[feature request] More methods for PackedSequence

[dmitriy-serdyuk]

As a person who works a lot with recurrent networks and sequences, I wish it was easier to work with PackedSequence. I frequently find myself packing/unpacking sequence to perform some simple operation such as

sequence, lengths = pad_packed_sequence(input, batch_first=True)
sequence = elementwise_function(sequence)
pack_padded_sequence(sequence, lengths, batch_first=True)

This complicates code, introduces bugs (setting batch_first incorrectly), and makes me write two versions of code for tensors and sequences.

PaddedSequence can be extended to act more like torch.Tensor. More specifically, it can be extended with following methods (I tried to choose ones which make sense):

Element-wise unary operations (wrt to the tensor)

• abs, acos, asin, atan, atan2, ceil, clamp, contiguous, cos, cosh, exp, expm1, frac, log, log10, log1p, log2, mul, pow, reciprocal, neg, renorm, round, rsqrt, sigmoid, sign, sin, sinh, sqrt, tan, tanh, trunc
• fill_, pin_memory

Boolean functions

• is_contiguous, is_pinned

Element-wise binary operations

Two sequences can be added, subtracted, compared, etc when they have same shape and lengths.

• add, div, eq, fmod, ge, gt, le, lt, ne, remainder, sub
• map_

Autograd functions

• detach

cc @albanD @mruberry @jbschlosser

[t-vi]

The binary operations are outright dangerous unless you implement erring out when feeding sequences of different length.
Personally, I use an apply method

def elementwise_applyi(fn, arg):
    return torch.nn.utils.rnn.PackedSequence(fn(arg.data, args[0].batch_sizes))

def elementwise_apply_multi(fn, *args):
    # DOES NOT CHECK compatibility of packed sequences!
    return torch.nn.utils.rnn.PackedSequence(fn(*[(arg.data if type(arg)==torch.nn.utils.rnn.PackedSequence else arg) for arg in args]), args[0].batch_sizes)

but that has been all I ever needed. Can be used with torch.x functions, user-defined or lambdas.

[dmitriy-serdyuk]

Checking that sequences are the same length doesn't seem very difficult.

In 4.0 PackedSequence received cuda(), cpu() and type conversion methods, in 5.0 to(). It would be logical to continue this direction of making it to behave like Tensor.

[shengpu-tang]

The documentation seems to suggest that users should never invoke the __init__ method of PackedSequence. I agree that the elementwise_apply functions are a way to make this work, but it would be nice if we make the interface of PackedSequence and Tensor consistent.

pytorch/torch/nn/utils/rnn.py

Line 18 in 0656ef4

Instances of this class should never be created manually. They are meant

[jluo41]

I have a question about using torch.nn.utils.rrn.pack_padded_sequence( input, lengths, batch_first). When using this function, the input sequences should be sorted by length in a decreasing order, as well as the length. What if I want to keep the original order of sequences in the order? Can this function support the input which is not sorted by length in a decreasing order?

[denkorzh]

also some aggregating operations would be quite useful, e.g. mean() with respect to varying lengths of sequences et cetera

[RylanSchaeffer]

What's the status of this? Coming soon, I hope :D

[jannessm]

I would be very interested in concatenation of packedsequences with other packedsequences or tensors!