pad_token=0,
use_vocab=False,
sequential=True)
GRAPH = GraphField(batch_first=True)
DOC.vocab = torch.load(args.vocab)
print('vocab {} loaded'.format(args.vocab))
args.__dict__.update({'doc_vocab': len(DOC.vocab)})
args_str = json.dumps(args.__dict__, indent=4, sort_keys=True)
print(args_str)
test_data = DocDataset(path=args.test,
text_field=DOC,
order_field=ORDER,
graph_field=GRAPH)
test_real = DocIter(test_data,
1,
device="cuda" if args.gpu else "cpu",
batch_size_fn=None,
train=False,
repeat=False,
shuffle=False,
sort=False)
print('{} Load data done'.format(curtime()))
start = time.time()
decode(args, test_real, (DOC, ORDER), checkpoint)
print('{} Decode done, time {} mins'.format(
curtime(), (time.time() - start) / 60))
def run_model(args):
if args.mode == 'train' or args.mode=='example':
if args.load_from is not None and len(args.load_from) == 1:
load_from = args.load_from[0]
print('{} load the checkpoint from {} for initilize or resume'.
format(curtime(), load_from))
checkpoint = torch.load(load_from, map_location='cpu')
else:
checkpoint = None
# if not resume(initilize), only need model parameters
if args.resume:
print('update args from checkpoint')
load_dict = checkpoint['args'].__dict__
except_name = ['mode', 'resume', 'maximum_steps']
override(args, load_dict, tuple(except_name))
main_path = Path(args.main_path)
model_path = main_path / args.model_path
decoding_path = main_path / args.decoding_path
for path in [model_path, decoding_path]:
path.mkdir(parents=True, exist_ok=True)
args.model_path = str(model_path)
args.decoding_path = str(decoding_path)
if args.model == '[time]':
args.model = time.strftime("%m.%d_%H.%M.", time.gmtime())
# setup random seeds
set_seeds(args.seed)
# special process, shuffle each document
# DOC = DocField(batch_first=True, include_lengths=True, eos_token='<eos>', init_token='<bos>')
DOC = DocField(batch_first=True, include_lengths=True)
ORDER = data.Field(batch_first=True, include_lengths=True, pad_token=0, use_vocab=False,
sequential=True)
GRAPH = GraphField(batch_first=True)
LABEL_FIELD=data.Field(batch_first=True, include_lengths=True, use_vocab=False)
train_data = DocDataset(path=args.corpus, text_field=DOC, order_field=ORDER, graph_field=GRAPH)
dev_data = DocDataset(path=args.valid, text_field=DOC, order_field=ORDER, graph_field=GRAPH)
DOC.vocab = torch.load(args.vocab)
print('vocab {} loaded'.format(args.vocab))
args.__dict__.update({'doc_vocab': len(DOC.vocab)})
train_flag = True
train_real = DocIter(train_data, args.batch_size, device='cuda',
train=train_flag,
shuffle=train_flag,
sort_key=lambda x: len(x.doc))
devbatch = 1
dev_real = DocIter(dev_data, devbatch, device='cuda', batch_size_fn=None,
train=False, repeat=False, shuffle=False, sort=False)
args_str = json.dumps(args.__dict__, indent=4, sort_keys=True)
print(args_str)
print('{} Start training'.format(curtime()))
train(args, train_real, dev_real, (DOC, ORDER, GRAPH), checkpoint)
else:
if len(args.load_from) == 1:
load_from = '{}.coqa_best.pt'.format(args.load_from[0])
print('{} load the best checkpoint from {}'.format(curtime(), load_from))
checkpoint = torch.load(load_from, map_location='cpu')
else:
raise RuntimeError('must load model')
# when translate load_dict update args except some
print('update args from checkpoint')
load_dict = checkpoint['args'].__dict__
except_name = ['mode', 'load_from', 'test', 'writetrans', 'beam_size', 'batch_size']
override(args, load_dict, tuple(except_name))
print('{} Load test set'.format(curtime()))
DOC = DocField(batch_first=True, include_lengths=True)
ORDER = data.Field(batch_first=True, include_lengths=True, pad_token=0, use_vocab=False,
sequential=True)
GRAPH = GraphField(batch_first=True)
DOC.vocab = torch.load(args.vocab)
print('vocab {} loaded'.format(args.vocab))
args.__dict__.update({'doc_vocab': len(DOC.vocab)})
args_str = json.dumps(args.__dict__, indent=4, sort_keys=True)
print(args_str)
test_data = DocDataset(path=args.test, text_field=DOC, order_field=ORDER, graph_field=GRAPH)
test_real = DocIter(test_data, 1, device='cuda', batch_size_fn=None,
train=False, repeat=False, shuffle=False, sort=False)
print('{} Load data done'.format(curtime()))
start = time.time()
decode(args, test_real, (DOC, ORDER), checkpoint)
print('{} Decode done, time {} mins'.format(curtime(), (time.time() - start) / 60))