navsim/agents/dreamer/hydra_dreamer_planning_agent.py

import os
import pickle
from typing import Any, Union

import numpy as np
from pytorch_lightning.callbacks import ModelCheckpoint
from torch.optim import Optimizer
from torch.optim.lr_scheduler import LRScheduler

from navsim.agents.dreamer.hydra_dreamer_config import HydraDreamerConfig
from navsim.agents.dreamer.hydra_dreamer_planning_model import HydraDreamerPlanningModel
from navsim.agents.dreamer.hydra_dreamer_wm_features import HydraDreamerWmFeatureBuilder
from navsim.agents.hydra.hydra_features import HydraFeatureBuilder, HydraTargetBuilder
from navsim.agents.hydra.hydra_loss_fn import hydra_kd_imi_agent_loss
from navsim.common.dataclasses import SensorConfig
from navsim.planning.training.abstract_feature_target_builder import (
    AbstractFeatureBuilder,
    AbstractTargetBuilder,
)

DEVKIT_ROOT = os.getenv('NAVSIM_DEVKIT_ROOT')
TRAJ_PDM_ROOT = os.getenv('NAVSIM_TRAJPDM_ROOT')

from typing import Dict, List

import pytorch_lightning as pl
import torch

from navsim.agents.abstract_agent import AbstractAgent


class HydraDreamerPlanningAgent(AbstractAgent):
    def __init__(
            self,
            config: HydraDreamerConfig,
            lr: float,
            checkpoint_path: str = None,
            dreamer_ckpt_path: str = None,
            pdm_split=None,
            metrics=None,
    ):
        super().__init__()
        config.trajectory_pdm_weight = {
            'noc': 3.0,
            'da': 3.0,
            'ttc': 2.0,
            'progress': config.progress_weight,
            'comfort': 1.0,
        }
        self._config = config
        self._lr = lr
        self.metrics = metrics
        self._checkpoint_path = checkpoint_path
        self.dreamer_ckpt_path = dreamer_ckpt_path
        self.vadv2_model = HydraDreamerPlanningModel(config)
        self.vocab_size = config.vocab_size
        self.backbone_wd = config.backbone_wd
        new_pkl_dir = f'vocab_score_full_{self.vocab_size}_navtrain'
        self.vocab_pdm_score_full = pickle.load(
            open(f'{TRAJ_PDM_ROOT}/{new_pkl_dir}/{pdm_split}.pkl', 'rb'))

    def name(self) -> str:
        """Inherited, see superclass."""

        return self.__class__.__name__

    def initialize(self) -> None:
        """Inherited, see superclass."""
        planner_state_dict: Dict[str, Any] = torch.load(
            self._checkpoint_path,
            map_location=torch.device("cpu")
        )["state_dict"]
        dreamer_state_dict: Dict[str, Any] = torch.load(
            self.dreamer_ckpt_path,
            map_location=torch.device("cpu")
        )["state_dict"]
        state_dict = {}
        for k, v in planner_state_dict.items():
            # ignore backbone
            if '_backbone' not in k:
                state_dict[k] = v
        for k, v in dreamer_state_dict.items():
            new_k = k.replace('agent.', 'agent.vadv2_model.')
            state_dict[new_k] = v

        self.load_state_dict({k.replace("agent.", ""): v for k, v in state_dict.items()})

    def get_sensor_config(self) -> SensorConfig:
        """Inherited, see superclass."""
        return SensorConfig(
            cam_f0=True,
            cam_l0=True,
            cam_l1=True,
            cam_l2=True,
            cam_r0=True,
            cam_r1=True,
            cam_r2=True,
            cam_b0=True,
            lidar_pc=[],
        )

    def get_target_builders(self) -> List[AbstractTargetBuilder]:
        return [HydraTargetBuilder(config=self._config)]

    def get_feature_builders(self) -> List[AbstractFeatureBuilder]:
        return [HydraDreamerWmFeatureBuilder(config=self._config)]

    def forward(self, features: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
        return self.vadv2_model(features)

    def forward_train(self, features, interpolated_traj):
        return self.vadv2_model(features, interpolated_traj)

    def compute_loss(
            self,
            features: Dict[str, torch.Tensor],
            targets: Dict[str, torch.Tensor],
            predictions: Dict[str, torch.Tensor],
            tokens=None
    ) -> Union[torch.Tensor, Dict[str, torch.Tensor]]:
        # get the pdm score by tokens
        scores = {}
        for k in self.metrics:
            tmp = [self.vocab_pdm_score_full[token][k][None] for token in tokens]
            scores[k] = (torch.from_numpy(np.concatenate(tmp, axis=0))
                         .to(predictions['trajectory'].device))
        return hydra_kd_imi_agent_loss(targets, predictions, self._config, scores)

    def get_optimizers(self) -> Union[Optimizer, Dict[str, Union[Optimizer, LRScheduler]]]:
        backbone_params_name = '_backbone.image_encoder'
        img_backbone_params = list(
            filter(lambda kv: backbone_params_name in kv[0], self.vadv2_model.named_parameters()))
        default_params = list(filter(lambda kv: backbone_params_name not in kv[0], self.vadv2_model.named_parameters()))
        params_lr_dict = [
            {'params': [tmp[1] for tmp in default_params]},
            {
                'params': [tmp[1] for tmp in img_backbone_params],
                'lr': self._lr * self._config.lr_mult_backbone,
                'weight_decay': self.backbone_wd
            }
        ]
        return torch.optim.Adam(params_lr_dict, lr=self._lr)

    def get_training_callbacks(self) -> List[pl.Callback]:
        return [
            # TransfuserCallback(self._config),
            ModelCheckpoint(
                save_top_k=30,
                monitor="val/loss_epoch",
                mode="min",
                dirpath=f"{os.environ.get('NAVSIM_EXP_ROOT')}/{self._config.ckpt_path}/",
                filename="{epoch:02d}-{step:04d}",
            )
        ]