def build_model(args, i_cv):
args.net = ARCHI(n_in=29, n_out=2, n_unit=args.n_unit)
args.adv_net = ARCHI(n_in=2, n_out=2, n_unit=args.n_unit)
args.net_optimizer = get_optimizer(args)
args.adv_optimizer = get_optimizer(args, args.adv_net)
args.net_criterion = WeightedCrossEntropyLoss()
args.adv_criterion = WeightedGaussEntropyLoss()
model = get_model(args, PivotClassifier)
model.set_info(DATA_NAME, BENCHMARK_NAME, i_cv)
return model
def build_model(args, i_cv):
args.net = ARCHI(n_in=29, n_out=2, n_unit=args.n_unit)
args.optimizer = get_optimizer(args)
model = get_model(args, Regressor)
model.base_name = CALIB
model.set_info(DATA_NAME, BENCHMARK_NAME, i_cv)
return model
def build_model(args, i_cv):
args.net = ARCHI(n_in=3, n_out=args.n_bins)
args.optimizer = get_optimizer(args)
args.criterion = S3DLoss()
model = get_model(args, Inferno)
model.set_info(DATA_NAME, BENCHMARK_NAME, i_cv)
return model
def main():
# BASIC SETUP
logger = set_logger()
args = REG_parse_args(
main_description="Training launcher for Regressor on S3D2 benchmark")
logger.info(args)
flush(logger)
# INFO
args.net = AR5R5E(n_in=3, n_out=2, n_extra=2)
args.optimizer = get_optimizer(args)
model = get_model(args, Regressor)
model.set_info(DATA_NAME, BENCHMARK_NAME, -1)
pb_config = S3D2Config()
# RUN
results = [run(args, i_cv) for i_cv in range(N_ITER)]
results = pd.concat(results, ignore_index=True)
results.to_csv(os.path.join(model.results_directory, 'results.csv'))
# EVALUATION
eval_table = evaluate_estimator(pb_config.INTEREST_PARAM_NAME, results)
print_line()
print_line()
print(eval_table)
print_line()
print_line()
eval_table.to_csv(os.path.join(model.results_directory, 'evaluation.csv'))
gather_images(model.results_directory)
def build_model(args, i_cv):
args.net = ARCHI(n_in=3, n_out=2, n_unit=args.n_unit)
args.optimizer = get_optimizer(args)
model = get_model(args, NeuralNetClassifier)
model.base_name = "DataAugmentation"
model.set_info(DATA_NAME, BENCHMARK_NAME, i_cv)
return model
def add_training_op(self, loss):
variables = tf.trainable_variables()
gradients = tf.gradients(loss, variables)
gradients, _ = tf.clip_by_global_norm(gradients,
self.config.max_grad_norm)
if self.config.learning_rate_annealing:
global_step = tf.Variable(0, trainable=False)
learning_rate = tf.train.exponential_decay(
self.config.learning_rate,
global_step,
1250,
0.96,
staircase=False)
global_step = tf.add(1, global_step)
else:
learning_rate = self.config.learning_rate
optimizer = get_optimizer(self.config.optimizer, learning_rate)
train_op = optimizer.apply_gradients(zip(gradients, variables))
# For applying EMA for trained parameters
if self.config.ema_for_weights:
ema = tf.train.ExponentialMovingAverage(0.999)
ema_op = ema.apply(variables)
with tf.control_dependencies([train_op]):
train_op = tf.group(ema_op)
return train_op
def build_model(args, i_cv):
args.net = ARCHI(n_in=29, n_out=N_BINS, n_unit=args.n_unit)
args.optimizer = get_optimizer(args)
args.criterion = HiggsLoss()
model = get_model(args, Inferno)
model.set_info(DATA_NAME, BENCHMARK_NAME, i_cv)
return model
def load_calib_rescale():
args = lambda: None
args.n_unit = 80
args.optimizer_name = "adam"
args.beta1 = 0.5
args.beta2 = 0.9
args.learning_rate = 1e-4
args.n_samples = 1000
args.n_steps = 1000
args.batch_size = 20
args.net = CALIB_ARCHI(n_in=1, n_out=2, n_unit=args.n_unit)
args.optimizer = get_optimizer(args)
model = get_model(args, Regressor)
model.base_name = CALIB_RESCALE
model.set_info(DATA_NAME, BENCHMARK_NAME, 0)
model.load(model.model_path)
return model
def load_calib_lam(DATA_NAME, BENCHMARK_NAME):
from model.regressor import Regressor
from archi.reducer import A1AR8MR8L1 as CALIB_ARCHI
args = lambda: None
args.n_unit = 200
args.optimizer_name = "adam"
args.beta1 = 0.5
args.beta2 = 0.9
args.learning_rate = 1e-4
args.n_samples = 1000
args.n_steps = 2000
args.batch_size = 20
args.net = CALIB_ARCHI(n_in=3, n_out=2, n_unit=args.n_unit)
args.optimizer = get_optimizer(args)
model = get_model(args, Regressor)
model.base_name = "Calib_lam"
model.set_info(DATA_NAME, BENCHMARK_NAME, 0)
model.load(model.model_path)
return model
def model_fn(features, labels, mode, params):
if mode == tf.estimator.ModeKeys.TRAIN:
params.update({"training": True})
else:
params.update({"training": False})
features = get_processed_feature(features)
output = model_fn_core(features, params)
if (mode == tf.estimator.ModeKeys.TRAIN or mode == tf.estimator.ModeKeys.EVAL):
loss = get_loss(output, labels)
eval_metric_ops = get_metric(output, labels)
else:
loss = None
eval_metric_ops = None
if mode == tf.estimator.ModeKeys.TRAIN:
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops): # do not forget this | important for batch_normalization
train_op = get_optimizer(opt_algo=params["opt_algo"],
learning_rate=params["learning_rate"],
loss=loss,
global_step=tf.train.get_global_step())
else:
train_op = None
if mode == tf.estimator.ModeKeys.PREDICT:
predictions = get_predict(output)
export_outputs = get_export_outputs(output)
else:
predictions = None
export_outputs = None
return tf.estimator.EstimatorSpec(
mode=mode,
predictions=predictions,
loss=loss,
train_op=train_op,
export_outputs=export_outputs,
eval_metric_ops=eval_metric_ops)
def run(args, i_cv):
logger = logging.getLogger()
print_line()
logger.info('Running iter n°{}'.format(i_cv))
print_line()
result_row = {'i_cv': i_cv}
result_table = []
# LOAD/GENERATE DATA
logger.info('Set up data generator')
pb_config = AP1Config()
seed = config.SEED + i_cv * 5
train_generator = Generator(param_generator, AP1(seed))
valid_generator = AP1(seed + 1)
test_generator = AP1(seed + 2)
# SET MODEL
logger.info('Set up rergessor')
args.net = F3R3(n_in=1, n_out=2)
args.optimizer = get_optimizer(args)
model = get_model(args, Regressor)
model.set_info(BENCHMARK_NAME, i_cv)
flush(logger)
# TRAINING / LOADING
if not args.retrain:
try:
logger.info('loading from {}'.format(model.path))
model.load(model.path)
except Exception as e:
logger.warning(e)
args.retrain = True
if args.retrain:
logger.info('Training {}'.format(model.get_name()))
model.fit(train_generator)
logger.info('Training DONE')
# SAVE MODEL
save_model(model)
# CHECK TRAINING
logger.info('Plot losses')
plot_REG_losses(model)
plot_REG_log_mse(model)
result_row['loss'] = model.losses[-1]
result_row['mse_loss'] = model.mse_losses[-1]
# MEASUREMENT
for mu in pb_config.TRUE_APPLE_RATIO_RANGE:
pb_config.TRUE_APPLE_RATIO = mu
logger.info('Generate testing data')
X_test, y_test, w_test = test_generator.generate(
apple_ratio=pb_config.TRUE_APPLE_RATIO,
n_samples=pb_config.N_TESTING_SAMPLES)
pred, sigma = model.predict(X_test, w_test)
name = pb_config.INTEREST_PARAM_NAME
result_row[name] = pred
result_row[name + _ERROR] = sigma
result_row[name + _TRUTH] = pb_config.TRUE_APPLE_RATIO
logger.info('{} =vs= {} +/- {}'.format(pb_config.TRUE_APPLE_RATIO,
pred, sigma))
result_table.append(result_row.copy())
result_table = pd.DataFrame(result_table)
logger.info('Plot params')
param_names = pb_config.PARAM_NAMES
for name in param_names:
plot_params(name, result_table, model)
logger.info('DONE')
return result_table
def run(args, i_cv):
logger = logging.getLogger()
print_line()
logger.info('Running iter n°{}'.format(i_cv))
print_line()
result_row = {'i_cv': i_cv}
result_table = []
# LOAD/GENERATE DATA
logger.info('Set up data generator')
pb_config = S3D2Config()
seed = config.SEED + i_cv * 5
train_generator = S3D2(seed)
valid_generator = S3D2(seed + 1)
test_generator = S3D2(seed + 2)
# SET MODEL
logger.info('Set up rergessor')
args.net = AR5R5E(n_in=3, n_out=2, n_extra=2)
args.optimizer = get_optimizer(args)
model = get_model(args, Regressor)
model.set_info(BENCHMARK_NAME, i_cv)
model.param_generator = param_generator
flush(logger)
# TRAINING / LOADING
if not args.retrain:
try:
logger.info('loading from {}'.format(model.model_path))
model.load(model.model_path)
except Exception as e:
logger.warning(e)
args.retrain = True
if args.retrain:
logger.info('Training {}'.format(model.get_name()))
model.fit(train_generator)
logger.info('Training DONE')
# SAVE MODEL
save_model(model)
# CHECK TRAINING
logger.info('Plot losses')
plot_REG_losses(model)
plot_REG_log_mse(model)
result_row['loss'] = model.losses[-1]
result_row['mse_loss'] = model.mse_losses[-1]
# MEASUREMENT
for mu in pb_config.TRUE_MU_RANGE:
pb_config.TRUE_MU = mu
logger.info('Generate testing data')
test_generator.reset()
X_test, y_test, w_test = test_generator.generate(
# pb_config.TRUE_R,
# pb_config.TRUE_LAMBDA,
pb_config.CALIBRATED_R,
pb_config.CALIBRATED_LAMBDA,
pb_config.TRUE_MU,
n_samples=pb_config.N_TESTING_SAMPLES)
p_test = np.array(
(pb_config.CALIBRATED_R, pb_config.CALIBRATED_LAMBDA))
pred, sigma = model.predict(X_test, w_test, p_test)
name = pb_config.INTEREST_PARAM_NAME
result_row[name] = pred
result_row[name + _ERROR] = sigma
result_row[name + _TRUTH] = pb_config.TRUE_MU
logger.info('{} =vs= {} +/- {}'.format(pb_config.TRUE_MU, pred, sigma))
result_table.append(result_row.copy())
result_table = pd.DataFrame(result_table)
logger.info('Plot params')
name = pb_config.INTEREST_PARAM_NAME
plot_params(name,
result_table,
title=model.full_name,
directory=model.results_path)
logger.info('DONE')
return result_table
def build_model(args, i_cv):
args.net = ARCHI(n_in=1, n_out=2, n_unit=args.n_unit)
args.optimizer = get_optimizer(args)
model = get_model(args, TangentPropClassifier)
model.set_info(DATA_NAME, BENCHMARK_NAME, i_cv)
return model
def build_model(args, i_cv):
args.net = ARCHI(n_in=1, n_out=2, n_unit=args.n_unit)
args.optimizer = get_optimizer(args)
model = get_model(args, NeuralNetClassifier)
model.set_info(DATA_NAME, f"{BENCHMARK_NAME}/learning_curve", i_cv)
return model
def build_model(args, i_cv):
args.net = ARCHI(n_in=29, n_out=2, n_extra=3, n_unit=args.n_unit)
args.optimizer = get_optimizer(args)
model = get_model(args, FilterRegressor)
model.set_info(DATA_NAME, BENCHMARK_NAME, i_cv)
return model
def build_model(args, i_cv):
args.net = ARCHI(n_in=31, n_out=2, n_unit=args.n_unit)
args.optimizer = get_optimizer(args)
model = get_model(args, Model)
model.set_info(DATA_NAME, BENCHMARK_NAME, i_cv)
return model
def train(self, train_loader, dev_loader, eval_loader):
self._init_embeddings(self.token2id)
# 对数据集按照长度进行排序
epoch = 1 # epoch
patience_count = 0
start = time.time()
while True:
self.step = 0
losses = 0.
if epoch == self.embedding_le:
print("Start Learning Embedding!")
self.optimizer = change_embedding_lr(self.optimizer,
self.embedding_lr)
for articles, tags, lengths in train_loader:
losses += self.train_step(articles, tags, lengths)
if self.step % TrainingConfig.print_step == 0:
total_step = len(train_loader)
print(
"Epoch {}, step/total_step: {}/{} {:.2f}% Loss:{:.4f}".
format(epoch, self.step, total_step,
100. * self.step / total_step,
losses / self.print_step))
losses = 0.
# 每轮结束测试在验证集上的性能，保存最好的一个
val_loss = self.validate(dev_loader)
print("Epoch {}, Val Loss:{:.4f}".format(epoch, val_loss))
# 计算验证集上的f1分数，保存最好的一个
metrics = self.cal_scores(eval_loader)
metrics.report()
if metrics.final_f1 > self.best_f1_score:
print("更新f1并保存模型...")
self.best_metrics = metrics
self.best_f1_model = copy.deepcopy(self.model)
assert id(self.best_f1_model) != id(self.model) # 确认有复制
self.best_f1_score = metrics.final_f1
patience_count = 0 # 有改进，patience_count清零
else:
patience_count += 1
# 如果连续lr_patience个回合f1分数下降,将当前模型换成best_f1_model
# 并且重新初始化学习器 ,降低学习率为原来的一半
if patience_count >= self.lr_patience:
print("Reduce Learning Rate....")
self.embedding_lr *= self.lr_decay
self.lr *= self.lr_decay
self.optimizer = get_optimizer(self.model,
self.embedding_lr, self.lr)
self.model = copy.deepcopy(self.best_f1_model)
# 如果连续patience个回合分数下降，则结束
if patience_count >= self.patience:
end = time.time()
print("在Epoch {} 训练终止,用时: {}".format(
epoch, time_format(end - start)))
break # 终止训练
epoch += 1
def __init__(self, vocab_size, out_size, token2id, tag2id, method="lstm"):
"""功能：对LSTM的模型进行训练与测试
参数:
vocab_size:词典大小
out_size:标注种类
method: 三种方法，["lstm", "lstm_crf", "lstm_lstm"]
crf选择是否添加CRF层"""
self.method = method
self.out_size = out_size
self.token2id = token2id
self.tag2id = tag2id
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
assert method in ["lstm", "lstm_crf", "lstm_lstm"]
# 加载模型参数
self.emb_size = LSTMConfig.emb_size
self.max_len = TrainingConfig.max_len
self.clip = TrainingConfig.clip
# 根据是否添加crf初始化不同的模型 选择不一样的损失计算函数
if method == "lstm":
self.hidden_size = LSTMConfig.hidden_size
self.model = BiLSTM(vocab_size, self.emb_size, self.hidden_size,
out_size).to(self.device)
self.alpha = TrainingConfig.alpha
self.cal_loss_func = cal_weighted_loss
elif method == "lstm_crf":
self.hidden_size = LSTMConfig.hidden_size
self.model = BiLSTM_CRF(vocab_size, self.emb_size,
self.hidden_size, out_size).to(self.device)
self.cal_loss_func = cal_lstm_crf_loss
elif method == "lstm_lstm":
self.enc_hsize = LSTMConfig.enc_hidden_size
self.dec_hsize = LSTMConfig.dec_hidden_size
self.dropout_p = LSTMConfig.dropout_p
self.model = DoubleLSTM(vocab_size, self.emb_size, self.enc_hsize,
self.dec_hsize, out_size, self.dropout_p)
self.alpha = TrainingConfig.alpha
self.cal_loss_func = cal_weighted_loss
self.model = self.model.to(self.device)
# 加载训练参数：
self.patience = TrainingConfig.training_patience
self.print_step = TrainingConfig.print_step
self.lr = TrainingConfig.lr
self.lr_patience = TrainingConfig.lr_patience
self.lr_decay = TrainingConfig.lr_decay
self.batch_size = TrainingConfig.batch_size
self.decoding_batch_size = TrainingConfig.decoding_batch_size
self.embedding_lr = TrainingConfig.embedding_lr
self.embedding_le = TrainingConfig.embedding_learning_epoch
self.token_method = TrainingConfig.token_method
# 初始化优化器 一开始embedding层学习速率设为0
self.optimizer = get_optimizer(self.model, 0.0, self.lr)
# 初始化其他指标
self.step = 0
self.best_val_loss = 1e18
self.best_model = None
# 最好的分数(f1分数)
self.best_metrics = None
self.best_f1_model = None
self.best_f1_score = 0.