def train_epoch(self, epoch_idx, model, train_batches):
# Turn on training mode which enables dropout and BNorm.
model.train()
np.random.shuffle(train_batches) # shuffling data
running_losses = []
epoch_losses = []
num_train_batches = len(train_batches)
bar = create_progress_bar('train_loss')
for batch_idx in bar(range(num_train_batches)):
self.optimizer.zero_grad()
loss_var = self.train_step(model, train_batches[batch_idx])
loss_data = loss_var.data[0]
loss_var.backward() # compute gradients
self.optimizer.step() # update weights
running_losses = ([loss_data] + running_losses)[:20]
bar.dynamic_messages['train_loss'] = np.mean(running_losses)
epoch_losses.append(loss_data)
epoch_loss_avg = np.mean(epoch_losses)
return epoch_loss_avg
def draw_prediction_errors(bad_dgs, tgt_dir):
if not os.path.exists(tgt_dir):
os.makedirs(tgt_dir)
num_items = len(bad_dgs)
logger.info('Drawing %d graphs with errors in --> %s', len(bad_dgs),
tgt_dir)
bar = create_progress_bar()
# for idx in bar(range(num_items // 100)):
for idx in bar(range(num_items)):
dg = bad_dgs[idx]
error_data = dg.graph['errors']
if not error_data is None:
for child, head, target in error_data[:2]:
nx.set_node_attributes(dg, values={child: 'red'}, name='color')
nx.set_node_attributes(dg, values={head: 'blue'}, name='color')
# nx.set_node_attributes(dg, values={head: 'setlinewidth(3)'}, name='style')
# if 'lo' in dg.node[child]:
# nx.set_node_attributes(dg, values={dg.node[child]['lo']: 'green'}, name='color')
#
# if 'ro' in dg.node[child]:
# nx.set_node_attributes(dg, values={dg.node[child]['ro']: 'green'}, name='color')
png_fname = os.path.join(tgt_dir, '%d.png' % idx)
gold_snt = dg.graph['snt']
pred_snt = ' '.join(
[dg.node[n]['GOLD_FORM'] for n in dg.graph['node_order']])
draw_graph(dg,
label='%s\n%s' % (gold_snt, pred_snt),
layout='dot',
fname=png_fname,
rename_nodes=True)
logger.info('Done')
def compute_val_loss(self, model, dev_batches):
total_loss = 0
running_losses = []
num_dev_batches = len(dev_batches)
bar = create_progress_bar('dev_loss')
for batch_idx in bar(range(num_dev_batches)):
loss_var = self.train_step(model, dev_batches[batch_idx])
loss_data = loss_var.data[0]
# Record loss
running_losses = ([loss_data] + running_losses)[:20]
bar.dynamic_messages['dev_loss'] = np.mean(running_losses)
total_loss += loss_data
total_loss_avg = total_loss / num_dev_batches
return total_loss_avg
