return sum_loss
def save_model(model, epoch, model_weight_path, checkpoint_path):
"""
save model weight without embedding
:param model:
:param epoch:
:param model_weight_path:
:param checkpoint_path:
:return:
"""
# save model weight
model_weight = model.state_dict()
del model_weight['embedding.embedding_layer.weight']
torch.save(model_weight, model_weight_path)
with open(checkpoint_path, 'w') as checkpoint_f:
checkpoint_f.write('epoch=%d' % epoch)
if __name__ == '__main__':
init_logging()
parser = argparse.ArgumentParser(description="train on the model")
parser.add_argument('--config', '-c', required=False, dest='config_path', default='config/global_config.yaml')
args = parser.parse_args()
train(args.config_path)
__author__ = 'han'
import os
import torch
import logging
import argparse
import torch.optim as optim
from dataset.squad_dataset import SquadDataset
from models.match_lstm import MatchLSTMModel
from models.loss import MyNLLLoss
from utils.load_config import init_logging, read_config
from utils.eval import eval_on_model
from utils.functions import pop_dict_keys
init_logging()
logger = logging.getLogger(__name__)
def main(config_path):
logger.info('------------Match-LSTM Train--------------')
logger.info('loading config file...')
global_config = read_config(config_path)
# set random seed
seed = global_config['model']['global']['random_seed']
torch.manual_seed(seed)
enable_cuda = global_config['train']['enable_cuda']
device = torch.device("cuda" if enable_cuda else "cpu")
if torch.cuda.is_available() and not enable_cuda:
parser = argparse.ArgumentParser(description="preprocess/train/test the model")
parser.add_argument('mode', help='preprocess or train or test')
parser.add_argument('--config',
'-c',
required=False,
dest='config_path',
default='config/global_config.yaml')
parser.add_argument('--output', '-o', required=False, dest='out_path')
parser.add_argument('--remark', required=False, dest='remark', default="")
args = parser.parse_args()
cur_time = time.strftime('%Y-%m-%d-%H_%M_%S', time.localtime())
host_name = socket.gethostname()
experiment_info = host_name + "_" + cur_time + "_" + args.remark
init_logging(experiment_info)
logger = logging.getLogger(__name__)
logger.info('======================== %s =================================' %
experiment_info)
if args.mode == 'preprocess':
preprocess(args.config_path)
elif args.mode == 'train':
train(args.config_path, experiment_info, training_info)
elif args.mode == 'debug':
debug(args.config_path, experiment_info)
elif args.mode == 'test':
test(args.config_path, experiment_info)
elif args.mode == 'train_5c':
train_5c(args.config_path, experiment_info, training_info)
elif args.mode == 'test_5c':
test_5c(args.config_path, experiment_info)
