tf.data WindowDataset flat_map给出'dict'对象没有属性'batch'错误

问题描述

我正在尝试处理(batch_size,time_steps,my_data)类型的批次

为什么在flat_map步骤上我得到AttributeError: 'dict' object has no attribute 'batch'

 x_train = np.random.normal(size=(60000,768))
    token_type_ids = np.ones(shape=(len(x_train)))
    position_ids = np.random.normal(size=(x_train.shape[0],5))

    features_ds = tf.data.Dataset.from_tensor_slices({'inputs_embeds': x_train,'token_type_ids': token_type_ids,'position_ids': position_ids})
    y_ds = tf.data.Dataset.from_tensor_slices(y_train)
    ds = tf.data.Dataset.zip((features_ds,y_ds))
    # result = list(ds.as_numpy_iterator())

    result_ds = ds.window(size=time_steps,shift=time_steps,stride=1,drop_remainder=True). \
        flat_map(lambda x,y: tf.data.Dataset.zip((x.batch(time_steps),y.batch(time_steps))))

任何主意是什么问题？以及如何解决？

解决方法

您可以将批次添加为单独的步骤：

x_train = np.random.normal(size=(60000,768))
token_type_ids = np.ones(shape=(len(x_train)))
position_ids = np.random.normal(size=(x_train.shape[0],5))

features_ds = tf.data.Dataset.from_tensor_slices({'inputs_embeds': x_train,'token_type_ids': token_type_ids,'position_ids': position_ids})
y_train = np.random.normal(size=(60000,1))
y_ds = tf.data.Dataset.from_tensor_slices(y_train)
ds = tf.data.Dataset.zip((features_ds,y_ds))

result_ds = ds.window(size=time_steps,shift=time_steps,stride=1,drop_remainder=True).\
    flat_map(lambda x,y: tf.data.Dataset.zip((x,y)))

time_steps=3
result_ds=result_ds.batch(time_steps)

for i in result_ds.take(1):
    print(i)