def evaluate(model, data_loader, global_stats, mode='train'):
# Use precision for classify
eval_time = util.Timer()
start_acc = util.AverageMeter()
# Make predictions
examples = 0
for ex in data_loader:
batch_size = ex[0].size(0)
pred_s = model.predict(ex)
answer = ex[5]
# We get metrics for independent start/end and joint start/end
start_acc.update(Evaluate.accuracies(pred_s, answer.cpu().data.numpy()), 1)
# If getting train accuracies, sample max 10k
examples += batch_size
if mode == 'train' and examples >= 1e4:
break
logger.info('%s valid unofficial use Accuracy: Epoch = %d | acc = %.2f | ' %
(mode, global_stats['epoch'], start_acc.avg) +
' = %d | ' %
(examples) +
'valid time = %.2f (s)' % eval_time.time())
return {'acc': start_acc.avg}
def main(args):
# data, word2ids = util.load_train_data(train_file, word2vec_file)
data, word2ids, embed_arr = util.load_data()
feature_dict = util.build_feature_dict(args, data)
model = init_model(words_dict=word2ids, feature_dict=feature_dict, args=args)
model.quick_load_embed(embed_arr)
data_loader = make_dataset(data, model)
start_epoch = 0
# TRAIN/VALID LOOP
logger.info('-' * 100)
logger.info('Train now! Output loss every %d batch...' % args.display_iter)
stats = {'timer': util.Timer(), 'epoch': 0, 'best_valid': 0}
for epoch in range(start_epoch, args.num_epochs):
stats['epoch'] = epoch
train(args, data_loader, model, stats)
result = evaluate(model, data_loader, global_stats=stats)
if result[args.valid_metric] > stats['best_valid']:
logger.info('Best valid: %s = %.2f (epoch %d, %d updates)' %
(args.valid_metric, result[args.valid_metric],
stats['epoch'], model.updates))
model.save(args.model_file)
stats['best_valid'] = result[args.valid_metric]
def train(args, data_loader, model, global_stats):
"""Run through one epoch of model training with the provided data loader."""
# Initialize meters + timers
train_loss = util.AverageMeter()
epoch_time = util.Timer()
# Run one epoch
for idx, ex in enumerate(data_loader):
train_loss.update(*model.update(ex)) # run on one batch
if idx % args.display_iter == 0:
logger.info('train: Epoch = %d | iter = %d/%d | ' %
(global_stats['epoch'], idx, len(data_loader)) +
'loss = %.2f | elapsed time = %.2f (s)' %
(train_loss.avg, global_stats['timer'].time()))
train_loss.reset()
logger.info('train: Epoch %d done. Time for epoch = %.2f (s)' %
(global_stats['epoch'], epoch_time.time()))
# Checkpoint
if args.checkpoint:
model.checkpoint(args.model_file + '.checkpoint',
global_stats['epoch'] + 1)