trajectron/environment/node.py

@@ -31,15 +31,6 @@ class Node(object):
 
         self.forward_in_time_on_next_override = False
 
-    @property
-    def first_timestep(self):
-        return self._first_timestep
-
-    @first_timestep.setter
-    def first_timestep(self, value):
-        self._first_timestep = value
-        self._last_timestep = None # reset
-
     def __eq__(self, other):
         return ((isinstance(other, self.__class__)
                  or isinstance(self, other.__class__))

trajectron/model/online/online_mgcvae.py

@@ -209,7 +209,7 @@ class OnlineMultimodalGenerativeCVAE(MultimodalGenerativeCVAE):
             if self.node not in maps:
                 # This means the node was removed (it is only being kept around because of the edge removal filter).
                 me_params = self.hyperparams['map_encoder'][self.node_type]
-                self.TD['encoded_map'] = torch.zeros((1, me_params['output_size'])).to(self.TD['node_history_encoded'].device)
+                self.TD['encoded_map'] = torch.zeros((1, me_params['output_size']))
             else:
                 encoded_map = self.node_modules[self.node_type + '/map_encoder'](maps[self.node] * 2. - 1.,
                                                                                  (mode == ModeKeys.TRAIN))

trajectron/train.py

@@ -1,5 +1,3 @@
-import logging
-from typing import List
 import torch
 from torch import nn, optim, utils
 import numpy as np
@@ -19,8 +17,6 @@ from trajectron.model.trajectron import Trajectron
 from trajectron.model.model_registrar import ModelRegistrar
 from trajectron.model.model_utils import cyclical_lr
 from trajectron.model.dataset import EnvironmentDataset, collate
-from trajectron.environment import Environment, Scene, Node
-
 from tensorboardX import SummaryWriter
 # torch.autograd.set_detect_anomaly(True)
 
@@ -138,7 +134,7 @@ def main():
                                        min_future_timesteps=hyperparams['prediction_horizon'],
                                        return_robot=not args.incl_robot_node)
     train_data_loader = dict()
-    logging.debug(f"{train_scenes=}")
+    print(train_scenes)
     for node_type_data_set in train_dataset:
         if len(node_type_data_set) == 0:
             continue
@@ -169,7 +165,7 @@ def main():
             for scene in eval_env.scenes:
                 scene.add_robot_from_nodes(eval_env.robot_type)
 
-        eval_scenes: List[Scene] = eval_env.scenes
+        eval_scenes = eval_env.scenes
         eval_scenes_sample_probs = eval_env.scenes_freq_mult_prop if args.scene_freq_mult_eval else None
 
         eval_dataset = EnvironmentDataset(eval_env,
@@ -182,7 +178,6 @@ def main():
                                           min_future_timesteps=hyperparams['prediction_horizon'],
                                           return_robot=not args.incl_robot_node)
         eval_data_loader = dict()
-        logging.debug(f"{eval_scenes=}")
         for node_type_data_set in eval_dataset:
             if len(node_type_data_set) == 0:
                 continue
@@ -392,7 +387,6 @@ def main():
                 # Predict batch timesteps for evaluation dataset evaluation
                 eval_batch_errors = []
                 for scene in tqdm(eval_scenes, desc='Sample Evaluation', ncols=80):
-                    logging.debug(f"{scene}, {scene.timesteps=}, {len(scene.nodes)}")
                     timesteps = scene.sample_timesteps(args.eval_batch_size)
 
                     predictions = eval_trajectron.predict(scene,
@@ -419,7 +413,6 @@ def main():
                 # Predict maximum likelihood batch timesteps for evaluation dataset evaluation
                 eval_batch_errors_ml = []
                 for scene in tqdm(eval_scenes, desc='MM Evaluation', ncols=80):
-                    logging.debug(f"{scene}, {scene.timesteps=}, {len(scene.nodes)}")
                     timesteps = scene.sample_timesteps(scene.timesteps)
 
                     predictions = eval_trajectron.predict(scene,
@@ -430,6 +423,7 @@ def main():
                                                           z_mode=True,
                                                           gmm_mode=True,
                                                           full_dist=False)
+
                     eval_batch_errors_ml.append(evaluation.compute_batch_statistics(predictions,
                                                                                     scene.dt,
                                                                                     max_hl=max_hl,