def main():
parser = argparse.ArgumentParser()
parser.add_argument('--char-emb-dim', default=8, type=int)
parser.add_argument('--char-channel-width', default=5, type=int)
parser.add_argument('--char-channel-num', default=100, type=int)
parser.add_argument('--dev-batch-size', default=60, type=int)
parser.add_argument('--disable-c2q', type=ast.literal_eval)
parser.add_argument('--disable-q2c', type=ast.literal_eval)
parser.add_argument('--dropout', default=0.2, type=float)
parser.add_argument('--epoch', default=12, type=int)
parser.add_argument('--gpu', default=0, type=int)
parser.add_argument('--hidden-size', default=100, type=int)
parser.add_argument('--learning-rate', default=0.5, type=float)
parser.add_argument('--moving-average-decay', default=0.999, type=float)
parser.add_argument('--squad-version', default='1.1')
parser.add_argument('--train-batch-size', default=60, type=int)
parser.add_argument('--word-vec-dim', default=100, type=int)
parser.add_argument('--validation-freq', default=500, type=int)
args = parser.parse_args()
device = torch.device(f"cuda:{args.gpu}" if torch.cuda.is_available() else "cpu")
print("Device is set to", device)
print(f"Load SQuAD {args.squad_version}")
data = SQuAD(device=device,
squad_version=args.squad_version,
word_vec_dim=args.word_vec_dim,
train_batch_size=args.train_batch_size,
dev_batch_size=args.dev_batch_size)
setattr(args, 'char_vocab_size', len(data.CHAR_NESTING.vocab))
setattr(args, 'word_vocab_size', len(data.WORD.vocab))
print(f"Load BiDAF Model")
model = BiDAF(pretrain_embedding=data.WORD.vocab.vectors,
char_vocab_size=len(data.CHAR_NESTING.vocab),
consider_na=True if args.squad_version == "2.0" else False,
enable_c2q= not args.disable_c2q if args.disable_c2q is not None else True,
enable_q2c= not args.disable_q2c if args.disable_q2c is not None else True,
hidden_size=args.hidden_size,
char_emb_dim=args.char_emb_dim,
char_channel_num=args.char_channel_num,
char_channel_width=args.char_channel_width,
dropout=args.dropout)
print(f"Training start")
trained_model = train(device=device,
data=data,
model=model,
epoch=args.epoch,
lr=args.learning_rate,
moving_average_decay=args.moving_average_decay,
validation_freq=args.validation_freq)
model_time = int(time.time())
if not os.path.exists('models'):
os.makedirs('models')
torch.save(trained_model.state_dict(), f'trained_models/BiDAF_{model_time}.pt')
def train(args, data):
device = torch.device(f"cuda:{args.gpu}" if torch.cuda.is_available() else "cpu")
model = BiDAF(args, data.WORD.vocab.vectors).to(device)
ema = EMA(args.exp_decay_rate)
for name, param in model.named_parameters():
if param.requires_grad:
ema.register(name, param.data)
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = optim.Adadelta(parameters, lr=args.learning_rate)
criterion = nn.CrossEntropyLoss()
model.train()
loss, last_epoch = 0, -1
max_dev_exact, max_dev_f1 = -1, -1
iterator = data.train_iter
for i, batch in enumerate(iterator):
present_epoch = int(iterator.epoch)
if present_epoch == args.epoch:
break
if present_epoch > last_epoch:
print('epoch:', present_epoch + 1)
last_epoch = present_epoch
p1, p2 = model(batch)
optimizer.zero_grad()
batch_loss = criterion(p1, batch.s_idx) + criterion(p2, batch.e_idx)
loss += batch_loss.item()
batch_loss.backward()
optimizer.step()
for name, param in model.named_parameters():
if param.requires_grad:
ema.update(name, param.data)
if (i + 1) % args.print_freq == 0:
dev_loss, dev_exact, dev_f1 = test(model, ema, args, data)
c = (i + 1) // args.print_freq
print(f'train loss: {loss:.3f} / dev loss: {dev_loss:.3f}'
f' / dev EM: {dev_exact:.3f} / dev F1: {dev_f1:.3f}')
if dev_f1 > max_dev_f1:
max_dev_f1 = dev_f1
max_dev_exact = dev_exact
best_model = copy.deepcopy(model)
loss = 0
model.train()
print(f'max dev EM: {max_dev_exact:.3f} / max dev F1: {max_dev_f1:.3f}')
return best_model
def train_pipeline(args):
field = data.SpanFieldCollection()
train_loader = data.SpanDataLoader(args.pre_train_path, field)
dev_loader = data.SpanDataLoader(args.pre_dev_path, field)
field.build_vocab(None, train_loader.dataset, dev_loader.dataset)
model = BiDAF(**argument_wrapper.get_model_args(field))
if args.pretrained_wv_path is not None:
pretrained = load_pretrained_wv(args.pretrained_wv_path)
model.set_word_embedding(pretrained)
train_iter = train_loader.get_batchiter(args.batch_size, args.gpu_id)
dev_iter = dev_loader.get_batchiter(args.batch_size, args.gpu_id)
test_func = rc.create_test_func('foo')
best_model = rc.train_by_span(model,train_iter,dev_iter,RawCorpus(args.dev_path).get_answers(),test_func,\
args.epoch_num,args.learning_rate,args.exp_decay_rate,args.save_freq)
best_model.dump('model.pt')
setattr(options, 'dataset_file',
'.data/squad/{}'.format(options.dev_file))
setattr(options, 'prediction_file',
os.path.join(root_dir, 'prediction.json'))
setattr(options, 'model_time', strftime('%H:%M:%S', gmtime()))
logger.info('data loading complete!')
options.old_model = best_model_file_name
options.old_ema = best_ema
answer_append_sentences = joblib.load(
'sampled_perturb_answer_sentences.pkl')
question_append_sentences = joblib.load(
'sampled_perturb_question_sentences.pkl')
model = BiDAF(options, data.WORD.vocab.vectors).to(device)
if options.old_model is not None:
model.load_state_dict(
torch.load(options.old_model,
map_location="cuda:{}".format(options.gpu)))
if options.old_ema is not None:
# ema = pickle.load(open(options.old_ema, "rb"))
ema = torch.load(options.old_ema, map_location=device)
else:
ema = EMA(options.exp_decay_rate)
for name, param in model.named_parameters():
if param.requires_grad:
ema.register(name, param.data)
torch.manual_seed(args.seed)
if torch.cuda.is_available():
def train(args, data):
device = torch.device(
"cuda:{}".format(args.gpu) if torch.cuda.is_available() else "cpu")
model = BiDAF(args, data.WORD.vocab.vectors).to(device)
ema = EMA(args.exp_decay_rate)
for name, param in model.named_parameters():
if param.requires_grad:
ema.register(name, param.data)
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = optim.Adadelta(parameters, lr=args.learning_rate)
criterion = nn.CrossEntropyLoss()
writer = SummaryWriter(log_dir='runs/' + args.model_time)
model.train()
loss, last_epoch = 0, -1
max_dev_exact, max_dev_f1 = -1, -1
iterator = data.train_iter
for i, batch in enumerate(iterator):
present_epoch = int(iterator.epoch)
if present_epoch == args.epoch:
break
if present_epoch > last_epoch:
print('epoch:', present_epoch + 1)
last_epoch = present_epoch
p1, p2 = model(batch)
optimizer.zero_grad()
batch_loss = criterion(p1, batch.s_idx) + criterion(p2, batch.e_idx)
loss += batch_loss.item()
batch_loss.backward()
optimizer.step()
for name, param in model.named_parameters():
if param.requires_grad:
ema.update(name, param.data)
if (i + 1) % args.print_freq == 0:
dev_loss, dev_exact, dev_f1 = test(model, ema, args, data)
c = (i + 1) // args.print_freq
writer.add_scalar('loss/train', loss, c)
writer.add_scalar('loss/dev', dev_loss, c)
writer.add_scalar('exact_match/dev', dev_exact, c)
writer.add_scalar('f1/dev', dev_f1, c)
print('train loss: {} / dev loss: {}'.format(loss, dev_loss) +
' / dev EM: {} / dev F1: {}'.format(dev_exact, dev_f1))
if dev_f1 > max_dev_f1:
max_dev_f1 = dev_f1
max_dev_exact = dev_exact
best_model = copy.deepcopy(model)
loss = 0
model.train()
writer.close()
print('max dev EM: {} / max dev F1: {}'.format(max_dev_exact, max_dev_f1))
return best_model
def train(args, data):
if args.load_model != "":
model = BiDAF(args, data.WORD.vocab.vectors)
model.load_state_dict(torch.load(args.load_model))
else:
model = BiDAF(args, data.WORD.vocab.vectors)
device = torch.device(f"cuda:{args.gpu}" if torch.cuda.is_available() else "cpu")
model = model.to(device)
ema = EMA(args.exp_decay_rate)
for name, param in model.named_parameters():
if param.requires_grad:
ema.register(name, param.data)
for name, i in model.named_parameters():
if not i.is_leaf:
print(name,i)
writer = SummaryWriter(log_dir='runs/' + args.model_name)
best_model = None
for iterator, dev_iter, dev_file_name, index, print_freq, lr in zip(data.train_iter, data.dev_iter, args.dev_files, range(len(data.train)), args.print_freq, args.learning_rate):
# print
# (iterator[0])
embed()
exit(0)
optimizer = optim.Adadelta(model.parameters(), lr=lr)
criterion = nn.CrossEntropyLoss()
model.train()
loss, last_epoch = 0, 0
max_dev_exact, max_dev_f1 = -1, -1
print(f"Training with {dev_file_name}")
print()
for i, batch in tqdm(enumerate(iterator), total=len(iterator) * args.epoch[index], ncols=100):
present_epoch = int(iterator.epoch)
eva = False
if present_epoch == args.epoch[index]:
break
if present_epoch > last_epoch:
print('epoch:', present_epoch + 1)
eva = True
last_epoch = present_epoch
p1, p2 = model(batch)
optimizer.zero_grad()
batch_loss = criterion(p1, batch.s_idx) + criterion(p2, batch.e_idx)
loss += batch_loss.item()
batch_loss.backward()
optimizer.step()
for name, param in model.named_parameters():
if param.requires_grad:
ema.update(name, param.data)
torch.cuda.empty_cache()
if (i + 1) % print_freq == 0 or eva:
dev_loss, dev_exact, dev_f1 = test(model, ema, args, data, dev_iter, dev_file_name)
c = (i + 1) // print_freq
writer.add_scalar('loss/train', loss, c)
writer.add_scalar('loss/dev', dev_loss, c)
writer.add_scalar('exact_match/dev', dev_exact, c)
writer.add_scalar('f1/dev', dev_f1, c)
print()
print(f'train loss: {loss:.3f} / dev loss: {dev_loss:.3f}'
f' / dev EM: {dev_exact:.3f} / dev F1: {dev_f1:.3f}')
if dev_f1 > max_dev_f1:
max_dev_f1 = dev_f1
max_dev_exact = dev_exact
best_model = copy.deepcopy(model)
loss = 0
model.train()
writer.close()
print(f'max dev EM: {max_dev_exact:.3f} / max dev F1: {max_dev_f1:.3f}')
print("testing with test batch on best model")
test_loss, test_exact, test_f1 = test(best_model, ema, args, data, list(data.test_iter)[-1], args.test_files[-1])
print(f'test loss: {test_loss:.3f}'
f' / test EM: {test_exact:.3f} / test F1: {test_f1:.3f}')
return best_model
def train(args, data):
device = torch.device(f"cuda:{args.gpu}" if torch.cuda.is_available() else "cpu")
model = BiDAF(args, data.CONTEXT_WORD.vocab.vectors).to(device)
num = count_parameters(model)
print(f'paramter {num}')
if torch.cuda.device_count() > 1:
print("Let's use", torch.cuda.device_count(), "GPUs!")
model = nn.DataParallel(model)
ema = EMA(args.exp_decay_rate)
for name, param in model.named_parameters():
if param.requires_grad:
ema.register(name, param.data)
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = optim.Adadelta(parameters, lr=args.learning_rate)
criterion = nn.CrossEntropyLoss()
writer = SummaryWriter(log_dir='runs/' + args.model_time)
model.train()
loss, last_epoch = 0, -1
max_dev_exact, max_dev_f1 = -1, -1
print('totally {} epoch'.format(args.epoch))
sys.stdout.flush()
iterator = data.train_iter
iterator.repeat = True
for i, batch in enumerate(iterator):
present_epoch = int(iterator.epoch)
if present_epoch == args.epoch:
print('present_epoch value:',present_epoch)
break
if present_epoch > last_epoch:
print('epoch:', present_epoch + 1)
last_epoch = present_epoch
p1, p2 = model(batch.c_char,batch.q_char,batch.c_word[0],batch.q_word[0],batch.c_word[1],batch.q_word[1])
optimizer.zero_grad()
batch_loss = criterion(p1, batch.s_idx) + criterion(p2, batch.e_idx)
loss += batch_loss.item()
batch_loss.backward()
optimizer.step()
for name, param in model.named_parameters():
if param.requires_grad:
ema.update(name, param.data)
if (i + 1) % args.print_freq == 0:
dev_loss, dev_exact, dev_f1, dev_hasans_exact, dev_hasans_f1, dev_noans_exact,dev_noans_f1 = test(model, ema, args, data)
c = (i + 1) // args.print_freq
writer.add_scalar('loss/train', loss, c)
writer.add_scalar('loss/dev', dev_loss, c)
writer.add_scalar('exact_match/dev', dev_exact, c)
writer.add_scalar('f1/dev', dev_f1, c)
print(f'train loss: {loss:.3f} / dev loss: {dev_loss:.3f}'
f' / dev EM: {dev_exact:.3f} / dev F1: {dev_f1:.3f}'
f' / dev hasans EM: {dev_hasans_exact} / dev hasans F1: {dev_hasans_f1}'
f' / dev noans EM: {dev_noans_exact} / dev noans F1: {dev_noans_f1}')
if dev_f1 > max_dev_f1:
max_dev_f1 = dev_f1
max_dev_exact = dev_exact
best_model = copy.deepcopy(model)
loss = 0
model.train()
sys.stdout.flush()
writer.close()
args.max_f1 = max_dev_f1
print(f'max dev EM: {max_dev_exact:.3f} / max dev F1: {max_dev_f1:.3f}')
return best_model
def predict():
parser = argparse.ArgumentParser()
parser.add_argument('--char-dim', default=8, type=int)
parser.add_argument('--char-channel-width', default=5, type=int)
parser.add_argument('--char-channel-size', default=100, type=int)
parser.add_argument('--context-threshold', default=400, type=int)
parser.add_argument('--dev-batch-size', default=100, type=int)
parser.add_argument('--test-batch-size', default=100, type=int)
parser.add_argument('--dev-file', default='dev-v1.1.json')
parser.add_argument('--test-file', default='test1.json')
parser.add_argument('--dropout', default=0.2, type=float)
parser.add_argument('--epoch', default=12, type=int)
parser.add_argument('--exp-decay-rate', default=0.999, type=float)
parser.add_argument('--gpu', default=0, type=int)
parser.add_argument('--hidden-size', default=200, type=int)
parser.add_argument('--learning-rate', default=0.5, type=float)
parser.add_argument('--print-freq', default=250, type=int)
parser.add_argument('--train-batch-size', default=60, type=int)
parser.add_argument('--train-file', default='train-v1.1.json')
parser.add_argument('--word-dim', default=300, type=int)
args = parser.parse_args()
print('loading SQuAD data...')
current_dir = os.getcwd()
current_dir = os.path.join(current_dir, 'BiDAF')
path = os.path.join(current_dir, 'testing_files')
data = SQuAD(args, path)
setattr(args, 'char_vocab_size', len(data.CHAR.vocab))
setattr(args, 'word_vocab_size', len(data.WORD.vocab))
setattr(args, 'dataset_file', f'testing_files/{args.dev_file}')
setattr(args, 'prediction_file', f'output/prediction{time.time()}.out')
setattr(args, 'model_time', strftime('%H:%M:%S', gmtime()))
print('data loading complete!')
device = torch.device(
f"cuda:{args.gpu}" if torch.cuda.is_available() else "cpu")
model = BiDAF(args, data.WORD.vocab.vectors).to(device)
# load trained the parameters to the model
model_path = os.path.join(current_dir, 'saved_models')
model.load_state_dict(
torch.load(os.path.join(model_path, 'BiDAF_tl_new.pt')))
model.eval() # show the model result
answers = dict() # answers to be saved in dictionary format
with torch.set_grad_enabled(False):
for batch in iter(data.test_iter):
# print(batch)
p1, p2 = model(batch)
# batch_loss = criterion(p1, batch.s_idx) + criterion(p2, batch.e_idx)
# loss += batch_loss.item()
batch_size, c_len = p1.size()
ls = nn.LogSoftmax(dim=1)
mask = (torch.ones(c_len, c_len) *
float('-inf')).to(device).tril(-1).unsqueeze(0).expand(
batch_size, -1, -1)
score = (ls(p1).unsqueeze(2) + ls(p2).unsqueeze(1)) + mask
score, s_idx = score.max(dim=1)
score, e_idx = score.max(dim=1)
s_idx = torch.gather(s_idx, 1, e_idx.view(-1, 1)).squeeze()
for i in range(batch_size):
id = batch.id[i]
answer = batch.c_word[0][i][s_idx[i]:e_idx[i] + 1]
answer = ' '.join(
[data.WORD.vocab.itos[idx] for idx in answer])
answers[id] = answer
print(answers)
return answers
def train(args, data):
device = torch.device(
f"cuda:{args.gpu}" if torch.cuda.is_available() else "cpu")
model = BiDAF(args, data.WORD.vocab.vectors).to(device)
ema = EMA(args.exp_decay_rate)
for name, param in model.named_parameters():
if param.requires_grad:
ema.register(name, param.data)
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = optim.Adadelta(parameters, lr=args.learning_rate)
criterion = nn.CrossEntropyLoss()
writer = SummaryWriter(log_dir='runs/' + args.model_time)
model.train()
loss, last_epoch = 0, -1
max_dev_exact, max_dev_f1 = -1, -1
iterator = data.train_iter
num_batch = len(iterator)
for present_epoch in range(args.epoch):
print('epoch', present_epoch + 1)
for i, batch in enumerate(iterator):
# present_epoch = int(iterator.epoch)
"""
if present_epoch == args.epoch:
print(present_epoch)
print()
print(args.epoch)
break
if present_epoch > last_epoch:
print('epoch:', present_epoch + 1)
last_epoch = present_epoch
"""
p1, p2 = model(batch)
optimizer.zero_grad()
"""
print(p1)
print()
print(batch.s_idx)
"""
if len(p1.size()) == 1:
p1 = p1.reshape(1, -1)
if len(p2.size()) == 1:
p2 = p2.reshape(1, -1)
batch_loss = criterion(p1, batch.s_idx) + criterion(
p2, batch.e_idx)
loss += batch_loss.item()
batch_loss.backward()
optimizer.step()
for name, param in model.named_parameters():
if param.requires_grad:
ema.update(name, param.data)
best_model = copy.deepcopy(model)
if i + 1 == num_batch:
dev_loss, dev_exact, dev_f1 = test(model, ema, args, data)
c = (i + 1) // args.print_freq
writer.add_scalar('loss/train', loss / num_batch, c)
writer.add_scalar('loss/dev', dev_loss, c)
writer.add_scalar('exact_match/dev', dev_exact, c)
writer.add_scalar('f1/dev', dev_f1, c)
print(
f'train loss: {loss/num_batch:.3f} / dev loss: {dev_loss:.3f}'
f' / dev EM: {dev_exact:.3f} / dev F1: {dev_f1:.3f}')
if dev_f1 > max_dev_f1:
max_dev_f1 = dev_f1
max_dev_exact = dev_exact
best_model = copy.deepcopy(model)
loss = 0
model.train()
writer.close()
print(f'max dev EM: {max_dev_exact:.3f} / max dev F1: {max_dev_f1:.3f}')
return best_model
def train(args, data):
device = torch.device(
f"cuda:{args.gpu}" if torch.cuda.is_available() else "cpu")
model = BiDAF(args).to(device)
D_batch = args.train_batch_size
ema = EMA(args.exp_decay_rate)
for name, param in model.named_parameters():
if param.requires_grad:
ema.register(name, param.data)
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = optim.Adadelta(parameters, lr=args.learning_rate)
criterion = nn.CrossEntropyLoss()
# writer = SummaryWriter(log_dir='runs/' + args.model_time)
model.train()
loss, last_epoch = 0, -1
max_dev_exact, max_dev_f1 = -1, -1
i = 0
# iterator = data.train_iter
while i + D_batch < len(data.data):
b_id = i
e_id = i + D_batch
# present_epoch = int(iterator.epoch)
# if present_epoch == args.epoch:
# break
# if present_epoch > last_epoch:
# print('epoch:', present_epoch + 1)
# last_epoch = present_epoch
p1, p2 = model(data, b_id, e_id)
optimizer.zero_grad()
s_idx, e_idx = data.get_targ(b_id, e_id)
batch_loss = criterion(p1, s_idx) + criterion(p2, e_idx)
loss += batch_loss.item()
batch_loss.backward()
optimizer.step()
for name, param in model.named_parameters():
if param.requires_grad:
ema.update(name, param.data)
# if (i + 1) % args.print_freq == 0:
# dev_loss, dev_exact, dev_f1 = test(model, ema, args, data)
# c = (i + 1) // args.print_freq
# # writer.add_scalar('loss/train', loss, c)
# # writer.add_scalar('loss/dev', dev_loss, c)
# # writer.add_scalar('exact_match/dev', dev_exact, c)
# # writer.add_scalar('f1/dev', dev_f1, c)
# # print(f'train loss: {loss:.3f} / dev loss: {dev_loss:.3f}'
# # f' / dev EM: {dev_exact:.3f} / dev F1: {dev_f1:.3f}')
# if dev_f1 > max_dev_f1:
# max_dev_f1 = dev_f1
# max_dev_exact = dev_exact
# best_model = copy.deepcopy(model)
# loss = 0
# model.train()
i += D_batch
# writer.close()
print(f'max dev EM: {max_dev_exact:.3f} / max dev F1: {max_dev_f1:.3f}')
return best_model
def train(args, data):
device = torch.device("cuda:{}".format(args.gpu) if torch.cuda.is_available() else "cpu")
model = BiDAF(args).to(device)
# 如果载入的这些参数中，有些参数不要求被更新，即固定不变，不参与训练，需要手动设置这些参数的梯度属性为Fasle，
# 并且在optimizer传参时筛选掉这些参数：
# ema = EMA(args.exp_decay_rate)
# for name, param in model.named_parameters():
# if param.requires_grad:
# ema.register(name, param.data)
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = optim.Adam(parameters, lr=args.learning_rate)
criterion = nn.CrossEntropyLoss()
writer = SummaryWriter(logdir='runs/' + args.model_time)
model.train()
loss, last_epoch = 0, -1
max_dev_exact, max_dev_f1 = -1, -1
iterator = data.train_iter
for i, batch in enumerate(iterator):
present_epoch = int(iterator.epoch)
if present_epoch == args.epoch:
break
if present_epoch > last_epoch:
print('epoch:', present_epoch + 1)
last_epoch = present_epoch
try:
p1, p2 = model(batch)
except OSError:
pass
optimizer.zero_grad()
batch_loss = criterion(p1, batch.s_idx) + criterion(p2, batch.e_idx)
loss = batch_loss.item()
batch_loss.backward()
optimizer.step()
print("loss", loss)
# for name, param in model.named_parameters():
# if param.requires_grad:
# ema.update(name, param.data)
# if (i + 1) % args.print_freq == 0:
# dev_loss, dev_exact, dev_f1 = test(model, args, data)
# c = (i + 1) // args.print_freq
# writer.add_scalar('loss/train', loss, c)
# writer.add_scalar('loss/dev', dev_loss, c)
# writer.add_scalar('exact_match/dev', dev_exact, c)
# writer.add_scalar('f1/dev', dev_f1, c)
# print('train loss: {:.3f} / dev loss: {:.3f} dev EM: {:.3f} dev F1: {:.3f}'.format(loss, dev_loss, dev_exact, dev_f1))
# if dev_f1 > max_dev_f1:
# max_dev_f1 = dev_f1
# max_dev_exact = dev_exact
# best_model = copy.deepcopy(model)
# loss = 0
# model.train()
# writer.close()
# print('max dev EM: {:.3f} / max dev F1: {:.3f}'.format(max_dev_exact, max_dev_f1))
return best_model
parser.add_argument('--context-threshold', default=400, type=int)
parser.add_argument('--dev-batch-size', default=100, type=int)
parser.add_argument('--dev-file', default='dev-v1.1.json')
parser.add_argument('--dropout', default=0.2, type=float)
parser.add_argument('--epoch', default=12, type=int)
parser.add_argument('--exp-decay-rate', default=0.999, type=float)
parser.add_argument('--gpu', default=0, type=int)
parser.add_argument('--hidden-size', default=100, type=int)
parser.add_argument('--learning-rate', default=0.5, type=float)
parser.add_argument('--print-freq', default=250, type=int)
parser.add_argument('--train-batch-size', default=60, type=int)
parser.add_argument('--train-file', default='train-v1.1.json')
parser.add_argument('--word-dim', default=100, type=int)
args = parser.parse_args()
print('loading SQuAD data...')
data = SQuAD(args)
setattr(args, 'char_vocab_size', len(data.CHAR.vocab))
setattr(args, 'word_vocab_size', len(data.WORD.vocab))
setattr(args, 'dataset_file', f'.data/squad/{args.dev_file}')
setattr(args, 'prediction_file', f'prediction{args.gpu}.out')
setattr(args, 'model_time', strftime('%H:%M:%S', gmtime()))
print('data loading complete!')
device = torch.device(f"cuda:{args.gpu}" if torch.cuda.is_available() else "cpu")
model = BiDAF(args, data.WORD.vocab.vectors).to(device)
# load trained the parameters to the model
# model.load_state_dict(torch.load(r'saved_models/BiDAF_08:03:17.pt'))
model.eval() # show the model result
print(model)