Many problems in the field of deep learning involve the processing of the sequence. From Wiki’s definition, sequences can represent a variety of physical meanings, but in deep learning, the most common is still “time sequence” - a sequence containing information of multiple time steps.
In Paddle Fluid, we represent the sequence as api_fluid_LoDTensor. Because the general neural network performs computing batch by batch, we use a LoDTensor to store a mini batch of sequences. The 0th dimension of a LoDTensor contains all the time steps of all sequences in the mini batch, and LoD is used to record the length of each sequence to distinguish different sequences. In the calculation, it is also necessary to split the 0th dimension of a mini batch in the LoDTensor into a number of sequences according to the LoD information. (Please refer to the LoD related documents for details. ) Therefore, the operation for the 0th dimension of LoDTensor cannot be performed simply by a general layer. The operation of this dimension must be combined with the information of LoD. (For example, you can’t reshape the 0th dimension of a sequence with
In order to correctly implement various sequence-oriented operations, we have designed a series of sequence-related APIs. In practice, because a LoDTensor contains a mini batch of sequences, and different sequences in the same mini batch usually belong to multiple samples, they do not and should not interact with each other. Therefore, if a layer is input with two (or more) LoDTensors (or with a list of LoDTensors), and each LoDTensor represents a mini batch of sequences, the first sequence in the first LoDTensor will be only calculated with the first sequence in the second LoDTensor, and the second sequence in the first LoDTensor will only be calculated with the second sequence in the second LoDTensor. To conclude with, the i’th sequence in the first LoDTensor will only be calculated with the i’th sequence in the second LoDTensor, and so on.
In summary, a LoDTensor stores multiple sequences in a mini batch, where the number of sequences is batch size; when multiple LoDTensors are calculated, the i’th sequence in each LoDTensor will only be calculated with the i’th of the other LoDTensors. Understanding this is critical to understand the following associated operations.
This layer takes a mini batch of sequences as input and does a softmax operation in each sequence. The output is a mini batch of sequences in the same shape, but it is normalized by softmax within the sequence. This layer is often used to do softmax normalization within each sequence.
Please refer to api_fluid_layers_sequence_softmax
The layer takes a list as input, which can contain multiple LoDTensors, and every LoDTensors is a mini batch of sequences. The layer will concatenate the i-th sequence in each batch into a new sequence in the time dimension as the i’th sequence in the returned batch. Of course, the sequences of each LoDTensor in the list must have the same batch size.
Please refer to api_fluid_layers_sequence_concat
This layer takes a LoDTensor as input and takes the first element in each sequence (the element of the first time step) as the return value.
Please refer to api_fluid_layers_sequence_first_step
sequence_first_step except that this layer takes the last element in each sequence (i.e. the last time step) as the return value.
Please refer to api_fluid_layers_sequence_last_step
This layer has two LoDTensors of sequences as input and extends the sequence in the first batch according to the LoD information of the sequence in the second LoDTensor. It is usually used to extend a sequence with only one time step (for example, the return result of
sequence_first_step) into a sequence with multiple time steps, which is convenient for calculations with sequences composed of multiple time steps.
Please refer to api_fluid_layers_sequence_expand
This layer takes two LoDTensors of sequences as input and then extends each sequence in the first Tensor to a sequence with the same length as the corresponding one in the second Tensor. Unlike
sequence_expand , this layer will strictly extend the sequence in the first LoDTensor to have the same length as the corresponding one in the second LoDTensor. If it cannot be extended to the same length (for example, the sequence length of the second batch is not an integer multiple of the sequence length of the first batch), an error will be reported.
Please refer to api_fluid_layers_sequence_expand_as
This layer takes a LodTensor of sequences as input and also specifies the length of a
win_size. This layer will take a subsequence of length
win_size in all sequences and combine them into a new sequence.
Please refer to api_fluid_layers_sequence_enumerate
This layer requires a LoDTensor of sequences as input, and you need to specify a
new_dim as the dimension of the new sequence. The layer will reshape each sequence in the mini batch to the dimension given by new_dim. Note that the length of each sequence will be changed (so does the LoD information) to accommodate the new shape.
Please refer to api_fluid_layers_sequence_reshape
This layer can scatter a sequence of data onto another tensor. This layer has three inputs, one is a target tensor to be scattered
input; One is the sequence of data to scatter
update ; One is the upper coordinate of the target tensor
index . Output is the tensor after scatter, whose shape is the same as
Please refer to api_fluid_layers_sequence_scatter
This layer can pad sequences of unequal length into equal-length sequences. To use this layer you need to provide a
PadValue and a
padded_length. The former is the element used to pad the sequence, it can be a number or a tensor; the latter is the target length of the sequence. This layer will return the padded sequence, and a tensor
Length of the length for each sequence before padding.
Please refer to api_fluid_layers_sequence_pad
This layer will generate a mask based on
input, where the
input is a tensor that records the length of each sequence. In addition, this layer requires a parameter
maxlen to specify the largest sequence length in the sequence. Usually, this layer is used to generate a mask that will filter away the portion of the paddings in the sequence. The
input tensor can usually directly use the returned
Please refer to api_fluid_layers_sequence_mask