def dev(use_gpu, batch_size, model_path):
dataset = build_dataset_pytorch.QuoraDataset_train(
'data/dev_data_pad.pkl', 'data/dev_data_char_pad.pkl',
'data/dev_data_flags_pad.pkl')
dataloader = DataLoader(dataset, batch_size, shuffle=False, num_workers=0)
mm1 = mp.MatchModel(vocab_size=100000,
tr_w_embed_dim=300,
tr_w_emb_dim_flag=100,
char_size=50,
char_embed_dim=20,
output_dim=100,
hidden_dim=200,
use_gpu=use_gpu,
yt_layer_num=3,
mode='test')
mm1.load_state_dict(t.load(model_path, map_location='cpu'))
mm1.eval()
with open('data/vector_n.pkl', 'rb') as handle:
data = pickle.load(
handle
) # Warning: If adding something here, also modifying saveDataset
word_vec = t.from_numpy(np.array(data['word_vec']))
if use_gpu: mm1.cuda()
def data_convert(*data):
datas = []
for i in data:
datas.append(i.cuda())
return datas
data_num = 0
correct_all = 0
for i, data in enumerate(dataloader):
if i == 5:
p_inputs, q_inputs, targets, p_inputs_char, q_inputs_char, p_flags, q_flags = data
print(t.nonzero(p_inputs).shape[0], t.nonzero(q_inputs).shape[0])
if use_gpu:
[p_inputs, q_inputs, targets, p_inputs_char, q_inputs_char, p_flags, q_flags, word_vec] = \
data_convert(p_inputs, q_inputs, targets, p_inputs_char, q_inputs_char, p_flags, q_flags, word_vec)
p_input = [p_inputs, p_inputs_char, word_vec, p_flags]
q_input = [q_inputs, q_inputs_char, q_flags]
out_p, out_q, cos = mm1(100, p_input, q_input)
vis = visdom.Visdom(env=u'test1')
for x in range(len(out_p)):
# print('\n%d:'%(x), cos[x].detach().numpy())
# vis.image(out_p[x].detach().numpy()[:,:400], win='p_layer_%d_%d'%(x, 0), opts={'title':'p_layer%d'%(x)})
# vis.image(out_q[x].detach().numpy()[:,:400], win='q_layer_%d_%d'%(x, 0), opts={'title':'q_layer%d'%(x)})
vis.image(cos[x].detach().numpy(),
win='q_layer_%d_%d' % (x, 0),
opts={'title': 'cos_layer%d' % (x)})
break
def dev(use_gpu, batch_size, model_path):
dataset = build_dataset_pytorch.QuoraDataset_train(
'data/dev_data_pad.pkl', 'data/dev_data_char_pad.pkl',
'data/dev_data_flags_pad.pkl')
dataloader = DataLoader(dataset, batch_size, shuffle=True, num_workers=1)
mm1 = mp.MatchModel(vocab_size=100000,
tr_w_embed_dim=300,
tr_w_emb_dim_flag=0,
char_size=50,
char_embed_dim=20,
output_dim=100,
hidden_dim=200,
use_gpu=use_gpu,
yt_layer_num=3,
mode='test')
mm1.load_state_dict(t.load(model_path))
mm1.eval()
with open('data/vector_n.pkl', 'rb') as handle:
data = pickle.load(
handle
) # Warning: If adding something here, also modifying saveDataset
word_vec = t.from_numpy(np.array(data['word_vec']))
if use_gpu: mm1.cuda()
def data_convert(*data):
datas = []
for i in data:
datas.append(i.cuda())
return datas
data_num = 0
correct_all = 0
for i, data in enumerate(dataloader):
p_inputs, q_inputs, targets, p_inputs_char, q_inputs_char, p_flags, q_flags = data
if use_gpu:
[p_inputs, q_inputs, targets, p_inputs_char, q_inputs_char, p_flags, q_flags, word_vec] = \
data_convert(p_inputs, q_inputs, targets, p_inputs_char, q_inputs_char, p_flags, q_flags, word_vec)
p_input = [p_inputs, p_inputs_char, word_vec, p_flags]
q_input = [q_inputs, q_inputs_char, q_flags]
output = mm1(100, p_input, q_input)
if use_gpu:
output = output.cpu()
targets = targets.cpu()
# print(output, targets)
correct_num = (t.argmax(output, 1) == t.argmax(targets, 1)).numpy()
correct_sum = correct_num.sum()
data_num = data_num + len(correct_num)
correct_all = correct_all + correct_sum
print("-- Dev Acc %.3f" % (correct_all / data_num))
return correct_all / data_num
def train(use_gpu, model_path, lr, epochs, model_prefix, batch_size):
dataset = build_dataset_pytorch.QuoraDataset_train(
'data/train_data_pad.pkl', 'data/train_data_char_pad.pkl',
'data/train_data_flags_pad.pkl')
dataloader = DataLoader(dataset, batch_size, shuffle=True, num_workers=1)
mm = mp.MatchModel(vocab_size=100000,
tr_w_embed_dim=300,
tr_w_emb_dim_flag=0,
char_size=50,
char_embed_dim=20,
output_dim=100,
hidden_dim=200,
use_gpu=use_gpu,
yt_layer_num=3)
# for name, param in mm.named_parameters():
# print(name, param.size())
optimizer = t.optim.Adam(mm.parameters(), lr=lr)
criterion = nn.CrossEntropyLoss()
t.manual_seed(2)
if model_path:
mm.load_state_dict(t.load(model_path))
if use_gpu:
# device_ids = [1,2]
mm.cuda()
# mm = nn.DataParallel(mm, device_ids = device_ids)
# optimizer = nn.DataParallel(optimizer, device_ids = device_ids)
criterion.cuda()
t.cuda.manual_seed_all(2)
# loss_meter = meter.AverageValueMeter()
with open('data/vector_n.pkl', 'rb') as handle:
data = pickle.load(
handle
) # Warning: If adding something here, also modifying saveDataset
word_vec = t.from_numpy(np.array(data['word_vec']))
def data_convert(*data):
datas = []
for i in data:
datas.append(i.cuda())
return datas
max_dev = 0
max_test = 0
for e in range(0, epochs):
correct_all = 0
data_num = 0
# loss_meter.reset()
print("----- Epoch {}/{} -----".format(e + 1, epochs))
# optimizer.zero_grad()
for i, data_ in tqdm(enumerate(dataloader), desc="Training"):
p_inputs, q_inputs, targets, p_inputs_char, q_inputs_char, p_flags, q_flags = data_
if use_gpu:
[p_inputs, q_inputs, targets, p_inputs_char, q_inputs_char, p_flags, q_flags, word_vec] = \
data_convert(p_inputs, q_inputs, targets, p_inputs_char, q_inputs_char, p_flags, q_flags, word_vec)
optimizer.zero_grad()
p_input = [p_inputs, p_inputs_char, word_vec, p_flags]
q_input = [q_inputs, q_inputs_char, q_flags]
output = mm(100, p_input, q_input)
loss = criterion(output, t.argmax(targets, 1))
# print(loss)
# print(output)
if use_gpu:
output = output.cpu()
targets = targets.cpu()
correct_num = (t.argmax(output, 1) == t.argmax(targets, 1)).numpy()
correct_sum = correct_num.sum()
# print('correct_sum:', correct_num)
accuracy = correct_sum / len(correct_num)
data_num = data_num + len(correct_num)
correct_all = correct_all + correct_sum
loss.backward()
optimizer.step()
# loss_meter.add(loss.data[0])
if (i + 1) % 600 == 0:
tqdm.write("----- Epoch %d Batch %d -- Loss %.2f -- Acc %.3f" %
(e + 1, i + 1, loss, accuracy))
tqdm.write("----- epoch %d -- Loss %.2f -- Acc %.3f" %
(e + 1, loss, correct_all / data_num))
t.save(mm.state_dict(), '%s_%s.pth' % (model_prefix, e))
dev_acc = dev(True, 50, '%s_%s.pth' % (model_prefix, e))
test_acc = test(True, 50, '%s_%s.pth' % (model_prefix, e))
max_dev = max(max_dev, dev_acc)
max_test = max(max_test, test_acc)
print('max test acc:', max_test, '\nmax dev acc:', max_dev)
