def evaluate_all(my_arg, pr=True):
emb = LoadEmbedding('res/emb.txt')
print 'finish loading embedding'
# batch_getter = BatchGetter('data/dev', 'GPE_NAM', 1, False)
batch_getter_lst = []
if my_arg == 0:
pernam_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.testa',
'PER', 1, False)
batch_getter_lst.append(pernam_batch_getter)
loc_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.testa',
'LOC', 1, False)
batch_getter_lst.append(loc_batch_getter)
misc_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.testa',
'MISC', 1, False)
batch_getter_lst.append(misc_batch_getter)
org_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.testa',
'ORG', 1, False)
batch_getter_lst.append(org_batch_getter)
if my_arg == 1:
pernam_batch_getter = BatchGetter('data/dev', 'PER_NAM', 1, False)
batch_getter_lst.append(pernam_batch_getter)
fac_batch_getter = BatchGetter('data/dev', 'FAC_NAM', 1, False)
batch_getter_lst.append(fac_batch_getter)
loc_batch_getter = BatchGetter('data/dev', 'LOC_NAM', 1, False)
batch_getter_lst.append(loc_batch_getter)
gpe_batch_getter = BatchGetter('data/dev', 'GPE_NAM', 1, False)
batch_getter_lst.append(gpe_batch_getter)
org_batch_getter = BatchGetter('data/dev', 'ORG_NAM', 1, False)
batch_getter_lst.append(org_batch_getter)
if my_arg == 2:
pernam_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.testa',
'PER', 1, False)
batch_getter_lst.append(pernam_batch_getter)
loc_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.testa',
'LOC', 1, False)
batch_getter_lst.append(loc_batch_getter)
misc_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.testa',
'MISC', 1, False)
batch_getter_lst.append(misc_batch_getter)
org_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.testa',
'ORG', 1, False)
batch_getter_lst.append(org_batch_getter)
batch_getter = MergeBatchGetter(batch_getter_lst, 1, False)
print 'finish loading dev data'
embedding_layer = EmbeddingLayer(emb, 0)
d = embedding_layer.get_out_dim()
att_layer = AttentionFlowLayer(2 * d)
# if my_arg == 2:
model_out_layer = ModelingOutLayer(8 * d, d, 2, 3, 0)
# else:
# model_out_layer = ModelingOutLayer(8*d, d, 2, 2, 0)
model_dir = 'model' + str(my_arg)
embedding_layer.load_state_dict(
torch.load(model_dir + '/embedding_layer.pkl'))
att_layer.load_state_dict(torch.load(model_dir + '/att_layer.pkl'))
model_out_layer.load_state_dict(
torch.load(model_dir + '/model_out_layer.pkl'))
# models = [embedding_layer, att_layer, model_out_layer]
# opts = [emb_opt, att_opt, model_out_opt]
ner_tag = Vocab('res/ner_xx',
unk_id=config['UNK_token'],
pad_id=config['PAD_token'])
# if my_arg == 2:
evaluator = ConllBoundaryPerformance(ner_tag)
# else:
# evaluator = BoundaryPerformance(ner_tag)
evaluator.reset()
if config['USE_CUDA']:
att_layer.cuda(config['cuda_num'])
embedding_layer.cuda(config['cuda_num'])
model_out_layer.cuda(config['cuda_num'])
emb_opt = torch.optim.Adam(embedding_layer.parameters())
att_opt = torch.optim.Adam(att_layer.parameters())
model_out_opt = torch.optim.Adam(model_out_layer.parameters())
out_file = codecs.open('data/eva_result' + str(my_arg),
mode='wb',
encoding='utf-8')
ex_iterations = 0
for iteration, this_batch in enumerate(batch_getter):
target, rec = evaluate_one(ex_iterations + iteration, embedding_layer,
att_layer, model_out_layer, emb_opt,
att_opt, model_out_opt, this_batch)
evaluator.evaluate(iteration,
target.numpy().tolist(),
rec.numpy().tolist(),
out_file,
pr=pr)
if iteration % 100 == 0:
print '{} sentences processed'.format(iteration)
evaluator.get_performance()
return evaluator.get_performance()
def main(my_arg):
logger = Logger('./logs' + str(my_arg))
emb = LoadEmbedding('res/embedding.txt')
print 'finish loading embedding'
# batch_getter = BatchGetter('data/train', 'GPE_NAM', config['batch_size'])
batch_getter_lst = []
if my_arg == 0:
pernam_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.train',
'PER', 1, True)
batch_getter_lst.append(pernam_batch_getter)
loc_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.train',
'LOC', 1, True)
batch_getter_lst.append(loc_batch_getter)
misc_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.train',
'MISC', 1, True)
batch_getter_lst.append(misc_batch_getter)
org_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.train',
'ORG', 1, True)
batch_getter_lst.append(org_batch_getter)
if my_arg == 1:
pernam_batch_getter = BatchGetter('data/train', 'PER_NAM', 1)
batch_getter_lst.append(pernam_batch_getter)
fac_batch_getter = BatchGetter('data/train', 'FAC_NAM', 1)
batch_getter_lst.append(fac_batch_getter)
loc_batch_getter = BatchGetter('data/train', 'LOC_NAM', 1)
batch_getter_lst.append(loc_batch_getter)
gpe_batch_getter = BatchGetter('data/train', 'GPE_NAM', 1)
batch_getter_lst.append(gpe_batch_getter)
# if my_arg == 1:
org_batch_getter = BatchGetter('data/train', 'ORG_NAM', 1)
batch_getter_lst.append(org_batch_getter)
if my_arg == 2:
pernam_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.train',
'PER', 1, True)
batch_getter_lst.append(pernam_batch_getter)
loc_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.train',
'LOC', 1, True)
batch_getter_lst.append(loc_batch_getter)
misc_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.train',
'MISC', 1, True)
batch_getter_lst.append(misc_batch_getter)
org_batch_getter = ConllBatchGetter('data/conll2003/bio_eng.train',
'ORG', 1, True)
batch_getter_lst.append(org_batch_getter)
batch_getter = MergeBatchGetter(batch_getter_lst, config['batch_size'],
True)
print 'finish loading train data'
embedding_layer = EmbeddingLayer(emb)
d = embedding_layer.get_out_dim()
att_layer = AttentionFlowLayer(2 * d)
# if my_arg == 2:
model_out_layer = ModelingOutLayer(8 * d, d, 2, 3)
# else:
# model_out_layer = ModelingOutLayer(8*d, d, 2, 2)
# models = [embedding_layer, att_layer, model_out_layer]
# opts = [emb_opt, att_opt, model_out_opt]
if config['USE_CUDA']:
att_layer.cuda(config['cuda_num'])
embedding_layer.cuda(config['cuda_num'])
model_out_layer.cuda(config['cuda_num'])
emb_opt = torch.optim.Adadelta(embedding_layer.parameters())
att_opt = torch.optim.Adadelta(att_layer.parameters())
model_out_opt = torch.optim.Adadelta(model_out_layer.parameters())
log_file = open('log_file' + str(my_arg), 'w')
f_max = 0
low_epoch = 0
ex_iterations = 0
model_dir = 'model' + str(my_arg)
for epoch in range(config['max_epoch']):
for iteration, this_batch in enumerate(batch_getter):
time0 = time.time()
print 'epoch: {}, iteraton: {}'.format(epoch,
ex_iterations + iteration)
train_iteration(logger, ex_iterations + iteration, embedding_layer,
att_layer, model_out_layer, emb_opt, att_opt,
model_out_opt, this_batch)
time1 = time.time()
print 'this iteration time: ', time1 - time0, '\n'
if (ex_iterations + iteration) % config['save_freq'] == 0:
torch.save(embedding_layer.state_dict(),
model_dir + '/embedding_layer.pkl')
torch.save(att_layer.state_dict(),
model_dir + '/att_layer.pkl')
torch.save(model_out_layer.state_dict(),
model_dir + '/model_out_layer.pkl')
ex_iterations += iteration + 1
batch_getter.reset()
f, p, r = evaluate_all(my_arg, False)
log_file.write('epoch: {} f: {} p: {} r: {}\n'.format(epoch, f, p, r))
log_file.flush()
if f >= f_max:
f_max = f
low_epoch = 0
# torch.save(embedding_layer.state_dict(), model_dir+'/early_embedding_layer.pkl')
# torch.save(att_layer.state_dict(), model_dir+'/early_att_layer.pkl')
# torch.save(model_out_layer.state_dict(), model_dir+'/early_model_out_layer.pkl')
os.system('cp {}/embedding_layer.pkl {}/early_embedding_layer.pkl'.
format(model_dir, model_dir))
os.system('cp {}/att_layer.pkl {}/early_att_layer.pkl'.format(
model_dir, model_dir))
os.system('cp {}/model_out_layer.pkl {}/early_model_out_layer.pkl'.
format(model_dir, model_dir))
else:
low_epoch += 1
log_file.write('low' + str(low_epoch) + '\n')
log_file.flush()
if low_epoch >= config['early_stop']:
break
log_file.close()