jmercat's picture
Removed history to avoid any unverified information being released
5769ee4
from dataclasses import dataclass
from functools import partial
from typing import Callable, List, Optional, Tuple, Union
from einops import repeat
from mmcv import Config
import pytorch_lightning as pl
import torch
from risk_biased.models.cvae_params import CVAEParams
from risk_biased.models.biased_cvae_model import (
cvae_factory,
)
from risk_biased.utils.cost import TTCCostTorch, TTCCostParams
from risk_biased.utils.risk import get_risk_estimator
from risk_biased.utils.risk import get_risk_level_sampler
@dataclass
class LitTrajectoryPredictorParams:
"""
cvae_params: CVAEParams class defining the necessary parameters for the CVAE model
risk distribution: dict of string and values defining the risk distribution to use
risk_estimator: dict of string and values defining the risk estimator to use
kl_weight: float defining the weight of the KL term in the loss function
kl_threshold: float defining the threshold to apply when computing kl divergence (avoid posterior collapse)
risk_weight: float defining the weight of the risk term in the loss function
n_mc_samples_risk: int defining the number of Monte Carlo samples to use when estimating the risk
n_mc_samples_biased: int defining the number of Monte Carlo samples to use when estimating the expected biased cost
dt: float defining the duration between two consecutive time steps
learning_rate: float defining the learning rate for the optimizer
use_risk_constraint: bool defining whether to use the risk constrained optimization procedure
risk_constraint_update_every_n_epoch: int defining the number of epochs between two risk weight updates
risk_constraint_weight_update_factor: float defining the factor by which the risk weight is multiplied at each update
risk_constraint_weight_maximum: float defining the maximum value of the risk weight
num_samples_min_fde: int defining the number of samples to use when estimating the minimum FDE
condition_on_ego_future: bool defining whether to condition the biasing on the ego future trajectory (else on the ego past)
"""
cvae_params: CVAEParams
risk_distribution: dict
risk_estimator: dict
kl_weight: float
kl_threshold: float
risk_weight: float
n_mc_samples_risk: int
n_mc_samples_biased: int
dt: float
learning_rate: float
use_risk_constraint: bool
risk_constraint_update_every_n_epoch: int
risk_constraint_weight_update_factor: float
risk_constraint_weight_maximum: float
num_samples_min_fde: int
condition_on_ego_future: bool
@staticmethod
def from_config(cfg: Config):
cvae_params = CVAEParams.from_config(cfg)
return LitTrajectoryPredictorParams(
risk_distribution=cfg.risk_distribution,
risk_estimator=cfg.risk_estimator,
kl_weight=cfg.kl_weight,
kl_threshold=cfg.kl_threshold,
risk_weight=cfg.risk_weight,
n_mc_samples_risk=cfg.n_mc_samples_risk,
n_mc_samples_biased=cfg.n_mc_samples_biased,
dt=cfg.dt,
learning_rate=cfg.learning_rate,
cvae_params=cvae_params,
use_risk_constraint=cfg.use_risk_constraint,
risk_constraint_update_every_n_epoch=cfg.risk_constraint_update_every_n_epoch,
risk_constraint_weight_update_factor=cfg.risk_constraint_weight_update_factor,
risk_constraint_weight_maximum=cfg.risk_constraint_weight_maximum,
num_samples_min_fde=cfg.num_samples_min_fde,
condition_on_ego_future=cfg.condition_on_ego_future,
)
class LitTrajectoryPredictor(pl.LightningModule):
"""Pytorch Lightning Module for Trajectory Prediction with the biased cvae model
Args:
params : dataclass object containing the necessary parameters
cost_params: dataclass object defining the TTC cost function
unnormalizer: function that takes in a trajectory and an offset and that outputs the
unnormalized trajectory
"""
def __init__(
self,
params: LitTrajectoryPredictorParams,
cost_params: TTCCostParams,
unnormalizer: Callable[[torch.Tensor, torch.Tensor], torch.Tensor],
) -> None:
super().__init__()
model = cvae_factory(
params.cvae_params,
cost_function=TTCCostTorch(cost_params),
risk_estimator=get_risk_estimator(params.risk_estimator),
training_mode="cvae",
)
self.model = model
self.params = params
self._unnormalize_trajectory = unnormalizer
self.set_training_mode("cvae")
self.learning_rate = params.learning_rate
self.num_samples_min_fde = params.num_samples_min_fde
self.dynamic_state_dim = params.cvae_params.dynamic_state_dim
self.dt = params.cvae_params.dt
self.use_risk_constraint = params.use_risk_constraint
self.risk_weight = params.risk_weight
self.risk_weight_ratio = params.risk_weight / params.kl_weight
self.kl_weight = params.kl_weight
if self.use_risk_constraint:
self.risk_constraint_update_every_n_epoch = (
params.risk_constraint_update_every_n_epoch
)
self.risk_constraint_weight_update_factor = (
params.risk_constraint_weight_update_factor
)
self.risk_constraint_weight_maximum = params.risk_constraint_weight_maximum
self._risk_sampler = get_risk_level_sampler(params.risk_distribution)
def set_training_mode(self, training_mode: str):
self.model.set_training_mode(training_mode)
self.partial_get_loss = partial(
self.model.get_loss,
kl_threshold=self.params.kl_threshold,
n_samples_risk=self.params.n_mc_samples_risk,
n_samples_biased=self.params.n_mc_samples_biased,
dt=self.params.dt,
unnormalizer=self._unnormalize_trajectory,
)
def _get_loss(
self,
x: torch.Tensor,
mask_x: torch.Tensor,
map: torch.Tensor,
mask_map: torch.Tensor,
y: torch.Tensor,
mask_y: torch.Tensor,
mask_loss: torch.Tensor,
x_ego: torch.Tensor,
y_ego: torch.Tensor,
offset: Optional[torch.Tensor] = None,
risk_level: Optional[torch.Tensor] = None,
) -> Tuple[Union[torch.Tensor, Tuple[torch.Tensor, ...]], dict]:
"""Compute loss based on trajectory history x and future y
Args:
x: (batch_size, num_agents, num_steps, state_dim) tensor of history
mask_x: (batch_size, num_agents, num_steps) tensor of bool mask
map: (batch_size, num_objects, object_sequence_length, map_feature_dim) tensor of encoded map objects
mask_map: (batch_size, num_objects, object_sequence_length) tensor True where map features are good False where it is padding
y: (batch_size, num_agents, num_steps_future, state_dim) tensor of future trajectory.
mask_y: (batch_size, num_agents, num_steps_future) tensor of bool mask.
mask_loss: (batch_size, num_agents, num_steps_future) tensor of bool mask set to True where the loss
should be computed and to False where it shouldn't
offset : (batch_size, num_agents, state_dim) offset position from ego
risk_level : (batch_size, num_agents) tensor of risk levels desired for future trajectories
Returns:
Union[torch.Tensor, Tuple[torch.Tensor, ...]]: (1,) loss tensor or tuple of
loss tensors
dict: dict that contains values to be logged
"""
return self.partial_get_loss(
x=x,
mask_x=mask_x,
map=map,
mask_map=mask_map,
y=y,
mask_y=mask_y,
mask_loss=mask_loss,
offset=offset,
risk_level=risk_level,
x_ego=x_ego,
y_ego=y_ego,
risk_weight=self.risk_weight,
kl_weight=self.kl_weight,
)
def log_with_prefix(
self,
log_dict: dict,
prefix: Optional[str] = None,
on_step: Optional[bool] = None,
on_epoch: Optional[bool] = None,
) -> None:
"""log entries in log_dict while optinally adding "<prefix>/" to its keys
Args:
log_dict: dict that contains values to be logged
prefix: prefix to be added to keys
on_step: if True logs at this step. None auto-logs at the training_step but not
validation/test_step
on_epoch: if True logs epoch accumulated metrics. None auto-logs at the val/test
step but not training_step
"""
if prefix is None:
prefix = ""
else:
prefix += "/"
for (metric, value) in log_dict.items():
metric = prefix + metric
self.log(metric, value, on_step=on_step, on_epoch=on_epoch)
def configure_optimizers(
self,
) -> Union[torch.optim.Optimizer, List[torch.optim.Optimizer]]:
"""Configure optimizer for PyTorch-Lightning
Returns:
torch.optim.Optimizer: optimizer to be used for training
"""
if isinstance(self.model.get_parameters(), list):
self._optimizers = [
torch.optim.Adam(params, lr=self.learning_rate)
for params in self.model.get_parameters()
]
else:
self._optimizers = [
torch.optim.Adam(self.model.get_parameters(), lr=self.learning_rate)
]
return self._optimizers
def training_step(
self,
batch: Tuple[
torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor
],
batch_idx: int,
) -> dict:
"""Training step definition for PyTorch-Lightning
Args:
batch : [(batch_size, num_agents, num_steps, state_dim), # past trajectories of all agents in the scene
(batch_size, num_agents, num_steps), # mask past False where past trajectories are padding data
(batch_size, num_agents, num_steps_future, state_dim), # future trajectory
(batch_size, num_agents, num_steps_future), # mask future False where future trajectories are padding data
(batch_size, num_agents, num_steps_future), # mask loss False where future trajectories are not to be predicted
(batch_size, num_objects, object_seq_len, state_dim), # map object sequences in the scene
(batch_size, num_objects, object_seq_len), # mask map False where map objects are padding data
(batch_size, num_agents, state_dim), # position offset of all agents relative to ego at present time
(batch_size, 1, num_steps, state_dim), # ego past trajectory
(batch_size, 1, num_steps_future, state_dim)] # ego future trajectory
batch_idx : batch_idx to be used by PyTorch-Lightning
Returns:
dict: dict of outputs containing loss
"""
x, mask_x, y, mask_y, mask_loss, map, mask_map, offset, x_ego, y_ego = batch
risk_level = repeat(
self._risk_sampler.sample(x.shape[0], x.device),
"b -> b num_agents",
num_agents=x.shape[1],
)
loss, log_dict = self._get_loss(
x=x,
mask_x=mask_x,
map=map,
mask_map=mask_map,
y=y,
mask_y=mask_y,
mask_loss=mask_loss,
offset=offset,
risk_level=risk_level,
x_ego=x_ego,
y_ego=y_ego,
)
if isinstance(loss, tuple):
loss = sum(loss)
self.log_with_prefix(log_dict, prefix="train", on_step=True, on_epoch=False)
return {"loss": loss}
def training_epoch_end(self, outputs: List[dict]) -> None:
"""Called at the end of the training epoch with the outputs of all training steps
Args:
outputs: list of outputs of all training steps in the current epoch
"""
if self.use_risk_constraint:
if (
self.model.training_mode == "bias"
and (self.trainer.current_epoch + 1)
% self.risk_constraint_update_every_n_epoch
== 0
):
self.risk_weight_ratio *= self.risk_constraint_weight_update_factor
if self.risk_weight_ratio < self.risk_constraint_weight_maximum:
sum_weight = self.risk_weight + self.kl_weight
self.risk_weight = (
sum_weight
* self.risk_weight_ratio
/ (1 + self.risk_weight_ratio)
)
self.kl_weight = sum_weight / (1 + self.risk_weight_ratio)
# self.risk_weight *= self.risk_constraint_weight_update_factor
# if self.risk_weight > self.risk_constraint_weight_maximum:
# self.risk_weight = self.risk_constraint_weight_maximum
def _get_risk_tensor(
self,
batch_size: int,
num_agents: int,
device: torch.device,
risk_level: Optional[torch.Tensor] = None,
):
"""This function is used to reformat different possible formattings of risk_level input arguments into a tensor of shape (batch_size).
If given a tensor the same tensor is returned.
If given a float value, a tensor of this value is returned.
If given None, a tensor filled with random samples is returned.
Args:
batch_size : desired batch size
device : device on which we want to store risk
risk_level : The risk level as a tensor, a float value or None
Returns:
_type_: _description_
"""
if risk_level is not None:
if isinstance(risk_level, float):
risk_level = (
torch.ones(batch_size, num_agents, device=device) * risk_level
)
else:
risk_level = risk_level.to(device)
else:
risk_level = None
return risk_level
def validation_step(
self,
batch: Tuple[
torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor
],
batch_idx: int,
risk_level: float = 1.0,
) -> dict:
"""Validation step definition for PyTorch-Lightning
Args:
batch : [(batch_size, num_agents, num_steps, state_dim), # past trajectories of all agents in the scene
(batch_size, num_agents, num_steps), # mask past False where past trajectories are padding data
(batch_size, num_agents, num_steps_future, state_dim), # future trajectory
(batch_size, num_agents, num_steps_future), # mask future False where future trajectories are padding data
(batch_size, num_agents, num_steps_future), # mask loss False where future trajectories are not to be predicted
(batch_size, num_objects, object_seq_len, state_dim), # map object sequences in the scene
(batch_size, num_objects, object_seq_len), # mask map False where map objects are padding data
(batch_size, num_agents, state_dim), # position offset of all agents relative to ego at present time
(batch_size, 1, num_steps, state_dim), # ego past trajectory
(batch_size, 1, num_steps_future, state_dim)] # ego future trajectory
batch_idx : batch_idx to be used by PyTorch-Lightning
risk_level : optional desired risk level
Returns:
dict: dict of outputs containing loss
"""
x, mask_x, y, mask_y, mask_loss, map, mask_map, offset, x_ego, y_ego = batch
risk_level = self._get_risk_tensor(
x.shape[0], x.shape[1], x.device, risk_level=risk_level
)
self.model.eval()
log_dict_accuracy = self.model.get_prediction_accuracy(
x=x,
mask_x=mask_x,
map=map,
mask_map=mask_map,
y=y,
mask_loss=mask_loss,
offset=offset,
x_ego=x_ego,
y_ego=y_ego,
unnormalizer=self._unnormalize_trajectory,
risk_level=risk_level,
num_samples_min_fde=self.num_samples_min_fde,
)
loss, log_dict_loss = self._get_loss(
x=x,
mask_x=mask_x,
map=map,
mask_map=mask_map,
y=y,
mask_y=mask_y,
mask_loss=mask_loss,
offset=offset,
risk_level=risk_level,
x_ego=x_ego,
y_ego=y_ego,
)
if isinstance(loss, tuple):
loss = sum(loss)
self.log_with_prefix(
dict(log_dict_accuracy, **log_dict_loss),
prefix="val",
on_step=False,
on_epoch=True,
)
self.model.train()
return {"loss": loss}
def test_step(
self,
batch: Tuple[
torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor
],
batch_idx: int,
risk_level: Optional[torch.Tensor] = None,
) -> dict:
"""Test step definition for PyTorch-Lightning
Args:
batch : [(batch_size, num_agents, num_steps, state_dim), # past trajectories of all agents in the scene
(batch_size, num_agents, num_steps), # mask past False where past trajectories are padding data
(batch_size, num_agents, num_steps_future, state_dim), # future trajectory
(batch_size, num_agents, num_steps_future), # mask future False where future trajectories are padding data
(batch_size, num_agents, num_steps_future), # mask loss False where future trajectories are not to be predicted
(batch_size, num_objects, object_seq_len, state_dim), # map object sequences in the scene
(batch_size, num_objects, object_seq_len), # mask map False where map objects are padding data
(batch_size, num_agents, state_dim), # position offset of all agents relative to ego at present time
(batch_size, 1, num_steps, state_dim), # ego past trajectory
(batch_size, 1, num_steps_future, state_dim)] # ego future trajectory
batch_idx : batch_idx to be used by PyTorch-Lightning
risk_level : optional desired risk level
Returns:
dict: dict of outputs containing loss
"""
x, mask_x, y, mask_y, mask_loss, map, mask_map, offset, x_ego, y_ego = batch
risk_level = self._get_risk_tensor(
x.shape[0], x.shape[1], x.device, risk_level=risk_level
)
loss, log_dict = self._get_loss(
x=x,
mask_x=mask_x,
map=map,
mask_map=mask_map,
y=y,
mask_y=mask_y,
mask_loss=mask_loss,
offset=offset,
risk_level=risk_level,
x_ego=x_ego,
y_ego=y_ego,
)
if isinstance(loss, tuple):
loss = sum(loss)
self.log_with_prefix(log_dict, prefix="test", on_step=False, on_epoch=True)
return {"loss": loss}
def predict_step(
self,
batch: Tuple[torch.Tensor, torch.Tensor],
batch_idx: int = 0,
risk_level: Optional[torch.Tensor] = None,
n_samples: int = 0,
return_weights: bool = False,
) -> torch.Tensor:
"""Predict step definition for PyTorch-Lightning
Args:
batch: [(batch_size, num_agents, num_steps, state_dim), # past trajectories of all agents in the scene
(batch_size, num_agents, num_steps), # mask past False where past trajectories are padding data
(batch_size, num_objects, object_seq_len, state_dim), # map object sequences in the scene
(batch_size, num_objects, object_seq_len), # mask map False where map objects are padding data
(batch_size, num_agents, state_dim), # position offset of all agents relative to ego at present time
(batch_size, 1, num_steps, state_dim), # past trajectory of the ego agent in the scene
(batch_size, 1, num_steps_future, state_dim),] # future trajectory of the ego agent in the scene
batch_idx : batch_idx to be used by PyTorch-Lightning (unused here)
risk_level : optional desired risk level
n_samples: Number of samples to predict per agent
With value of 0 does not include the `n_samples` dim in the output.
return_weights: If True, also returns the sample weights
Returns:
(batch_size, (n_samples), num_steps_future, state_dim) tensor
"""
x, mask_x, map, mask_map, offset, x_ego, y_ego = batch
risk_level = self._get_risk_tensor(
batch_size=x.shape[0],
num_agents=x.shape[1],
device=x.device,
risk_level=risk_level,
)
y_sampled, weights, _ = self.model(
x,
mask_x,
map,
mask_map,
offset=offset,
x_ego=x_ego,
y_ego=y_ego,
risk_level=risk_level,
n_samples=n_samples,
)
predict_sampled = self._unnormalize_trajectory(y_sampled, offset)
if return_weights:
return predict_sampled, weights
else:
return predict_sampled
def predict_loop_once(
self,
batch: Tuple[torch.Tensor, torch.Tensor],
batch_idx: int = 0,
risk_level: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""Predict with refinment:
A first prediction is done as in predict_step, however instead of unnormalize and return it,
it is fed to the encoder that wast trained to encode past and ground truth future.
Then the decoder is used again but its latent input sample is biased by the encoder
instead of being a sample of the prior distribution.
Then as in predict_step the result is unnormalized and returned.
Args:
batch: [(batch_size, num_agents, num_steps, state_dim), # past trajectories of all agents in the scene
(batch_size, num_agents, num_steps), # mask past False where past trajectories are padding data
(batch_size, num_objects, object_seq_len, state_dim), # map object sequences in the scene
(batch_size, num_objects, object_seq_len), # mask map False where map objects are padding data
(batch_size, num_agents, state_dim),] # position offset of all agents relative to ego at present time
batch_idx : batch_idx to be used by PyTorch-Lightning (Unused here). Defaults to 0.
risk_level : optional desired risk level
Returns:
torch.Tensor: (batch_size, num_steps_future, state_dim) tensor
"""
x, mask_x, map, mask_map, offset = batch
risk_level = self._get_risk_tensor(
x.shape[0], x.shape[1], x.device, risk_level=risk_level
)
y_sampled, _ = self.model(
x,
mask_x,
map,
mask_map,
offset=offset,
risk_level=risk_level,
)
mask_y = repeat(mask_x.any(-1), "b a -> b a f", f=y_sampled.shape[-2])
y_sampled, _ = self.model(
x,
mask_x,
map,
mask_map,
y_sampled,
mask_y,
offset=offset,
risk_level=risk_level,
)
predict_sampled = self._unnormalize_trajectory(y_sampled, offset=offset)
return predict_sampled