if self.args.multi_maps:
spatial_map = torch.cat(all_spatial_maps, dim=1).flatten(0, 1)
mask = torch.cat(all_masks, dim=1).flatten(0, 1)
seq = seq.unsqueeze(1).expand(-1, self.args.timesteps, -1).flatten(0, 1)
seq_lens = seq_lens.unsqueeze(1).expand(-1, self.args.timesteps).flatten(0, 1)
# Predict with unet
pred = self.navigator(seq, seq_lens, spatial_map)
# shape: (batch_size*timesteps, 2)
goal_pred = pred[:, 0, :, :]
path_pred = pred[:, 1, :, :]
return goal_pred, path_pred, mask
if __name__ == "__main__":
args = cfg.parse_args()
args_text = cfg.print_args(args)
if args.eval is False:
trainer = PVNBaselineTrainer(args=args)
trainer.train()
else:
filepath = "%s/%s/%s_%d" % (args.snapshot_dir, args.exp_name, args.exp_name, args.val_epoch)
trainer = PVNBaselineEvaluator(filepath=filepath, args=args)
trainer.evaluate("val_seen")
trainer.evaluate("val_unseen")
def __init__(self, args, filepath=None, load_sim=True):
self.args = copy.deepcopy(args)
if filepath:
# Copy args from the loaded config to self.args. This ensures old models can still
# be loaded even after new args are added to the config
self.loader = torch.load(filepath)
print("Loading from checkpoint: %s" % filepath)
for arg, value in vars(self.loader["args"]).items():
setattr(self.args, arg, value)
# eval parameters override
self.args.filepath = filepath
self.args.eval = args.eval
self.args.val_epoch = args.val_epoch
self.args.batch_size = args.batch_size
self.args.viz_eval = args.viz_eval
self.args.viz_gif = args.viz_gif
self.args.viz_iterations = args.viz_iterations
self.args.viz_folder = args.viz_folder
self.args.visdom = args.visdom
self.args.visdom_env = args.visdom_env
self.args.visdom_server = args.visdom_server
self.args.visdom_port = args.visdom_port
self.args.preloading = args.preloading
self.args.debug_scale = args.debug_scale
self.args.debug_mode = args.debug_mode
self.args.start_epoch = args.start_epoch
if self.args.visdom:
visdom_env = self.args.exp_name if self.args.visdom_env == "" else self.args.visdom_env
self.visdom = VisdomVisualize(env_name=visdom_env,
win_prefix=self.args.exp_name + "_",
server=self.args.visdom_server,
port=self.args.visdom_port)
self.visdom.text(
cfg.print_args(self.args, verbose=False).replace("\n", "<br>"))
else:
self.visdom = None
# Set random seed before anything
seed = self.args.random_seed
if self.args.start_epoch != 1 and self.args.eval == False:
seed = self.args.start_epoch + seed
utils.set_random_seed(seed)
if load_sim:
self.sim = PanoSimulator(self.args)
# Load data (trajectories, scans, graphs, vocab)
if not self.args.eval:
splits = ['train', 'val_seen', 'val_unseen'
] if self.args.train_split == 'trainval' else ['train']
self.traindata = DataLoader(self.args, splits=splits)
self.valseendata = DataLoader(self.args, splits=['val_seen'])
self.valunseendata = DataLoader(self.args, splits=['val_unseen'])
self.all_scans = self.valseendata.scans | self.valunseendata.scans
if not self.args.eval:
self.all_scans |= self.traindata.scans
self.floorplan = Floorplan(
self.args.debug_scale / self.args.gridcellsize,
list(self.all_scans))