Trait relearn::torch::seq_modules::SequenceModule[−][src]

pub trait SequenceModule {
    fn seq_serial(&self, inputs: &Tensor, seq_lengths: &[usize]) -> Tensor;
    fn seq_packed(&self, inputs: &Tensor, batch_sizes: &Tensor) -> Tensor;
}

Expand description

A network module that operates on a sequence of data.

Required methods

[src]

fn seq_serial(&self, inputs: &Tensor, seq_lengths: &[usize ]) -> Tensor

Apply the network over multiple sequences arranged in series one after another.

input.i(.., ..seq_lengths[0], ..) is the first batch of sequences, input.i(.., seq_lengths[0]..(seq_lengths[0]+seq_lengths[1]), ..) is the second, etc.

Args

inputs - Batched input sequences arranged in series. An f32 tensor of shape [BATCH_SIZE, TOTAL_SEQ_LENGTH, NUM_INPUT_FEATURES]
seq_lengths - Length of each sequence. The sequence length is the same across the batch dimension.

Returns

Batched output sequences arranged in series. A tensor of shape [BATCH_SHAPE, TOTAL_SEQ_LENGTH, NUM_OUTPUT_FEATURES].

[src]

fn seq_packed(&self, inputs: &Tensor, batch_sizes: &Tensor) -> Tensor

Apply the network over multiple sequences packed together in heterogeneous batches.

input.i(0..batch_sizes[0], ..) are the batched first steps of all sequences, input.i(batch_sizes[0]..(batch_sizes[0]+batch_sizes[1]), ..) are the second steps, etc.

Args

inputs - Packed input sequences. An f32 tensor of shape [TOTAL_STEPS, NUM_INPUT_FEATURES] where the TOTAL_STEPS dimension consists of the packed and batched steps ordered first by index within a sequence, then by batch index. Sequences must be ordered from longest to shortest.

If all sequences have the same length then the TOTAL_STEPS dimension corresponds to a flattend Tensor of shape [SEQ_LENGTH, BATCH_SIZE].
batch_sizes - The batch size of each in-sequence step index. A i64 tensor of shape [MAX_SEQ_LENGTH]. Must be on the CPU. Must be monotonically decreasing and positive.