def __init__(self, vocab, args, device): super().__init__() self.m_pad_idx = vocab.pad_idx self.m_Recon_loss_fn = None self.m_KL_loss_fn = None self.m_RRe_loss_fn = None self.m_ARe_loss_fn = None self.m_save_mode = True self.m_mean_en_train_loss = 0 self.m_mean_en_val_loss = 0 self.m_mean_de_train_loss = 0 self.m_mean_en_val_loss = 0 self.m_device = device self.m_epochs = args.epochs self.m_batch_size = args.batch_size self.m_x0 = args.x0 self.m_k = args.k self.m_anneal_func = args.anneal_func self.m_rec_loss = _REC_LOSS(self.m_device, ignore_index=self.m_pad_idx) self.m_rec_l_loss = _REC_LOSS(self.m_device, ignore_index=0) self.m_bow_loss = _BOW_LOSS(self.m_device) self.m_kl_loss_z = _KL_LOSS_CUSTOMIZE(self.m_device) self.m_kl_loss_s = _KL_LOSS_CUSTOMIZE(self.m_device) self.m_kl_loss_l = _KL_LOSS_STANDARD(self.m_device) self.m_train_step = 0 self.m_valid_step = 0 self.m_model_path = args.model_path self.m_model_file = args.model_file self.m_train_iteration = 0 self.m_valid_iteration = 0 self.m_print_interval = args.print_interval self.m_overfit_epoch_threshold = 3
def f_init_eval(self, network, model_file=None, reload_model=False):
if reload_model:
print("reload model")
if not model_file:
model_file = "model_best.pt"
model_name = os.path.join(self.m_model_path, model_file)
print("model name", model_name)
check_point = torch.load(model_name)
network.load_state_dict(check_point['model'])
self.m_network = network
self.m_Recon_loss_fn = _REC_LOSS(ignore_index=self.m_pad_idx,
device=self.m_device,
reduction=True)
self.m_KL_loss_z_fn = _KL_LOSS_STANDARD(self.m_device, reduction=True)
def __init__(self, vocab_obj, args, device):
super().__init__()
self.m_pad_idx = vocab_obj.pad_idx
self.m_save_mode = True
self.m_mean_train_loss = 0
self.m_mean_val_loss = 0
self.m_device = device
self.m_epochs = args.epochs
self.m_batch_size = args.batch_size
self.m_x0 = args.x0
self.m_k = args.k
self.m_anneal_func = args.anneal_func
print("args.anneal_func", args.anneal_func)
self.m_Recon_loss_fn = _REC_LOSS(ignore_index=self.m_pad_idx, device=self.m_device)
self.m_KL_loss_z_fn = _KL_LOSS_STANDARD(self.m_device)
self.m_train_step = 0
self.m_valid_step = 0
self.m_model_path = args.model_path
self.m_model_name = "VAE"
self.m_train_iteration = 0
self.m_valid_iteration = 0
self.m_KL_weight = 0.0
# self.m_KL_weight_list = []
self.m_grads = {}
self.m_model_file = args.model_file
self.m_train_loss_list = []
self.m_train_rec_loss_list = []
self.m_train_kl_loss_list = []
self.m_print_interval = args.print_interval