为什么DenseFeatures影响模型结果？

如何解决为什么DenseFeatures影响模型结果？

我正在研究ANN预测模型。过去，我不使用DenseFeatures来创建模型。现在，在SAME数据集和图层设置下，我尝试使用DenseFeatures创建模型以替换以前的预处理。但是模型结果有很大的不同。而且我希望这应该没什么不同，因为使用DesnseFeature可以帮助我控制/监视输入变量。

上一个方法：

# (after preprocessing,e.g one-hot encoder [0,1])
# (df -> array)

BATCH_SIZE = 100
EPOCHS = 1000

model = keras.Sequential()
[layer setting]
...
[later setting]

model.fit(T_x_train,T_y_train,batch_size=BATCH_SIZE,epochs=EPOCHS,verbose=1,validation_data=(T_x_valid,T_y_valid),callbacks = [es])

DenseFeature方法（在模型内部）

参考来自 https://www.tensorflow.org/tutorials/structured_data/feature_columns#create_target_variable

# (setting up the featureColumn for  DenseFeature Layer)
# (For example:)
feature_columns = []
# Numeric Columns
numeric_cols = ['x1','x2']
for header in numeric_cols :
  feature_columns.append(feature_column.numeric_column(header))

# # Categorical Columns
# # Hour
hour_type = feature_column.categorical_column_with_vocabulary_list(
            'hour',[hr for hr in range(24)])
hour_type_one_hot = feature_column.indicator_column(hour_type)
feature_columns.append(hour_type_one_hot)

feature_layer = tf.keras.layers.DenseFeatures(feature_columns,trainable = False)

# (Setting up the training dataset for tf format) 
def df_to_dataset(dataframe,target,shuffle=False,batch_size=100):
    dataframe = dataframe.copy()
    labels = dataframe.pop(target) # Copy and delete column in df
    ds = tf.data.Dataset.from_tensor_slices((dict(dataframe),labels))
    if shuffle:
        ds = ds.shuffle(buffer_size=len(dataframe))
    ds = ds.batch(batch_size)
    return ds

TRAIN = df_to_dataset(train_full_model_df,target = 'reading')
VALID = df_to_dataset(valid_full_model_df,target = 'reading')

# (model part)
model = keras.Sequential()
model.add(feature_layer)

[layer setting]
...
[later setting]

model.fit(TRAIN,validation_data=(VALID),callbacks = [es])

DenseFeature方法（在模型外部）

# (setting up the featureColumn for  DenseFeature Layer)
# (For example:)
feature_columns = []
# Numeric Columns
numeric_cols = ['x1',trainable = False)


# (model part)
model = keras.Sequential()

[layer setting]
...
[later setting]

model.fit(feature_layer(dict(train_full_model_df)),train_full_model_df['reading'].values,validation_data=(feature_layer(dict(valid_full_model_df)),valid_full_model_df['reading'].values),callbacks = [es])

我已经使用feature_layer(dict(train_full_model_df))对数组/输入值进行了交叉检查，它确认我们正在训练和验证数据上使用样本长度和值。

在Previous Method和DenseFeature Method (Outside the model)中可以产生相似的结果，但是DenseFeature Method (Inside the model)会产生很大的不同甚至更糟的结果。

请帮助，谢谢。我希望在模型中添加DenseFeatures的计算应该没有什么不同。