def train(model, optimizer, data_iterator, metrics, save_summary_steps,
num_steps, label_encoder):
# set model to training mode
model.train()
# summary for current training loop and a running average object for loss
summ = []
loss_avg = utils.RunningAverage()
preds_all = []
true_all = []
loss_all = []
# Use tqdm for progress bar
t = trange(num_steps)
for i in t:
model.train()
# zero the gradients
model.zero_grad()
# fetch the next training batch
train_batch, labels_batch = next(data_iterator)
# compute model output and log_lik
loss, logits = model(
input=train_batch,
labels=labels_batch,
label_pad_idx=label_encoder[ClassEncoder.FEATURE_NAME].pad_idx)
# clear previous gradients, compute gradients of all variables wrt loss
loss.backward()
# performs updates using calculated gradients
optimizer.step()
max_norm(model)
# Evaluate summaries only once in a while
if i % save_summary_steps == 0:
model.eval()
# 1. get the predictions from model
preds = model.predict(
logits=logits,
mask=(train_batch[WordEncoder.FEATURE_NAME] !=
data_encoder[WordEncoder.FEATURE_NAME].pad_idx).float())
labels_batch = labels_batch[
ClassEncoder.FEATURE_NAME].data.cpu().numpy().squeeze()
# 2. decode the predictions and the ground_truths
labels = label_encoder[ClassEncoder.FEATURE_NAME].decode(
labels_batch)
preds = label_encoder[ClassEncoder.FEATURE_NAME].decode(preds)
# 3. gather stats
preds_all.extend(preds)
true_all.extend(labels)
loss_all.append(loss.item())
# update the average loss
loss_avg.update(loss.item())
t.set_postfix(loss='{:05.3f}'.format(loss_avg()))
# compute mean of all metrics in summary
scores = {
metric: metrics[metric](preds_all, true_all)
for metric in metrics
}
scores['loss'] = np.mean(loss_all)
metrics_string = " ; ".join("{}: {:05.3f}".format(k, v)
for k, v in scores.items())
logging.info("- Train metrics: " + metrics_string)
return scores
def train(model, optimizer, data_iterator, metrics, params, num_steps, data_encoder, label_encoder):
# summary for current training loop and a running average object for loss
loss_avg = utils.RunningAverage()
preds_all = []
true_all = []
loss_all = []
# Use tqdm for progress bar
t = trange(num_steps)
for i in t:
model.train()
# zero the gradients
model.zero_grad()
# fetch the next training batch
train_batch, labels_batch = next(data_iterator)
# compute model output and log_lik
loss = model(train_X=train_batch,
train_Y=labels_batch,
mode='train')
# clear previous gradients, compute gradients of all variables wrt loss
loss.backward()
# gradient clipping
torch.nn.utils.clip_grad_norm_(model.parameters(), 5.0)
# performs updates using calculated gradients
optimizer.step()
# Evaluate summaries only once in a while
if i % params.save_summary_steps == 0:
model.eval()
# 1. get the predictions from model
preds = model.predict(train_X=train_batch, mode='train')
labels_batch = labels_batch[EntityEncoder.FEATURE_NAME].data.cpu().numpy()
labels_batch_trunc = [labels_batch[o_idx][:len(o)] for o_idx, o in enumerate(preds)]
# 2. decode the predictions and the ground_truths
labels_batch_trunc = [label_encoder[EntityEncoder.FEATURE_NAME].decode(l) for l in labels_batch_trunc]
preds = [label_encoder[EntityEncoder.FEATURE_NAME].decode(l) for l in preds]
# 3. gather stats
preds_all.extend(preds)
true_all.extend(labels_batch_trunc)
loss_all.append(loss.item())
# update the average loss
loss_avg.update(loss.item())
t.set_postfix(loss='{:05.3f}'.format(loss_avg()))
del loss
# compute mean of all metrics in summary
scores = {metric: metrics[metric](preds_all, true_all) for metric in metrics}
scores['loss'] = np.mean(loss_all)
metrics_string = " ; ".join("{}: {}".format(k, v) for k, v in scores.items())
logging.info("- Train metrics: " + metrics_string)
return scores