in multiple_futures_prediction/train_ngsim.py [0:0]
def __init__(self, log:bool=False, # save checkpoints?
modes:int=2, # how many latent modes
use_cuda:bool=True,
encoder_size:int=16, # encoder latent layer size
decoder_size:int=16, # decoder latent layer size
subsampling:int=2, # factor subsample in time
hist_len_orig_hz:int=30, # number of original history samples
fut_len_orig_hz:int=50, # number of original future samples
dyn_embedding_size:int=32, # dynamic embedding size
input_embedding_size:int=32, # input embedding size
dec_nbr_enc_size:int=8, # decoder neighbors encode size
nbr_atten_embedding_size:int=80, # neighborhood attention embedding size
seed:int=1234,
remove_y_mean:bool=False, # normalize by remove mean of the future trajectory
use_gru:bool=True, # GRUs instead of LSTMs
bi_direc:bool=False, # bidrectional
self_norm:bool=False, # normalize with respect to the current time
data_aug:bool=False, # data augment
use_context:bool=False, # use contextual image as additional input
nll:bool=True, # negative log-liklihood loss
use_forcing:int=0, # teacher forcing
iter_per_err:int=100, # iterations to display errors
iter_per_eval:int=1000, # iterations to eval on validation set
pre_train_num_updates:int=200000, # how many iterations for pretraining
updates_div_by_10:int=100000, # at what iteration to divide the learning rate by 10.0
nbr_search_depth:int=10, # how deep do we search for neighbors
lr_init:float=0.001, # initial learning rate
min_lr:float=0.00005, # minimal learning rate
iters_per_save:int=1500 ) -> None :
self.params = AttrDict(locals())