ValueError：column_name： input_tensor dtype必须为字符串或整数。 dtype：

Question

[监督分类]我正在尝试使用张量流和keras训练具有许多不同分类数据的模型。我无法使用“一键编码”，因为有数百种不同的值。因此，我尝试创建一个feature_columncategorical_column_with_hash_bucket，然后将其变成feature_column.embedding_column因此，我的数据中的字符串值被转换为整数，然后转换为3维浮点向量。训练时出现错误

ValueError: in converted code:
    relative to C:\Users\kremer\Anaconda3\lib\site-packages\tensorflow\python\feature_column:

    feature_column_v2.py:474 call
        self._state_manager)
    feature_column_v2.py:3121 get_dense_tensor
        transformation_cache, state_manager)
    feature_column_v2.py:3488 get_sparse_tensors
        transformation_cache.get(self, state_manager), None)
    feature_column_v2.py:2562 get
        transformed = column.transform_feature(self, state_manager)
    feature_column_v2.py:3466 transform_feature
        return self._transform_input_tensor(input_tensor)
    feature_column_v2.py:3444 _transform_input_tensor
        prefix='column_name: {} input_tensor'.format(self.key))
    utils.py:58 assert_string_or_int
        '{} dtype must be string or integer. dtype: {}.'.format(prefix, dtype))

    ValueError: column_name: Artikel input_tensor dtype must be string or integer. dtype: <dtype: 'float32'>.

这是我的代码：

#defining feature columns:

feature_columns = []

# numeric cols
for header in ['POS', 'DAUER_RUEST', 'UNTERBRECHUNGEN_RUEST', 'DAUER_PROD', 'UNTERBRECHUNGEN_PROD', 'GUTTEILE', 'Teile_Soll', 'Stueckzeit', 'Ruestzeit_Soll']:
  feature_columns.append(feature_column.numeric_column(header))

# categorical columns with embedding
artikel = feature_column.categorical_column_with_hash_bucket(key='Artikel' , hash_bucket_size=600, dtype=tf.dtypes.string)
artikel_embedding = feature_column.embedding_column(artikel, dimension=3)
feature_columns.append(artikel_embedding)

batchnumber = feature_column.categorical_column_with_hash_bucket(key='BA' , hash_bucket_size=600, dtype=tf.dtypes.string)
batchnumber_embedding = feature_column.embedding_column(batchnumber, dimension=3)
feature_columns.append(batchnumber_embedding)

...

#five embedding columns with this design in total

...

#building and training the model

model = tf.keras.Sequential()
model.add(feature_layer)
model.add(layers.Dense(28, activation='relu'))
model.add(layers.Dense(28, activation='relu'))
model.add(layers.Dense(1))

early_stopping = tf.keras.callbacks.EarlyStopping(patience=2)

model.compile(optimizer='adam',
              loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
              metrics=['accuracy'])

early_stopping = tf.keras.callbacks.EarlyStopping(patience=3)

model.fit(train_ds,
          validation_data=val_ds,
          epochs=5,
          callbacks=[early_stopping],
          verbose = 1,
         )

Answer 1

更改

[artikel = feature_column.categorical_column_with_hash_bucket(key='Artikel' , hash_bucket_size=600, dtype=tf.dtypes.string)至

artikel = feature_column.categorical_column_with_hash_bucket(key='Artikel' , hash_bucket_size=600, dtype=tf.dtypes.float)

因为您在categorical_column_with_hash_bucket中将artikel定义为string。一直以来，我对Keras都不熟悉，我认为在model.fit中，artikel中的train_ds是float的实例。在tensorflow估计器中，tf.estimator.TrainSpec input_fn需要特定的dtype。