.py
• Lines: 20-110

Signature

class BaseILTrainer(BaseTrainer):
    device: torch.device
    config: "DictConfig"
    video_option: List[str]
    _flush_secs: int

    def __init__(self, config: "DictConfig"):

Import

from habitat_baselines.common.base_il_trainer import BaseILTrainer

I/O Contract

Inputs

Name	Type	Required	Description
config	DictConfig	Yes	Hydra/OmegaConf configuration object containing habitat_baselines IL settings

Outputs

Name	Type	Description
(instance)	BaseILTrainer	Initialized trainer with created directories for logs, checkpoints, and results

Key Methods

save_checkpoint

def save_checkpoint(self, state_dict: OrderedDict, file_name: str) -> None

Saves a model state dictionary to the configured checkpoint folder.

_eval_checkpoint (abstract)

def _eval_checkpoint(
    self,
    checkpoint_path: str,
    writer: TensorboardWriter,
    checkpoint_index: int = 0,
) -> None

Evaluates a single checkpoint. Must be implemented by subclasses.

train (abstract)

def train(self) -> None

Runs the training loop. Must be implemented by subclasses.

load_checkpoint (abstract)

def load_checkpoint(self, checkpoint_path, *args, **kwargs) -> Dict

Loads a checkpoint from disk. Must be implemented by subclasses.

Usage Examples

Basic Usage

from collections import OrderedDict
from habitat_baselines.common.base_il_trainer import BaseILTrainer
from habitat_baselines.common.tensorboard_utils import TensorboardWriter

class MyILTrainer(BaseILTrainer):
    def train(self) -> None:
        # Custom IL training loop
        for epoch in range(100):
            loss = self._run_epoch()
            state_dict = self.model.state_dict()
            self.save_checkpoint(
                OrderedDict(state_dict),
                f"ckpt_{epoch}.pth"
            )

    def _eval_checkpoint(self, checkpoint_path, writer, checkpoint_index=0):
        ckpt = self.load_checkpoint(checkpoint_path)
        # Run evaluation with loaded checkpoint
        pass

    def load_checkpoint(self, checkpoint_path, *args, **kwargs):
        return torch.load(checkpoint_path)

# Instantiate and train
trainer = MyILTrainer(config=my_config)
trainer.train()