x-j6ddlZddlmZddlmZgZdZdZdS)N)core)DataToDenseTensorConvertercnt|tjr#tt j|||St|t rt|t|gtj j j }g}|D]9}| t|| |:|g|ks Jdt j|j}|g|jdR}tj}||||||St|tjr@tj}||||||St)d)a Create a DenseTensor from a numpy array, list or existing DenseTensor. The implementation is as follows: 1. Check whether the length-based LoD, i.e., :code:`recursive_seq_lens` is valid. 2. Convert :code:`recursive_seq_lens` to a offset-based LoD. 3. Based on :code:`place` , copy the :code:`data` from a numpy array, list or existing DenseTensor to CPU or GPU device. 4. Set offset-based LoD to the output DenseTensor. Suppose we want to create a DenseTensor to hold data for word sequences, where each word is represented by an integer. If we want to create a DenseTensor to represent two sentences, one of 2 words, and one of 3 words. Then :code:`data` would be a numpy array of integers with shape (5, 1). :code:`recursive_seq_lens` would be [[2, 3]], indicating the word number in each sentence. This length-based :code:`recursive_seq_lens` [[2, 3]] would be converted to offset-based LoD [[0, 2, 5]] inside the function call. Args: data (numpy.ndarray|list|DenseTensor): a numpy array, a list or ad DenseTensor holding the data to be copied. recursive_seq_lens (list[list[int]]): a list of lists indicating the length-based LoD info. place (CPUPlace|CUDAPlace): CPU or GPU place indicating where the data in the created DenseTensor will be stored. Returns: A DenseTensor with tensor data and recursive_seq_lens info. Examples: .. code-block:: python >>> import paddle.base as base >>> import numpy as np >>> t = base.create_lod_tensor(np.ndarray([5, 30]), [[2, 3]], base.CPUPlace()) )place lod_levelshapedtypez(data and recursive_seq_lens do not matchrz==hy~&& kk/CI/q//**!## 3--.@AAA D"* % % !## 4--.@AAA  J   ct|ts Jdt|dg|}tj|||d}t|||S)aL :api_attr: Static Graph Create a DenseTensor containing random integers. The implementation is as follows: 1. Obtain the shape of output DenseTensor based on :code:`recursive_seq_lens` and :code:`base_shape` . The first dimension of the shape is the total length of sequences, while the other dimensions are the same as :code:`base_shape` . 2. Create a numpy array of random integers, and parse the created numpy array as parameter :code:`data` of :ref:`api_paddle_base_create_lod_tensor` to create the output DenseTensor. Suppose we want to create a DenseTensor to hold data for 2 sequences, where the dimension of the sequences are [2, 30] and [3, 30] respectively. The :code:`recursive_seq_lens` would be [[2, 3]], and :code:`base_shape` would be [30] (the other dimensions excluding the sequence length). Therefore, the shape of the output DenseTensor would be [5, 30], where the first dimension 5 is the total lengths of the sequences, and the other dimensions are :code:`base_shape`. Args: recursive_seq_lens (list[list[int]]): a list of lists indicating the length-based LoD info. base_shape (list[int]): the shape of the output DenseTensor excluding the first dimension. place (CPUPlace|CUDAPlace): CPU or GPU place indicating where the data in the created DenseTensor will be stored. low (int): the lower bound of the random integers. high (int): the upper bound of the random integers. Returns: A DenseTensor with tensor data and recursive_seq_lens info, whose data is inside [low, high]. Examples: .. code-block:: pycon >>> import paddle.base as base >>> t = base.create_random_int_lodtensor( ... recursive_seq_lens=[[2, 3]], ... base_shape=[30], ... place=base.CPUPlace(), ... low=0, ... high=10, ... ) >>> print(t.shape()) paddle.Size([5, 30]) zbase_shape should be a listint64)r rsumrrandomrandom_integersastyper )r base_shaperlowhigh overall_shapers r#create_random_int_lodtensorr0psyr j$ ' 'FF)FFF '+B/00>:>M 9 $ $S$ > > E Eg N ND T#5u = ==r$) numpyrr data_feederr__all__r r0r$r#r6sh333333 V V V r>>>>>>>>>>r$