无法通过使用自定义块复制自动keras模型

如何解决无法通过使用自定义块复制自动keras模型

我正在使用StructuredDataRegressor进行一些预测。我假设生成的“ oracle.json”描述了 模型架构和超参数，以找到最佳模型。这就是“ oracle.json”的相关部分所说的：

"values": {
  "structured_data_block_1/normalize": false,"structured_data_block_1/dense_block_1/num_layers": 2,"structured_data_block_1/dense_block_1/use_batchnorm": false,"structured_data_block_1/dense_block_1/dropout": 0,"structured_data_block_1/dense_block_1/units_0": 32,"structured_data_block_1/dense_block_1/units_1": 32,"regression_head_1/dropout": 0,"optimizer": "adam","learning_rate": 0.001,"structured_data_block_1/dense_block_1/units_2": 32
}

我该如何使用Blocks复制代码，这样我只需要使用这些值精确地训练一次？

这是我尝试过的一些东西。

请注意，我尝试的每种方法都是分别尝试的。我没有尝试一起运行所有这些代码。这里都在一起，所以我不必打开多个代码部分。

如果这些代码节中的任何一个都可以工作，请告诉我，我可以详细说明确切的失败或不正常的地方。

# I assume I always need this
input_node = ak.StructuredDatainput()

# Played around with StructuredDataBlock,but there is no way to
#   pass `num_layers`,`use_batchnorm`,`dropout` to the
#   DenseBlock that it internally uses.
output_node = ak.StructuredDataBlock(
    normalize=False,# **{
    #     "num_layers": 2,#     "use_batchnorm": False,#     "dropout": 0,# }
    # num_layers=2,# use_batchnorm=False,# dropout=0,)(input_node)

# Tried replicate what StructuredDataBlock is doing. It runs,#   but the resulting "oracle.json" does not look like the
#   original one,and it does not respect
#   `num_layers`,`dropout`.
output_node = ak.CategoricalToNumerical()(input_node)
output_node = ak.normalization()(output_node)
output_node = ak.DenseBlock(
    num_layers=2,use_batchnorm=False,dropout=0.0,)(output_node)

# Tried to use DenseBlock exactly the way it's used in
#   "StructuredDataBlock",but that also didn't work.
hp = HyperParameters()
output_node = input_node
block = ak.CategoricalToNumerical()
output_node = block.build(hp,output_node)
output_node = ak.DenseBlock().build(hp,output_node)

# This is the regression head,which I assume I will always need anyway.
output_node = ak.RegressionHead(
    loss=DimaLoss(),metrics=[LOSS_REGISTRY[self.loss_name]],)(output_node)

# AutoModel
auto_model = ak.AutoModel(
    project_name=PROJECT_NAME,directory=dima_trainer_dir,overwrite=True,max_trials=self.max_trials,inputs=input_node,outputs=output_node,objective=kerastuner.Objective(f"val_{self.loss_name}",direction="min"),)

# fit
auto_model.fit(
    x=df_train_X,y=df_train_y,epochs=self.epochs,callbacks=[early_stopping],validation_split=0.15,**{
        "batch_size": self.batch_size,# "optimizer": "adam",# "learning_rate": 0.002,},# optimizer="adam",# learning_rate=0.002,)