def train():
if not os.path.exists('train_model/'):
os.makedirs('train_model/')
if not os.path.exists('result/'):
os.makedirs('result/')
train_data, dev_data, word2id, id2word, char2id, opts = load_data(
vars(args))
model = UNet(opts)
if args.use_cuda:
model = model.cuda()
dev_batches = get_batches(dev_data, args.batch_size, evaluation=True)
if args.eval:
print('load model...')
model.load_state_dict(torch.load(args.model_dir))
model.eval()
model.Evaluate(dev_batches,
args.data_path + 'dev_eval.json',
answer_file='result/' + args.model_dir.split('/')[-1] +
'.answers',
drop_file=args.data_path + 'drop.json',
dev=args.data_path + 'dev-v2.0.json')
exit()
if args.load_model:
print('load model...')
model.load_state_dict(torch.load(args.model_dir))
model.eval()
_, F1 = model.Evaluate(dev_batches,
args.data_path + 'dev_eval.json',
answer_file='result/' +
args.model_dir.split('/')[-1] + '.answers',
drop_file=args.data_path + 'drop.json',
dev=args.data_path + 'dev-v2.0.json')
best_score = F1
with open(args.model_dir + '_f1_scores.pkl', 'rb') as f:
f1_scores = pkl.load(f)
with open(args.model_dir + '_em_scores.pkl', 'rb') as f:
em_scores = pkl.load(f)
else:
best_score = 0.0
f1_scores = []
em_scores = []
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = torch.optim.Adamax(parameters, lr=args.lrate)
lrate = args.lrate
for epoch in range(1, args.epochs + 1):
train_batches = get_batches(train_data, args.batch_size)
dev_batches = get_batches(dev_data, args.batch_size, evaluation=True)
total_size = len(train_data) // args.batch_size
model.train()
for i, train_batch in enumerate(train_batches):
loss = model(train_batch)
model.zero_grad()
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(parameters, opts['grad_clipping'])
optimizer.step()
model.reset_parameters()
if i % 100 == 0:
print(
'Epoch = %d, step = %d / %d, loss = %.5f, lrate = %.5f best_score = %.3f'
% (epoch, i, total_size, model.train_loss.value, lrate,
best_score))
sys.stdout.flush()
model.eval()
exact_match_score, F1 = model.Evaluate(
dev_batches,
args.data_path + 'dev_eval.json',
answer_file='result/' + args.model_dir.split('/')[-1] + '.answers',
drop_file=args.data_path + 'drop.json',
dev=args.data_path + 'dev-v2.0.json')
f1_scores.append(F1)
em_scores.append(exact_match_score)
with open(args.model_dir + '_f1_scores.pkl', 'wb') as f:
pkl.dump(f1_scores, f)
with open(args.model_dir + '_em_scores.pkl', 'wb') as f:
pkl.dump(em_scores, f)
if best_score < F1:
best_score = F1
print('saving %s ...' % args.model_dir)
torch.save(model.state_dict(), args.model_dir)
if epoch > 0 and epoch % args.decay_period == 0:
lrate *= args.decay
for param_group in optimizer.param_groups:
param_group['lr'] = lrate
def train():
if not os.path.exists("train_model/"):
os.makedirs("train_model/")
if not os.path.exists("result/"):
os.makedirs("result/")
train_data, dev_data, word2id, id2word, char2id, opts = load_data(
vars(args))
model = UNet(opts)
if args.use_cuda:
model = model.cuda()
dev_batches = get_batches(dev_data, args.batch_size, evaluation=True)
if args.eval:
print("load model...")
model.load_state_dict(torch.load(args.model_dir))
model.eval()
model.Evaluate(
dev_batches,
os.path.join(args.prepro_dir, "dev_eval.json"),
answer_file="result/" + args.model_dir.split("/")[-1] + ".answers",
drop_file=os.path.join(args.prepro_dir, "drop.json"),
dev=args.dev_file,
)
exit()
if args.load_model:
print("load model...")
model.load_state_dict(torch.load(args.model_dir))
model.eval()
_, F1 = model.Evaluate(
dev_batches,
os.path.join(args.prepro_dir, "dev_eval.json"),
answer_file=os.path.join("result/",
args.model_dir.split("/")[-1],
".answers"),
drop_file=os.path.join(args.prepro_dir, "drop.json"),
dev=args.dev_file,
)
best_score = F1
with open(args.model_dir + "_f1_scores.pkl", "rb") as f:
f1_scores = pkl.load(f)
with open(args.model_dir + "_em_scores.pkl", "rb") as f:
em_scores = pkl.load(f)
else:
best_score = 0.0
f1_scores = []
em_scores = []
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = torch.optim.Adamax(parameters, lr=args.lrate)
lrate = args.lrate
for epoch in range(1, args.epochs + 1):
train_batches = get_batches(train_data, args.batch_size)
dev_batches = get_batches(dev_data, args.batch_size, evaluation=True)
total_size = len(train_data) // args.batch_size
model.train()
for i, train_batch in enumerate(train_batches):
loss = model(train_batch)
model.zero_grad()
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(parameters, opts["grad_clipping"])
optimizer.step()
model.reset_parameters()
if i % 100 == 0:
print(
"Epoch = %d, step = %d / %d, loss = %.5f, lrate = %.5f best_score = %.3f"
% (epoch, i, total_size, model.train_loss.value, lrate,
best_score))
sys.stdout.flush()
model.eval()
exact_match_score, F1 = model.Evaluate(
dev_batches,
os.path.join(args.prepro_dir, "dev_eval.json"),
answer_file=os.path.join("result/",
args.model_dir.split("/")[-1],
".answers"),
drop_file=os.path.join(args.prepro_dir, "drop.json"),
dev=args.dev_file,
)
f1_scores.append(F1)
em_scores.append(exact_match_score)
with open(args.model_dir + "_f1_scores.pkl", "wb") as f:
pkl.dump(f1_scores, f)
with open(args.model_dir + "_em_scores.pkl", "wb") as f:
pkl.dump(em_scores, f)
if best_score < F1:
best_score = F1
print("saving %s ..." % args.model_dir)
torch.save(model.state_dict(), args.model_dir)
if epoch > 0 and epoch % args.decay_period == 0:
lrate *= args.decay
for param_group in optimizer.param_groups:
param_group["lr"] = lrate
