"max_nge", "max_ngvl", "max_ngel", "share_emb", "share_arch"
]
]):
model.increase_input_size(finetune_config)
if not all([
train_config[key] == finetune_config[key] for key in [
"predict_net", "predict_net_hidden_dim",
"predict_net_num_heads", "predict_net_mem_len",
"predict_net_mem_init", "predict_net_recurrent_steps"
]
]):
new_predict_net = model.create_predict_net(
finetune_config["predict_net"],
pattern_dim=model.predict_net.pattern_dim,
graph_dim=model.predict_net.graph_dim,
hidden_dim=finetune_config["predict_net_hidden_dim"],
num_heads=finetune_config["predict_net_num_heads"],
recurrent_steps=finetune_config["predict_net_recurrent_steps"],
mem_len=finetune_config["predict_net_mem_len"],
mem_init=finetune_config["predict_net_mem_init"])
del model.predict_net
model.predict_net = new_predict_net
model = model.to(device)
torch.cuda.empty_cache()
logger.info("load the model based on the dev set (epoch: {:0>3d})".format(
best_epochs["dev"]))
logger.info(model)
logger.info("num of parameters: %d" %
(sum(p.numel()
for p in model.parameters() if p.requires_grad)))
