optuna在深度学习上自动调参

1.背景

最近在烦恼怎么对深度学习进行调参，发现在optuna上可以实现。

optuna可以和主流的机器学习框架进行融合，然后进行调参。同时调参后，还有查看结果对比的功能。

2.安装

直接使用pip进行安装:

pip install optuna

3.在tensorflow上使用

首先看一下官方案例：

import tensorflow as tf
import optuna

# 1. Define an objective function to be maximized.
def objective(trial):

    # 2. Suggest values of the hyperparameters using a trial object.
    n_layers = trial.suggest_int('n_layers', 1, 3)
    model = tf.keras.Sequential()
    model.add(tf.keras.layers.Flatten())
    for i in range(n_layers):
        num_hidden = trial.suggest_int(f'n_units_l{i}', 4, 128, log=True)
        model.add(tf.keras.layers.Dense(num_hidden, activation='relu'))
    model.add(tf.keras.layers.Dense(CLASSES))
    ...
    return accuracy

# 3. Create a study object and optimize the objective function.
study = optuna.create_study(direction='maximize', sampler=optuna.samplers.TPESampler())
study.optimize(objective, n_trials=100)

案例上分为三步：

• 1）定义objective：这个函数主要使定义“模型构建”，“模型训练”，“参数更新”等；同时需要拿到一个值，这里是accuracy，也就是保证准确率最大化maximize。

• 2）传入参数hyperparameters ：用来控制哪些参数需要进行遍历，来找到最优的参数。

• 3）定义study：

direction

控制是以最大化还是最小化为准

sampler

指示您希望 Optuna 实施的采样器方法。

4.实例

4.1 定义模型、训练过程

class Trainer(object):
    def __init__(self,
                transformer_encoder,
                discriminator_class,
                optimizer,
                epochs,
                train_dataset_white,
                val_dataset_white
                ) -> None:
        super(Trainer, self).__init__()

        self.train_dataset_white = train_dataset_white
        self.val_dataset_white = val_dataset_white
        self.epochs = epochs

        self.transformer_encoder = transformer_encoder
        self.discriminator_class = discriminator_class
        self.optimizer = optimizer

    train_step_signature = [
        tf.TensorSpec(shape=(None, duel_seq_nums, seq_len), dtype=tf.float32),
        tf.TensorSpec(shape=(None, sta_len), dtype=tf.float32),
        tf.TensorSpec(shape=(None, 1), dtype=tf.int32),
    ]

    @tf.function(input_signature=train_step_signature)
    def train_step(self, inp_seq, inp_sta, label):

        with tf.GradientTape() as tape:
            enc_output = self.transformer_encoder(inp_seq, training = True)
            predictions = self.discriminator_class(enc_output, inp_sta)
            loss = loss_function(label, predictions)

        trainable_variables = self.transformer_encoder.trainable_variables + self.discriminator_class.trainable_variables
        gradients = tape.gradient(loss, trainable_variables)
        self.optimizer.apply_gradients(zip(gradients, trainable_variables))

        train_loss_class(loss)

        predictions_label = tf.where(predictions >= 0.5, 1, 0)
        train_precision_class(label, predictions_label)
        train_recall_class(label, predictions_label)

        return predictions

    @tf.function(input_signature=train_step_signature)
    def val_step(self, inp_seq, inp_sta, label):

        enc_output = self.transformer_encoder(inp_seq, training = False)
        predictions = self.discriminator_class(enc_output, inp_sta)
        loss = loss_function(label, predictions)

        val_loss_class(loss)
        
        predictions_label = tf.where(predictions >= 0.5, 1, 0)
        val_precision_class(label, predictions_label)
        val_recall_class(label, predictions_label)

        return predictions
        
    def train(self):
        # checkpoint and tensorboard
        ckpt = tf.train.Checkpoint(transformer_encoder=self.transformer_encoder, discriminator_class=self.discriminator_class, optimizer=self.optimizer)
        current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
        _checkpoint_path = checkpoint_path + "_" + current_time
        # ckpt_manager = tf.train.CheckpointManager(ckpt, _checkpoint_path, max_to_keep=5)
        train_summary_writer, val_summary_writer = tensorboard()

        # # 如果检查点存在，则恢复最新的检查点。
        # if ckpt_manager.latest_checkpoint:
        #     ckpt.restore(ckpt_manager.latest_checkpoint)
        #     logging.info('Latest checkpoint restored!!')
            
        discriminator_step = 0
        max_val_f1 = 0

        discriminator_step = 0
        for epoch in range(10):
            start_epoch = time.time()
            train_loss_class.reset_states()
            train_precision_class.reset_states()
            train_recall_class.reset_states()
            val_loss_class.reset_states()
            val_precision_class.reset_states()
            val_recall_class.reset_states()

            # Train
            for start_index, parsed_dataset in enumerate(train_dataset_white):
                seq, sta, l = rd.read_dataset(parsed_dataset)

                predictions = self.train_step(seq, sta, l)
                discriminator_step += 1
                f1_score = (2 * train_precision_class.result() * train_recall_class.result()) / (train_precision_class.result() + train_recall_class.result())
                with train_summary_writer.as_default():
                    # tf.summary.scalar('Generator loss', train_loss_gen_seq.result(), step=epoch)
                    tf.summary.scalar('Class loss', train_loss_class.result(), step=discriminator_step)
                    tf.summary.scalar('precision', train_precision_class.result(), step=discriminator_step)
                    tf.summary.scalar('recall', train_recall_class.result(), step=discriminator_step)
                    tf.summary.scalar('F1', f1_score, step=discriminator_step)

                if start_index % 1000 == 0:
                    logging.info('Epoch {:2d} | Index {:5d} | Class Loss {:.4f} | Secs {:.4f}'.format(epoch, start_index, train_loss_class.result(), time.time() - start_epoch))
                    if start_index == 2000:
                        break

            logging.info('Epoch {:2d} | Train Loss {:.4f} | Precision {:.4f} Recall {:.4f} F1 {:.4f}'.format(epoch, train_loss_class.result(),
                                                            train_precision_class.result(),
                                                            train_recall_class.result(),
                                                            f1_score))

            # Val
            val_predicted_all = np.zeros([0, 1])
            val_label = np.zeros([0, 1])
            for start_index, parsed_dataset in enumerate(val_dataset_white):
                seq, sta, l = rd.read_dataset(parsed_dataset)

                predictions = self.val_step(seq, sta, l)

                val_predicted_all = np.concatenate([val_predicted_all, np.reshape(predictions.numpy(), [-1, 1])], axis=0)
                val_label = np.concatenate([val_label, np.reshape(l.numpy(), [-1, 1])], axis=0)

            pre, rec, val_f1, _ = precision_recall_fscore_support(val_label, np.where(val_predicted_all >= 0.5, 1, 0))
            val_f1 = val_f1[1]

            with val_summary_writer.as_default():
                tf.summary.scalar('Class loss', val_loss_class.result(), step=epoch)
                tf.summary.scalar('precision', val_precision_class.result(), step=epoch)
                tf.summary.scalar('recall', val_recall_class.result(), step=epoch)
                tf.summary.scalar('F1', val_f1, step=epoch)

            logging.info('Epoch {:2d} | Val Loss {:.4f} | Precision {:.4f} Recall {:.4f} F1 {:.4f}'.format(epoch, val_loss_class.result(),
                                                            val_precision_class.result(),
                                                            val_recall_class.result(),
                                                            val_f1))

            if max_val_f1 <= val_f1:
                max_val_f1 = val_f1
                # ckpt_save_path = ckpt_manager.save()
                # logging.info('Saving checkpoint for epoch {} at {}'.format(epoch+1, ckpt_save_path))
                logging.info("xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx")

            logging.info("Finish Epoch in %f \n" % (time.time() - start_epoch))
        return max_val_f1

Trainer

• transformer_encoder、discriminator_class：模型
• optimizer：优化器
• train_dataset_white、val_dataset_white：训练集和验证集

train_step

训练过程

val_step

验证过程

train

启动模型训练

4.2 objective函数定义

    def build_model(params):
        # model
        transformer_encoder = TransformerEncoder(
                        num_layers=params["num_layers"],
                        d_model=params["d_model"],
                        num_heads=hp.num_heads,
                        dff=params["dff"],
                        input_vocab_size=20,
                        target_vocab_size=22,
                        pe_input=100,
                        pe_target=100,
                        target_feature_len=21,
                        rate=params["dropout_rate"])
        discriminator_class = DecoderClass()

        optimizer = tf.keras.optimizers.Adam(params["learning_rate"])

        logging.info('Model bulid: %f' % (time.time() - start_time))
        
        trainer = Trainer(
            transformer_encoder=transformer_encoder,
            discriminator_class=discriminator_class,
            optimizer=optimizer,
            epochs=EPOCHS,
            train_dataset_white=train_dataset_white,
            val_dataset_white=val_dataset_white,
        )
        return trainer

def objective(trial):

     params = {
              'learning_rate': trial.suggest_float('learning_rate', 1e-4, 1e-3),
              'num_layers': trial.suggest_categorical('num_layers', [1, 2, 3, 4]),
              'd_model': trial.suggest_categorical('d_model', [64, 128, 256, 512]),
              'dff': trial.suggest_int('dff', 128, 1024),
              'dropout_rate': trial.suggest_categorical('dropout_rate', [0.0, 0.1, 0.2])
              }
    
     trainer = build_model(params)
    
     max_val_f1 = trainer.train()

     return max_val_f1

params

传入需要控制的参数

4.3 启动optuna

study = optuna.create_study(direction="maximize", sampler=optuna.samplers.TPESampler(), storage='sqlite:///db.sqlite3')
study.optimize(objective, n_trials=50)

storge

其中可以保存模型调参的结果

5.图形化显示

从上面保存文件db.sqlite3，然后再安装:optuna-dashboard

pip install optuna-dashboard

接着启动命令：

optuna-dashboard sqlite:///db.sqlite3 

然后在浏览器输入：127.0.0.1:8080，就可以看到具体的图像结果：

哪个参数比较重要：