def run_epoch(data_loader, train_model, model, gen, optimizer, step, args):
'''
Train model for one pass of train data, and return loss, acccuracy
'''
eval_model = not train_model
data_iter = data_loader.__iter__()
losses = []
obj_losses = []
k_selection_losses = []
k_continuity_losses = []
preds = []
probas = []
golds = []
losses = []
texts = []
rationales = []
if train_model:
model.train()
gen.train()
else:
gen.eval()
model.eval()
num_batches_per_epoch = len(data_iter)
if train_model:
num_batches_per_epoch = min(len(data_iter), 10000)
for _ in tqdm.tqdm(range(num_batches_per_epoch)):
batch = data_iter.next()
if train_model:
step += 1
if step % 100 == 0 or args.debug_mode:
args.gumbel_temprature = max(
np.exp((step + 1) * -1 * args.gumbel_decay), .05)
x_indx = learn.get_x_indx(batch, args, eval_model)
text = batch['text']
y = autograd.Variable(batch['y'], volatile=eval_model)
if args.cuda:
x_indx, y = x_indx.cuda(), y.cuda()
if train_model:
optimizer.zero_grad()
if args.get_rationales:
mask, z = gen(x_indx)
else:
mask = None
logit, _ = model(x_indx, mask=mask)
if args.use_as_tagger:
logit = logit.view(-1, 2)
y = y.view(-1)
loss = get_loss(logit, y, args)
obj_loss = loss
if args.get_rationales:
selection_cost, continuity_cost = gen.loss(mask, x_indx)
loss += args.selection_lambda * selection_cost
loss += args.continuity_lambda * continuity_cost
if train_model:
loss.backward()
optimizer.step()
if args.get_rationales:
k_selection_losses.append(generic.tensor_to_numpy(selection_cost))
k_continuity_losses.append(
generic.tensor_to_numpy(continuity_cost))
obj_losses.append(generic.tensor_to_numpy(obj_loss))
losses.append(generic.tensor_to_numpy(loss))
batch_softmax = F.softmax(logit, dim=-1).cpu()
max_preds = torch.max(batch_softmax, 1)
probas.extend(max_preds[0].view(y.size()).data.numpy())
preds.extend(max_preds[1].view(y.size()).data.numpy())
texts.extend(text)
rationales.extend(learn.get_rationales(mask, text))
if args.use_as_tagger:
golds.extend(batch['y'].view(-1).numpy())
else:
golds.extend(batch['y'].numpy())
epoch_metrics = metrics.get_metrics(preds, golds, args)
epoch_stat = {'loss': np.mean(losses), 'obj_loss': np.mean(obj_losses)}
for metric_k in epoch_metrics.keys():
epoch_stat[metric_k] = epoch_metrics[metric_k]
if args.get_rationales:
epoch_stat['k_selection_loss'] = np.mean(k_selection_losses)
epoch_stat['k_continuity_loss'] = np.mean(k_continuity_losses)
return epoch_stat, step, losses, preds, golds, rationales, probas
def run_epoch(data_loader, train_model, model, gen, optimizer, step, args):
'''
Train model for one pass of train data, and return loss, acccuracy
'''
eval_model = not train_model
data_iter = data_loader.__iter__()
losses = []
obj_losses = []
k_selection_losses = []
k_continuity_losses = []
preds = []
golds = []
losses = []
if train_model:
model.train()
gen.train()
else:
gen.eval()
model.eval()
num_batches_per_epoch = len(data_iter)
if train_model:
num_batches_per_epoch = min(len(data_iter), 10000)
for _ in tqdm.tqdm(range(num_batches_per_epoch)):
batch = data_iter.next()
if train_model:
step += 1
if step % 100 == 0 or args.debug_mode:
args.gumbel_temprature = max( np.exp((step+1) *-1* args.gumbel_decay), .05)
x_indx = utils.get_x_indx(batch, args, eval_model)
text = batch['text']
indices = batch['i']
y = autograd.Variable(batch['y'], volatile=eval_model)
if args.cuda:
x_indx, y = x_indx.cuda(), y.cuda()
if train_model:
optimizer.zero_grad()
if args.get_rationales:
mask, z = gen(x_indx)
else:
mask = None
logit, _ = model(x_indx, mask=mask)
loss = get_loss(logit, y, args)
obj_loss = loss
if args.get_rationales:
selection_cost, continuity_cost = gen.loss(mask, x_indx)
loss += args.selection_lambda* selection_cost
loss += args.continuity_lambda* continuity_cost
if train_model:
loss.backward()
optimizer.step()
if args.get_rationales:
k_selection_losses.append( generic.tensor_to_numpy(selection_cost))
k_continuity_losses.append( generic.tensor_to_numpy(continuity_cost))
obj_losses.append(generic.tensor_to_numpy(obj_loss))
losses.append( generic.tensor_to_numpy(loss) )
preds.extend(
torch.max(logit.data,
1)[1].view(y.size()).cpu().numpy()) # Record predictions
golds.extend(batch['y'].numpy())
if args.objective in ['cross_entropy', 'margin']:
metric = sklearn.metrics.accuracy_score(y_true=golds, y_pred=preds)
confusion_matrix = sklearn.metrics.confusion_matrix(y_true=golds,y_pred=preds)
elif args.objective == 'mse':
metric = sklearn.metrics.mean_squared_error(y_true=golds, y_pred=preds)
confusion_matrix = "NA"
epoch_stat = {
'loss' : np.mean(losses),
'obj_loss': np.mean(obj_losses),
'metric':metric,
'confusion_matrix': confusion_matrix
}
if args.get_rationales:
epoch_stat['k_selection_loss'] = np.mean(k_selection_losses)
epoch_stat['k_continuity_loss'] = np.mean(k_continuity_losses)
return epoch_stat, step, losses, preds, golds