def test(data_loader, model, device):
# set model in validating mode.
model.eval()
# model.to(device)
correct = 0
n = 0
for (batch_x, batch_len, batch_y, batch_x_tag) in data_loader:
n += len(batch_y)
bx, bx_len, bx_mask, by, bx_tag = prepare_input(
batch_x, batch_len, batch_y, batch_x_tag)
batch_x = bx.to(device)
batch_len = bx_len.to(device)
batch_y = by.to(device)
batch_max = bx_mask.to(device)
try:
output, _ = model(batch_x, batch_len, batch_max)
# get the index of the max log-probability
_, y_p = output.max(1)
correct += (y_p == batch_y).sum().item()
except:
print("in test ..... ")
print(batch_x.shape)
print(batch_len.shape)
print(batch_y.shape)
input()
pass
eval_acc = correct / n
# set model back to training mode.
model.train()
return eval_acc
def test(data_loader, model, device):
# set model in validating mode.
model.eval()
# model.to(device)
correct = 0
n = 0
for batch_x, batch_len, batch_y in data_loader:
n += len(batch_y)
bx, bx_len, bx_mask, by = prepare_input(batch_x, batch_len,
batch_y)
batch_x = bx.to(device)
batch_x_mask = bx_mask.to(device)
batch_len = bx_len.to(device)
batch_y = by.to(device)
output = model(batch_x, batch_x_mask, batch_len)
# get the index of the max log-probability
y_p = output.max(1)[1]
correct += (y_p == batch_y).sum().item()
eval_acc = correct / n
# set model back to training mode.
model.train()
return eval_acc
def train_epoch(self, data_loader, report_step=100):
stop_next_epoch = False
train_loss = 0.
train_acc = 0.
n_sample = 0.
for step, (batch_x, batch_len, batch_y,
batch_x_tag) in enumerate(data_loader):
n_sample += len(batch_y)
bx, bx_len, bx_mask, by, bx_tag = prepare_input(
batch_x, batch_len, batch_y, batch_x_tag)
batch_x = bx.to(self.device)
batch_x_mask = bx_mask.to(self.device)
batch_len = bx_len.to(self.device)
batch_y = by.to(self.device)
batch_x_tag = bx_tag.to(self.device)
self.pred, tags_logits = self.model(batch_x, batch_len,
batch_x_mask)
# Step 4. Compute the loss, gradients, and update the parameters by
# calling optimizer.step()
loss = self.criterion(tags_logits, batch_x_tag, self.pred, batch_y)
self.optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.parameters, max_norm=self.clip)
self.optimizer.step()
train_loss += loss.item()
y_p = self.pred.max(dim=1)[1]
train_correct = (y_p == batch_y).sum().item()
train_acc += train_correct
if (step + 1) % report_step == 0:
train_loss /= n_sample
train_acc /= n_sample
if train_loss < 1.e-4:
stop_next_epoch = True
print("[step_%05d] loss: %0.4f, acc: %0.4f" %
(step + 1, train_loss, train_acc))
train_loss = 0.
train_acc = 0.
n_sample = 0.
return stop_next_epoch
def get_attention_scores(data_loader, model, device):
# set model in validating mode.
model.eval()
# model.to(device)
scores = []
data_lens = []
n = 0
for (batch_x, batch_len, batch_y, batch_x_tag) in data_loader:
n += len(batch_y)
bx, bx_len, bx_mask, by, bx_tag = prepare_input(
batch_x, batch_len, batch_y, batch_x_tag)
batch_x = bx.to(device)
batch_len = bx_len.to(device)
batch_max = bx_mask.to(device)
batch_scores = model.get_attention(batch_x, batch_len, batch_max)
batch_scores = batch_scores.cpu().detach().numpy()
scores.append(batch_scores)
data_lens.append(bx_len)
return scores, data_lens