def process_data(self):
batches = []
patient_dict = {}
index = 0
for patient in self.patients:
# get patient's visits
patient_dict['pid_' + patient['pid']] = index
index += 1
visits = patient['visits']
# sorting visits by admission date
sorted_visits = sorted(visits, key=lambda visit: visit['admsn_dt'])
# number of visits
no_visits = len(visits)
# generating batch sample: list of visits including concept codes,
# label of last visit mortality
ls_visits = []
label = [int(sorted_visits[no_visits - 1]['Death'])]
for visit in sorted_visits:
codes = visit['DXs']
if not self.dx_only:
codes.extend(visit['CPTs'])
code_size = len(codes)
# code padding
if code_size < self.max_len_visit:
list_zeros = [0] * (self.max_len_visit - code_size)
codes.extend(list_zeros)
ls_visits.append(codes)
# visit padding
if no_visits < self.max_visits:
for i in range(self.max_visits - no_visits):
list_zeros = [0] * self.max_len_visit
ls_visits.append(list_zeros)
# print(len(ls_visits))
batches.append([
np.array(ls_visits, dtype=np.int32),
np.array(label, dtype=np.int32)
])
b_patients = []
b_label = []
for batch in batches:
b_patients.append(batch[0])
b_label.append(batch[1])
save_file({
'patient': b_patients,
'label': b_label
}, join(cfg.dataset_dir, 'mortality_' + cfg.data_source + '.pickle'))
dict_file = join(cfg.dataset_dir, 'mimic3_patient_dict')
print('patient dict file location: ', dict_file)
with open(dict_file + '.json', 'w') as fp:
json.dump(patient_dict, fp)
return b_patients, b_label
def save_preprocess_file(self):
random.shuffle(self.fixed_samples)
dev_rate = 0.1
train_len = int(len(self.fixed_samples) * (1 - dev_rate))
train = self.fixed_samples[:train_len]
dev = self.fixed_samples[train_len:]
save_file(train, pjoin(self.data_dir, 'train.json'), mode='json')
save_file(dev, pjoin(self.data_dir, 'dev.json'), mode='json')
def save_preprocess_file(self):
train_samples = []
dev_samples = []
test_samples = []
for sample in self.samples:
if sample['data_type'] == 1:
train_samples.append(sample)
elif sample['data_type'] == 2:
dev_samples.append(sample)
else:
test_samples.append(sample)
save_file(train_samples,
pjoin(self.data_dir, 'train.json'),
mode='json')
save_file(dev_samples, pjoin(self.data_dir, 'dev.json'), mode='json')
save_file(test_samples, pjoin(self.data_dir, 'test.json'), mode='json')
def train(cfg):
# ======= data ===========
loaded_data = load_file(cfg['processed_path'], 'processed_datasets', mode='pickle')
if loaded_data is None and cfg['load_preproc']:
train_data_obj = Dataset(cfg['train_data_path'], 'train', bpe_path=cfg['bpe_data_dir'])
dev_data_obj = Dataset(cfg['dev_data_path'], 'dev', train_data_obj.bpe_enc)
test_data_obj = Dataset(cfg['test_data_path'], 'test', train_data_obj.bpe_enc)
dev_data_obj.sort_example()
test_data_obj.sort_example()
save_file({'train_data_obj': train_data_obj, 'dev_data_obj': dev_data_obj, 'test_data_obj': test_data_obj},
cfg['processed_path'])
else:
train_data_obj = loaded_data['train_data_obj']
dev_data_obj = loaded_data['dev_data_obj']
test_data_obj = loaded_data['test_data_obj']
# 1. delete too long sentence, max len: 50
train_data_obj.filter_example(cfg['max_sent_len'])
# ========= build network ======
if cfg['model_class'] is None:
print('Did not find the model, please check '
'1) module name 2) class name and 3) implementation of get_default_model_paramters(). exit!')
exit()
with tf.variable_scope("model") as scp:
model = cfg['model_class'](
cfg, train_data_obj.bpe_enc.n_vocab, 3,
train_data_obj.bpe_enc.get_idx_from_token(Dataset.CLASSIFY_TOKEN),
scp.name
)
# ======= build session =======
graph_handler = GraphHandler(model, cfg)
evaluator = Evaluator(model, cfg)
performance_record = PerformanceRecoder(cfg['ckpt_dir'], cfg['save_model'], cfg['save_num'])
sess = graph_handler.initialize()
model.load_openai_pretrained_transformer_model(
sess, cfg['pretrained_transformer_dir'])
# ======== begin to train =======
loss_task_ma, loss_lm_ma, accuracy_ma = MovingAverage(), MovingAverage(), MovingAverage()
for example_batch, batch_num, data_round, idx_b in train_data_obj.generate_batch_iter(
cfg['train_batch_size'], cfg['n_steps']):
global_step_val = sess.run(model.global_step) + 1
step_out = model.step(sess, example_batch, cfg['summary_period'])
loss_task_ma(step_out['task_loss']), accuracy_ma(step_out['accuracy']), loss_lm_ma(step_out['lm_loss'])
graph_handler.add_summary(step_out['summary'], global_step_val)
if global_step_val % 100 == 0:
log.info('data round: %d: %d/%d, global step:%d -- loss: %.4f, accu: %.4f' %
(data_round, idx_b, batch_num, global_step_val,
loss_task_ma.value, accuracy_ma.value))
if 'lm_loss' in step_out:
log.info('\tauxiliary language model perplexity: %.4f' % math.exp(loss_lm_ma.value))
# eval
if global_step_val % cfg['eval_period'] == 0:
dev_res = evaluator.get_evaluation(sess, dev_data_obj, global_step_val)
log.info('==> for dev, loss: %.4f, accuracy: %.4f' %
(dev_res['loss'], dev_res['accuracy']))
criterion_metric = dev_res['accuracy']
if not test_data_obj.no_label:
test_res = evaluator.get_evaluation(sess, test_data_obj, global_step_val)
log.info('~~> for test, loss: %.4f, accuracy: %.4f' %
(test_res['loss'], test_res['accuracy']))
is_in_top, deleted_step = performance_record.update_top_list(global_step_val, criterion_metric, sess)
if is_in_top: # get prediction for non-labeled test data
if test_data_obj.no_label and global_step_val > 0.4 * cfg['n_steps']:
evaluator.get_evaluation(sess, test_data_obj, global_step_val, write_predict=True)
# todo: time count
log.info(str(performance_record.top_list))
def train():
output_model_params()
loadFile = True
ifLoad, data = False, None
if loadFile:
ifLoad, data = load_file(cfg.processed_path, 'processed data',
'pickle')
if not ifLoad or not loadFile:
train_data_obj = Dataset(cfg.train_data_path, 'train')
dev_data_obj = Dataset(cfg.dev_data_path, 'dev', train_data_obj.dicts)
save_file(
{
'train_data_obj': train_data_obj,
'dev_data_obj': dev_data_obj
}, cfg.processed_path)
else:
train_data_obj = data['train_data_obj']
dev_data_obj = data['dev_data_obj']
emb_mat_token, emb_mat_glove = train_data_obj.emb_mat_token, train_data_obj.emb_mat_glove
output_cls_num = len(
train_data_obj.dicts['sub_cls']) if cfg.fine_grained else len(
train_data_obj.dicts['cls'])
with tf.variable_scope(network_type) as scope:
if network_type in model_set:
model = Model(emb_mat_token,
emb_mat_glove,
len(train_data_obj.dicts['token']),
len(train_data_obj.dicts['char']),
train_data_obj.max_lens['token'],
output_cls_num,
scope=scope.name)
graphHandler = GraphHandler(model)
evaluator = Evaluator(model)
performRecoder = PerformRecoder(3)
if cfg.gpu_mem is None:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem, allow_growth=True)
graph_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True)
else:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem)
graph_config = tf.ConfigProto(gpu_options=gpu_options)
# graph_config.gpu_options.allow_growth = True
sess = tf.Session(config=graph_config)
graphHandler.initialize(sess)
# begin training
steps_per_epoch = int(
math.ceil(1.0 * train_data_obj.sample_num / cfg.train_batch_size))
num_steps = cfg.num_steps or steps_per_epoch * cfg.max_epoch
global_step = 0
for sample_batch, batch_num, data_round, idx_b in train_data_obj.generate_batch_sample_iter(
num_steps):
global_step = sess.run(model.global_step) + 1
if_get_summary = global_step % (cfg.log_period or steps_per_epoch) == 0
loss, summary, train_op = model.step(sess,
sample_batch,
get_summary=if_get_summary)
if global_step % 100 == 0:
_logger.add('data round: %d: %d/%d, global step:%d -- loss: %.4f' %
(data_round, idx_b, batch_num, global_step, loss))
if if_get_summary:
graphHandler.add_summary(summary, global_step)
# Occasional evaluation
if global_step % (cfg.eval_period or steps_per_epoch) == 0:
# ---- dev ----
dev_loss, dev_accu = evaluator.get_evaluation(
sess, dev_data_obj, global_step)
_logger.add('==> for dev, loss: %.4f, accuracy: %.4f' %
(dev_loss, dev_accu))
is_in_top, deleted_step = performRecoder.update_top_list(
global_step, dev_accu, sess)
this_epoch_time, mean_epoch_time = cfg.time_counter.update_data_round(
data_round)
# if this_epoch_time is not None and mean_epoch_time is not None:
# _logger.add('##> this epoch time: %f, mean epoch time: %f' % (this_epoch_time, mean_epoch_time))
do_analyse_qc(_logger.path)
def train():
output_model_params()
loadFile = True
ifLoad, data = False, None
if loadFile:
ifLoad, data = load_file(cfg.processed_path, 'processed data',
'pickle')
if not ifLoad or not loadFile:
train_data_obj = Dataset(cfg.train_data_path, 'train')
dev_data_obj = Dataset(cfg.dev_data_path,
'dev',
dicts=train_data_obj.dicts)
test_data_obj = Dataset(cfg.test_data_path,
'test',
dicts=train_data_obj.dicts)
save_file(
{
'train_data_obj': train_data_obj,
'dev_data_obj': dev_data_obj,
'test_data_obj': test_data_obj
}, cfg.processed_path)
train_data_obj.save_dict(cfg.dict_path)
else:
train_data_obj = data['train_data_obj']
dev_data_obj = data['dev_data_obj']
test_data_obj = data['test_data_obj']
train_data_obj.filter_data()
dev_data_obj.filter_data()
test_data_obj.filter_data()
emb_mat_token, emb_mat_glove = train_data_obj.emb_mat_token, train_data_obj.emb_mat_glove
with tf.variable_scope(cfg.base_name) as scope:
model = Model(emb_mat_token, emb_mat_glove,
len(train_data_obj.dicts['token']),
len(train_data_obj.dicts['char']),
train_data_obj.max_lens['token'], scope.name)
graphHandler = GraphHandler(model)
evaluator = Evaluator(model)
performRecoder = PerformRecoder(cfg.save_num)
if cfg.gpu_mem is None:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem, allow_growth=True)
graph_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True)
elif cfg.gpu_mem < 1.:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem)
graph_config = tf.ConfigProto(gpu_options=gpu_options)
else:
gpu_options = tf.GPUOptions()
graph_config = tf.ConfigProto(gpu_options=gpu_options)
sess = tf.Session(config=graph_config)
graphHandler.initialize(sess)
# begin training
steps_per_epoch = int(
math.ceil(1.0 * train_data_obj.sample_num / cfg.train_batch_size))
num_steps = cfg.num_steps or steps_per_epoch * cfg.max_epoch
global_step = 0
for sample_batch, batch_num, data_round, idx_b in train_data_obj.generate_batch_sample_iter(
num_steps):
global_step = sess.run(model.global_step) + 1
if_get_summary = global_step % (cfg.log_period or steps_per_epoch) == 0
loss, summary = model.step(sess,
sample_batch,
get_summary=if_get_summary,
global_step_value=global_step)
if global_step % 100 == 0:
_logger.add(
'data round: %d: %d/%d, global step:%d -- loss_sl: %.4f, loss_rl: %.4f'
%
(data_round, idx_b, batch_num, global_step, loss[0], loss[1]))
if if_get_summary:
graphHandler.add_summary(summary, global_step)
# Occasional evaluation
evaluation = False
if cfg.model_dir_suffix == 'test':
if global_step % (cfg.eval_period or steps_per_epoch) == 0:
evaluation = True
elif is_base_training:
if global_step > cfg.num_steps - 350000 and (
global_step % (cfg.eval_period or steps_per_epoch) == 0):
evaluation = True
else:
if global_step % (cfg.eval_period or steps_per_epoch) == 0:
if cfg.load_model:
evaluation = True
else:
if global_step > 250000:
evaluation = True
if evaluation:
# ---- dev ----
dev_loss, dev_accu, dev_perc = evaluator.get_evaluation(
sess, dev_data_obj, global_step)
_logger.add(
'==> for dev, loss: %.4f %.4f, perc: %.4f, accuracy: %.4f' %
(dev_loss[0], dev_loss[1], dev_perc, dev_accu))
# ---- test ----
test_loss, test_accu, test_perc = evaluator.get_evaluation(
sess, test_data_obj, global_step)
_logger.add(
'~~> for test, loss: %.4f %.4f, perc: %.4f, accuracy: %.4f' %
(test_loss[0], test_loss[1], test_perc, test_accu))
is_in_top, deleted_step = performRecoder.update_top_list(
global_step, dev_accu, sess)
this_epoch_time, mean_epoch_time = cfg.time_counter.update_data_round(
data_round)
if this_epoch_time is not None and mean_epoch_time is not None:
_logger.add('##> this epoch time: %f, mean epoch time: %f' %
(this_epoch_time, mean_epoch_time))
if is_base_training and global_step >= 200000 and global_step % 50000 == 0 and cfg.save_model:
graphHandler.save(sess, global_step)
_logger.writeToFile()
do_analyse_snli_rl(_logger.path)
def multi_test():
assert cfg.load_path is not None
output_model_params()
loadFile = True
ifLoad, data = False, None
if loadFile:
ifLoad, data = load_file(cfg.processed_path, 'processed data',
'pickle')
if not ifLoad or not loadFile:
train_data_obj = Dataset(cfg.train_data_path, 'train')
dev_data_obj = Dataset(cfg.dev_data_path,
'dev',
dicts=train_data_obj.dicts)
test_data_obj = Dataset(cfg.test_data_path,
'test',
dicts=train_data_obj.dicts)
save_file(
{
'train_data_obj': train_data_obj,
'dev_data_obj': dev_data_obj,
'test_data_obj': test_data_obj
}, cfg.processed_path)
train_data_obj.save_dict(cfg.dict_path)
else:
train_data_obj = data['train_data_obj']
dev_data_obj = data['dev_data_obj']
test_data_obj = data['test_data_obj']
train_data_obj.filter_data()
dev_data_obj.filter_data()
test_data_obj.filter_data()
emb_mat_token, emb_mat_glove = train_data_obj.emb_mat_token, train_data_obj.emb_mat_glove
with tf.variable_scope(cfg.base_name) as scope:
model = Model(emb_mat_token, emb_mat_glove,
len(train_data_obj.dicts['token']),
len(train_data_obj.dicts['char']),
train_data_obj.max_lens['token'], scope.name)
graphHandler = GraphHandler(model)
evaluator = Evaluator(model)
if cfg.gpu_mem is None:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem, allow_growth=True)
graph_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True)
elif cfg.gpu_mem < 1.:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem)
graph_config = tf.ConfigProto(gpu_options=gpu_options)
else:
gpu_options = tf.GPUOptions()
graph_config = tf.ConfigProto(gpu_options=gpu_options)
sess = tf.Session(config=graph_config)
graphHandler.initialize(sess)
repeat_num = 10
time_counter = TimeCounter()
for t in range(repeat_num):
# ---- dev ----
test_loss, test_accu, test_perc = evaluator.get_evaluation(
sess, test_data_obj, None, time_counter=time_counter)
_logger.add(
'==> for test, loss: %.4f %.4f, perc: %.4f, accuracy: %.4f' %
(test_loss[0], test_loss[1], test_perc, test_accu))
print(time_counter.update_data_round(t + 1))
def save_dict(self, path):
save_file(self.dicts, path, 'token and char dict data', 'pickle')
def test():
output_model_params()
loadFile = True
ifLoad, data = False, None
if loadFile:
ifLoad, data = load_file(cfg.processed_path, 'processed data',
'pickle')
if not ifLoad or not loadFile:
raw_data = RawDataProcessor(cfg.data_dir)
train_data_list = raw_data.get_data_list('train')
dev_data_list = raw_data.get_data_list('dev')
test_data_list = raw_data.get_data_list('test')
train_data_obj = Dataset(train_data_list, 'train')
dev_data_obj = Dataset(dev_data_list, 'dev', train_data_obj.dicts)
test_data_obj = Dataset(test_data_list, 'test', train_data_obj.dicts)
save_file(
{
'train_data_obj': train_data_obj,
'dev_data_obj': dev_data_obj,
'test_data_obj': test_data_obj
}, cfg.processed_path)
train_data_obj.save_dict(cfg.dict_path)
else:
train_data_obj = data['train_data_obj']
dev_data_obj = data['dev_data_obj']
test_data_obj = data['test_data_obj']
train_data_obj.filter_data(True)
dev_data_obj.filter_data(True)
test_data_obj.filter_data(True)
emb_mat_token, emb_mat_glove = train_data_obj.emb_mat_token, train_data_obj.emb_mat_glove
with tf.variable_scope(network_type) as scope:
if network_type in model_set:
model = Model(emb_mat_token, emb_mat_glove,
len(train_data_obj.dicts['token']),
len(train_data_obj.dicts['char']),
train_data_obj.max_lens['token'], scope.name)
graphHandler = GraphHandler(model)
evaluator = Evaluator(model)
if cfg.gpu_mem is None:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem, allow_growth=True)
graph_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True)
else:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem)
graph_config = tf.ConfigProto(gpu_options=gpu_options)
# graph_config.gpu_options.allow_growth = True
sess = tf.Session(config=graph_config)
graphHandler.initialize(sess)
# todo: test model
# ---- dev ----
dev_loss, dev_accu, dev_sent_accu = evaluator.get_evaluation(
sess, dev_data_obj, 1)
_logger.add(
'==> for dev, loss: %.4f, accuracy: %.4f, sentence accuracy: %.4f' %
(dev_loss, dev_accu, dev_sent_accu))
# ---- test ----
test_loss, test_accu, test_sent_accu = evaluator.get_evaluation(
sess, test_data_obj, 1)
_logger.add(
'~~> for test, loss: %.4f, accuracy: %.4f, sentence accuracy: %.4f' %
(test_loss, test_accu, test_sent_accu))
# ---- train ----
train_loss, train_accu, train_sent_accu = evaluator.get_evaluation(
sess, train_data_obj, 1)
_logger.add(
'--> for test, loss: %.4f, accuracy: %.4f, sentence accuracy: %.4f' %
(train_loss, train_accu, train_sent_accu))
def train():
n_fold_val = 10
output_model_params()
loadFile = True
ifLoad, data = False, None
if loadFile:
ifLoad, data = load_file(cfg.processed_path, 'processed data',
'pickle')
if not ifLoad or not loadFile:
data_obj = Dataset(cfg.data_path, cfg.dataset_type)
save_file({'data_obj': data_obj}, cfg.processed_path)
else:
data_obj = data['data_obj']
data_obj.split_dataset_to_blocks(n_fold_val)
# for block len
if cfg.block_len is None and cfg.context_fusion_method == 'block':
_logger.add()
_logger.add('calculating block length for dataset')
statistic = data_obj.get_statistic()
expected_n = statistic['mean'] + statistic['std'] * math.sqrt(
2. * math.log(1. * cfg.train_batch_size))
dy_block_len = math.ceil(math.pow(2 * expected_n,
1.0 / 3)) + 1 # fixme: change length
cfg.block_len = dy_block_len
_logger.add('block length is %d' % dy_block_len)
emb_mat_token, emb_mat_glove = data_obj.emb_mat_token, data_obj.emb_mat_glove
output_cls_num = data_obj.class_num
steps_per_epoch = int(
math.ceil(1.0 * data_obj.sample_num / cfg.train_batch_size))
num_steps = cfg.num_steps or steps_per_epoch * cfg.max_epoch
dev_performance_list = []
for n_th_fold in range(n_fold_val):
time_accu_recorder = TimeAccuRecorder(data_obj.dataset_type, n_th_fold,
cfg.answer_dir)
g = tf.Graph()
with g.as_default():
with tf.variable_scope("%s_%s" %
(cfg.dataset_type, network_type)) as scope:
if network_type in model_set:
model = Model(emb_mat_token,
emb_mat_glove,
len(data_obj.dicts['token']),
len(data_obj.dicts['char']),
data_obj.max_lens['token'],
output_cls_num,
scope=scope.name)
else:
assert RuntimeError
graphHandler = GraphHandler(model)
evaluator = Evaluator(model)
performRecoder = PerformRecoder(1)
if cfg.gpu_mem is None:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem,
allow_growth=True)
graph_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True)
else:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem)
graph_config = tf.ConfigProto(gpu_options=gpu_options)
# graph_config.gpu_options.allow_growth = True
sess = tf.Session(config=graph_config)
graphHandler.initialize(sess)
global_step = 0
for sample_batch, batch_num, data_round, idx_b in \
data_obj.generate_batch_sample_iter(n_th_fold, num_steps):
global_step = sess.run(model.global_step) + 1
if_get_summary = global_step % (cfg.log_period
or steps_per_epoch) == 0
loss, summary, train_op = model.step(
sess, sample_batch, get_summary=if_get_summary)
# if global_step % 100 == 0:
_logger.add('cross validation index: %d' % n_th_fold)
_logger.add(
'data round: %d: %d/%d, global step:%d -- loss: %.4f' %
(data_round, idx_b, batch_num, global_step, loss))
if if_get_summary:
graphHandler.add_summary(summary, global_step)
# Occasional evaluation
if global_step % (cfg.eval_period or steps_per_epoch) == 0:
# ---- dev ----
dev_loss, dev_accu = evaluator.get_evaluation(
sess, data_obj, n_th_fold, global_step)
_logger.add('==> for dev, loss: %.4f, accuracy: %.4f' %
(dev_loss, dev_accu))
# record time vs. accuracy
time_accu_recorder.add_data(
cfg.time_counter.global_training_time, dev_accu)
is_in_top, deleted_step = performRecoder.update_top_list(
global_step, dev_accu, sess)
this_epoch_time, mean_epoch_time = cfg.time_counter.update_data_round(
data_round)
# if this_epoch_time is not None and mean_epoch_time is not None:
# _logger.add('##> this epoch time: %f, mean epoch time: %f' % (this_epoch_time, mean_epoch_time))
dev_performance_list.append(performRecoder.best_result)
_logger.add("%d th x val accuracy is %.4f" %
(n_th_fold, performRecoder.best_result))
time_accu_recorder.save_to_file()
if len(dev_performance_list) > 0:
dev_performance_array = np.array(dev_performance_list)
xval_average = np.mean(dev_performance_array)
xval_std = np.std(dev_performance_array)
else:
xval_average = 0
xval_std = 0
dev_performance_list_str = [str(elem) for elem in dev_performance_list]
_logger.add("all accuracies: %s" % ', '.join(dev_performance_list_str))
_logger.add('%d fold cross validation average accuracy is %f, standard variance is %f' % \
(n_fold_val, xval_average, xval_std))
_logger.writeToFile()
def train():
output_model_params()
loadFile = True
ifLoad, data = False, None
if loadFile:
ifLoad, data = load_file(cfg.processed_path, 'data', 'pickle')
if not ifLoad or not loadFile:
data_object = Dataset(cfg.train_data_path, cfg.dev_data_path)
data_object.save_dict(cfg.dict_path)
save_file({'data_obj': data_object}, cfg.processed_path)
else:
data_object = data['data_obj']
emb_mat_token, emb_mat_glove = data_object.emb_mat_token, data_object.emb_mat_glove
with tf.variable_scope(network_type) as scope:
if network_type in model_set:
model = Model(emb_mat_token, emb_mat_glove,
len(data_object.dicts['token']),
len(data_object.dicts['char']),
data_object.max_lens['token'], scope.name)
graphHandler = GraphHandler(model)
evaluator = Evaluator(model)
performRecoder = PerformRecoder(5)
if cfg.gpu_mem < 1.:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem, allow_growth=True)
else:
gpu_options = tf.GPUOptions()
graph_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True)
sess = tf.Session(config=graph_config)
graphHandler.initialize(sess)
# begin training
steps_per_epoch = int(
math.ceil(1.0 * len(data_object.digitized_train_data_list) /
cfg.train_batch_size))
num_steps = steps_per_epoch * cfg.max_epoch or cfg.num_steps
global_step = 0
# debug or not
if cfg.debug:
sess = tf_debug.LocalCLIDebugWrapperSession(sess)
for sample_batch, batch_num, data_round, idx_b in Dataset.generate_batch_sample_iter(
data_object.digitized_train_data_list, num_steps):
global_step = sess.run(model.global_step) + 1
if_get_summary = global_step % (cfg.log_period or steps_per_epoch) == 0
loss, summary, train_op = model.step(sess,
sample_batch,
get_summary=if_get_summary)
if global_step % 10 == 0:
_logger.add('data round: %d: %d/%d, global step:%d -- loss: %.4f' %
(data_round, idx_b, batch_num, global_step, loss))
if if_get_summary:
graphHandler.add_summary(summary, global_step)
# Occasional evaluation
if global_step % (cfg.eval_period or steps_per_epoch) == 0:
# ---- dev ----
dev_loss, dev_accu = evaluator.get_evaluation(
sess, data_object.digitized_dev_data_list, 'dev', global_step)
_logger.add('==> for dev, loss: %.4f, accuracy: %.4f' %
(dev_loss, dev_accu))
# ---- test ----
if cfg.test_data_name != None:
test_loss, test_accu = evaluator.get_evaluation(
sess, data_object.digitized_test_data_list, 'test',
global_step)
_logger.add('~~> for test, loss: %.4f, accuracy: %.4f' %
(test_loss, test_accu))
is_in_top, deleted_step = performRecoder.update_top_list(
global_step, dev_accu, sess)
this_epoch_time, mean_epoch_time = cfg.time_counter.update_data_round(
data_round)
if this_epoch_time is not None and mean_epoch_time is not None:
_logger.add('##> this epoch time: %f, mean epoch time: %f' %
(this_epoch_time, mean_epoch_time))
def train():
output_model_params()
loadFile = True
ifLoad, data = load_file(cfg.processed_path, 'processed data', 'pickle')
if not ifLoad or not loadFile:
train_data_obj = Dataset(cfg.train_dataset_path, data_type='train')
dev_data_obj = Dataset(cfg.dev_dataset_path,
data_type='dev',
dicts=train_data_obj.dicts)
save_file({
'train_data': train_data_obj,
'dev_data': dev_data_obj
}, cfg.processed_path)
else:
train_data_obj = data['train_data']
dev_data_obj = data['dev_data']
train_data_obj.filter_data()
emb_mat_token, emb_mat_glove = train_data_obj.emb_mat_token, train_data_obj.emb_mat_glove
# for block len
if cfg.block_len is None and cfg.context_fusion_method == 'block':
_logger.add()
_logger.add('calculating block length for dataset')
statistic = train_data_obj.get_statistic()
expected_n = statistic['mean'] + statistic['std'] * math.sqrt(
2. * math.log(1. * cfg.train_batch_size))
dy_block_len = math.ceil(math.pow(2 * expected_n,
1.0 / 3)) + 1 # fixme: change length
cfg.block_len = dy_block_len
_logger.add('block length is %d' % dy_block_len)
with tf.variable_scope(network_type) as scope:
model = Model(emb_mat_token, emb_mat_glove,
len(train_data_obj.dicts['token']),
cfg.word_embedding_length, cfg.hidden_units_num,
scope.name)
graph_handler = GraphHandler(model)
evaluator = Evaluator(model)
perform_recoder = PerformRecoder(3)
if cfg.gpu_mem is None:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem, allow_growth=True)
graph_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True)
else:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem)
graph_config = tf.ConfigProto(gpu_options=gpu_options)
sess = tf.Session(config=graph_config)
graph_handler.initialize(sess)
# begin training
steps_per_epoch = int(
math.ceil(1.0 * train_data_obj.sample_num / cfg.train_batch_size))
num_steps = cfg.num_steps or steps_per_epoch * cfg.max_epoch
global_step = 0
for sample_batch, batch_num, data_round, idx_b in train_data_obj.generate_batch_sample_iter(
num_steps):
global_step = sess.run(model.global_step) + 1
if_get_summary = global_step % (cfg.log_period or steps_per_epoch) == 0
loss, summary, train_op = model.step(sess,
sample_batch,
get_summary=if_get_summary)
if global_step % 100 == 0:
_logger.add('data round: %d: %d/%d, global step:%d -- loss: %.4f' %
(data_round, idx_b, batch_num, global_step, loss))
if if_get_summary:
graph_handler.add_summary(summary, global_step)
# period evaluation
if (global_step > (cfg.num_steps - 20000) or cfg.model_dir_suffix == 'test') and \
global_step % (cfg.eval_period or steps_per_epoch) == 0:
# ---- dev ----
dev_loss, dev_accu = evaluator.get_evaluation(
sess, dev_data_obj, global_step)
_logger.add('==> for dev, loss: %.4f, accuracy: %.4f' %
(dev_loss, dev_accu))
is_in_top, deleted_step = perform_recoder.update_top_list(
global_step, dev_accu, sess)
this_epoch_time, mean_epoch_time = cfg.time_counter.update_data_round(
data_round)
if this_epoch_time is not None and mean_epoch_time is not None:
_logger.add('##> this epoch time: %f, mean epoch time: %f' %
(this_epoch_time, mean_epoch_time))
_logger.writeToFile()
do_analysis_squad_sim(_logger.path)
def train():
output_model_params()
loadFile = True
ifLoad, data = False, None
if loadFile:
ifLoad, data = load_file(cfg.processed_path, 'processed data',
'pickle')
if not ifLoad or not loadFile:
train_data_obj = Dataset(cfg.train_data_path, 'train')
dev_data_obj = Dataset(cfg.dev_data_path,
'dev',
dicts=train_data_obj.dicts)
test_data_obj = Dataset(cfg.test_data_path,
'test',
dicts=train_data_obj.dicts)
save_file(
{
'train_data_obj': train_data_obj,
'dev_data_obj': dev_data_obj,
'test_data_obj': test_data_obj
}, cfg.processed_path)
train_data_obj.save_dict(cfg.dict_path)
else:
train_data_obj = data['train_data_obj']
dev_data_obj = data['dev_data_obj']
test_data_obj = data['test_data_obj']
train_data_obj.filter_data()
dev_data_obj.filter_data()
test_data_obj.filter_data()
emb_mat_token, emb_mat_glove = train_data_obj.emb_mat_token, train_data_obj.emb_mat_glove
with tf.variable_scope(network_type) as scope:
model = Model(emb_mat_token, emb_mat_glove,
len(train_data_obj.dicts['token']),
len(train_data_obj.dicts['char']),
train_data_obj.max_lens['token'], scope.name)
graphHandler = GraphHandler(model)
evaluator = Evaluator(model)
performRecoder = PerformRecoder(5)
sess = tf.Session(config=sess_config_gene(cfg.gpu_mem))
graphHandler.initialize(sess)
# begin training
steps_per_epoch = int(
math.ceil(1.0 * train_data_obj.sample_num / cfg.train_batch_size))
num_steps = cfg.num_steps or steps_per_epoch * cfg.max_epoch
global_step = 0
for sample_batch, batch_num, data_round, idx_b in train_data_obj.generate_batch_sample_iter(
num_steps):
global_step = sess.run(model.global_step) + 1
if_get_summary = global_step % (cfg.log_period or steps_per_epoch) == 0
loss, summary, train_op = model.step(sess,
sample_batch,
get_summary=if_get_summary)
if global_step % 100 == 0:
_logger.add('data round: %d: %d/%d, global step:%d -- loss: %.4f' %
(data_round, idx_b, batch_num, global_step, loss))
if if_get_summary:
graphHandler.add_summary(summary, global_step)
# # occasional saving
# if global_step % (cfg.save_period or steps_per_epoch) == 0:
# graphHandler.save(sess, global_step)
# Occasional evaluation
if (global_step > (cfg.num_steps - 300000) or cfg.model_dir_prefix=='test') and \
global_step % (cfg.eval_period or steps_per_epoch) == 0:
# ---- dev ----
dev_loss, dev_accu = evaluator.get_evaluation(
sess, dev_data_obj, global_step)
_logger.add('==> for dev, loss: %.4f, accuracy: %.4f' %
(dev_loss, dev_accu))
# ---- test ----
test_loss, test_accu = evaluator.get_evaluation(
sess, test_data_obj, global_step)
_logger.add('~~> for test, loss: %.4f, accuracy: %.4f' %
(test_loss, test_accu))
is_in_top, deleted_step = performRecoder.update_top_list(
global_step, dev_accu, sess)
this_epoch_time, mean_epoch_time = cfg.time_counter.update_data_round(
data_round)
if this_epoch_time is not None and mean_epoch_time is not None:
_logger.add('##> this epoch time: %f, mean epoch time: %f' %
(this_epoch_time, mean_epoch_time))
# if global_step % (cfg.save_period or steps_per_epoch) != 0:
# graphHandler.save(sess, global_step)
_logger.writeToFile()
do_analyse_snli(_logger.path)
def train():
output_model_params()
loadFile = True
ifLoad, data = False, None
if loadFile:
ifLoad, data = load_file(cfg.processed_path, 'processed data',
'pickle')
if not ifLoad or not loadFile:
raw_data = RawDataProcessor(cfg.data_dir)
train_data_list = raw_data.get_data_list('train')
dev_data_list = raw_data.get_data_list('dev')
test_data_list = raw_data.get_data_list('test')
train_data_obj = Dataset(train_data_list, 'train')
dev_data_obj = Dataset(dev_data_list, 'dev', train_data_obj.dicts)
test_data_obj = Dataset(test_data_list, 'test', train_data_obj.dicts)
save_file(
{
'train_data_obj': train_data_obj,
'dev_data_obj': dev_data_obj,
'test_data_obj': test_data_obj
}, cfg.processed_path)
train_data_obj.save_dict(cfg.dict_path)
else:
train_data_obj = data['train_data_obj']
dev_data_obj = data['dev_data_obj']
test_data_obj = data['test_data_obj']
train_data_obj.filter_data(cfg.only_sentence, cfg.fine_grained)
dev_data_obj.filter_data(True, cfg.fine_grained)
test_data_obj.filter_data(True, cfg.fine_grained)
# for block len
if cfg.block_len is None and cfg.context_fusion_method == 'block':
_logger.add()
_logger.add('calculating block length for dataset')
statistic = train_data_obj.get_statistic()
expected_n = statistic['mean'] + statistic['std'] * math.sqrt(
2. * math.log(1. * cfg.train_batch_size))
dy_block_len = math.ceil(math.pow(2 * expected_n,
1.0 / 3)) + 1 # fixme: change length
cfg.block_len = dy_block_len
_logger.add('block length is %d' % dy_block_len)
emb_mat_token, emb_mat_glove = train_data_obj.emb_mat_token, train_data_obj.emb_mat_glove
with tf.variable_scope(network_type) as scope:
if network_type in model_set:
model = Model(emb_mat_token, emb_mat_glove,
len(train_data_obj.dicts['token']),
len(train_data_obj.dicts['char']),
train_data_obj.max_lens['token'], scope.name)
graphHandler = GraphHandler(model)
evaluator = Evaluator(model)
performRecoder = PerformRecoder(3)
if cfg.gpu_mem is None:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem, allow_growth=True)
graph_config = tf.ConfigProto(gpu_options=gpu_options,
allow_soft_placement=True)
else:
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=cfg.gpu_mem)
graph_config = tf.ConfigProto(gpu_options=gpu_options)
# graph_config.gpu_options.allow_growth = True
sess = tf.Session(config=graph_config)
graphHandler.initialize(sess)
# begin training
steps_per_epoch = int(
math.ceil(1.0 * train_data_obj.sample_num / cfg.train_batch_size))
num_steps = cfg.num_steps or steps_per_epoch * cfg.max_epoch
global_step = 0
for sample_batch, batch_num, data_round, idx_b in train_data_obj.generate_batch_sample_iter(
num_steps):
global_step = sess.run(model.global_step) + 1
if_get_summary = global_step % (cfg.log_period or steps_per_epoch) == 0
loss, summary, train_op = model.step(sess,
sample_batch,
get_summary=if_get_summary)
if global_step % 100 == 0:
_logger.add('data round: %d: %d/%d, global step:%d -- loss: %.4f' %
(data_round, idx_b, batch_num, global_step, loss))
if if_get_summary:
graphHandler.add_summary(summary, global_step)
# Occasional evaluation
if global_step % (cfg.eval_period or steps_per_epoch) == 0:
# ---- dev ----
dev_loss, dev_accu, dev_sent_accu = evaluator.get_evaluation(
sess, dev_data_obj, global_step)
_logger.add(
'==> for dev, loss: %.4f, accuracy: %.4f, sentence accuracy: %.4f'
% (dev_loss, dev_accu, dev_sent_accu))
# ---- test ----
test_loss, test_accu, test_sent_accu = evaluator.get_evaluation(
sess, test_data_obj, global_step)
_logger.add(
'~~> for test, loss: %.4f, accuracy: %.4f, sentence accuracy: %.4f'
% (test_loss, test_accu, test_sent_accu))
# ---- train ----
# train_loss, train_accu, train_sent_accu = evaluator.get_evaluation(
# sess, train_data_obj, global_step
# )
# _logger.add('--> for train, loss: %.4f, accuracy: %.4f, sentence accuracy: %.4f' %
# (train_loss, train_accu, train_sent_accu))
is_in_top, deleted_step = performRecoder.update_top_list(
global_step, dev_accu, sess)
if is_in_top and global_step > 30000: # todo-ed: delete me to run normally
# evaluator.get_evaluation_file_output(sess, dev_data_obj, global_step, deleted_step)
evaluator.get_evaluation_file_output(sess, test_data_obj,
global_step, deleted_step)
this_epoch_time, mean_epoch_time = cfg.time_counter.update_data_round(
data_round)
if this_epoch_time is not None and mean_epoch_time is not None:
_logger.add('##> this epoch time: %f, mean epoch time: %f' %
(this_epoch_time, mean_epoch_time))
# finally save
# if global_step % (cfg.save_period or steps_per_epoch) != 0:
# graphHandler.save(sess, global_step)
do_analyse_sst(_logger.path)
def process_data(self):
batches = []
for patient in self.patients:
visits = patient['visits']
selected_visits = sorted(visits,
key=lambda visit: visit['admsn_dt'])
selected_codes = []
# reduced window
if self.is_reduced_window:
reduced_window = random.RandomState(1).randint(
self.skip_window)
else:
reduced_window = 0
# actual window
actual_window = self.skip_window - reduced_window
# concat all codes togeter for one patient, and the format is [code, date]
for s_visit in selected_visits:
dt = datetime.datetime.strptime(s_visit['admsn_dt'], "%Y%m%d")
codes = s_visit['DXs']
if not self.dx_only:
codes.extend(s_visit['CPTs'])
for code in codes:
selected_codes.append([code, dt])
# sampling codes based on their frequncy in dataset
if self.is_sample:
sampled_codes = [
code for code in selected_codes
if self.word_sample[code[0]] > random.rand() * 2**32
]
else:
sampled_codes = selected_codes
# generate batch samples
for pos, word in enumerate(sampled_codes):
# now go over all words from the actual window, predicting each one in turn
start = max(0, pos - actual_window)
window_pos = enumerate(
sampled_codes[start:(pos + actual_window + 1)], start)
context_indices = [[word2[0], (word2[1] - word[1]).days]
for pos2, word2 in window_pos
if (word2[0] is not None and pos2 != pos)]
context_len = len(context_indices)
if context_len > 0:
# if context lenth is less than two times actual window, and padding
if context_len < 2 * actual_window:
for i in range(2 * actual_window - context_len):
context_indices.append([0, 0])
intervals = np.zeros(
(2 * actual_window, 2 * actual_window))
for i in range(2 * actual_window):
for j in range(2 * actual_window):
if i > j:
code_i = context_indices[i][0]
code_j = context_indices[j][0]
interval_i = context_indices[i][1]
interval_j = context_indices[j][1]
if code_i > 0 and code_j > 0:
intervals[i, j] = np.abs(interval_i -
interval_j) + 1
intervals[j, i] = np.abs(interval_i -
interval_j) + 1
batches.append([
np.array(context_indices, dtype=np.int32), intervals,
np.array([word[0]], dtype=np.int32)
])
contexts = []
intervals = []
labels = []
for batch in batches:
contexts.append(batch[0])
intervals.append(batch[1])
labels.append(batch[2])
save_file(
{
'contexts': contexts,
'intervals': intervals,
'labels': labels
}, cfg.processed_path)
return contexts, intervals, labels
