No project description provided
Project description
ENN Ragged Buffer
This Python package implements an efficient RaggedBuffer
datatype that is similar to
a 3D numpy array, but which allows for variable sequence length in the second
dimension. It was created primarily for use in ENN-PPO
and currently only supports a small selection of the numpy array methods.
User Guide
Install the package with pip install ragged-buffer
.
The package currently supports three RaggedBuffer
variants, RaggedBufferF32
, RaggedBufferI64
, and RaggedBufferBool
.
Creating a RaggedBuffer
There are three ways to create a RaggedBuffer
:
RaggedBufferF32(features: int)
creates an emptyRaggedBuffer
with the specified number of features.RaggedBufferF32.from_flattened(flattened: np.ndarray, lenghts: np.ndarray)
creates aRaggedBuffer
from a flattened 2D numpy array and a 1D numpy array of lengths.RaggedBufferF32.from_array
creates aRaggedBuffer
(with equal sequence lenghts) from a 3D numpy array.
Creating an empty buffer and pushing each row:
import numpy as np
from ragged_buffer import RaggedBufferF32
# Create an empty RaggedBuffer with a feature size of 3
buffer = RaggedBufferF32(3)
# Push sequences with 3, 5, 0, and 1 elements
buffer.push(np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]], dtype=np.float32))
buffer.push(np.array([[10, 11, 12], [13, 14, 15], [16, 17, 18], [19, 20, 21], [22, 23, 24]], dtype=np.float32))
buffer.push(np.array([], dtype=np.float32)) # Alternative: `buffer.push_empty()`
buffer.push(np.array([[25, 25, 27]], dtype=np.float32))
Creating a RaggedBuffer from a flat 2D numpy array which combines the first and second dimension, and an array of sequence lengths:
import numpy as np
from ragged_buffer import RaggedBufferF32
buffer = RaggedBufferF32.from_flattened(
np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9], [10, 11, 12], [13, 14, 15], [16, 17, 18], [19, 20, 21], [22, 23, 24], [25, 25, 27]], dtype=np.float32),
np.array([3, 5, 0, 1], dtype=np.int64))
)
Creating a RaggedBuffer from a 3D numpy array (all sequences have the same length):
import numpy as np
from ragged_buffer import RaggedBufferF32
buffer = RaggedBufferF32.from_array(np.zeros((4, 5, 3), dtype=np.float32))
Get size
The size0
, size1
, and size2
methods return the number of sequences, the number of elements in a sequence, and the number of features respectively.
import numpy as np
from ragged_buffer import RaggedBufferF32
buffer = RaggedBufferF32.from_flattened(
np.zeros((9, 64), dtype=np.float32),
np.array([3, 5, 0, 1], dtype=np.int64))
)
# Get size of the first/batch dimension.
assert buffer.size0() == 10
# Get size of individual sequences.
assert buffer.size1(1) == 5
assert buffer.size1(2) == 0
# Get size of the last/feature dimension.
assert buffer.size2() == 64
Convert to numpy array
as_aray
converts a RaggedBuffer
to a flat 2D numpy array that combines the first and second dimension.
import numpy as np
from ragged_buffer import RaggedBufferI64
buffer = RaggedBufferI64(1)
buffer.push(np.array([[1], [1], [1]], dtype=np.int64))
buffer.push(np.array([[2], [2]], dtype=np.int64))
assert np.all(buffer.as_array(), np.array([[1], [1], [1], [2], [2]], dtype=np.int64))
Indexing
You can index a RaggedBuffer
with a single integer (returning a RaggedBuffer
with a single sequence), or with a numpy array of integers selecting/permuting multiple sequences.
import numpy as np
from ragged_buffer import RaggedBufferF32
# Create a new `RaggedBufferF32`
buffer = RaggedBufferF32.from_flattened(
np.arange(0, 40, dtype=np.float32).reshape(10, 4),
np.array([3, 5, 0, 1], dtype=np.int64)
)
# Retrieve the first sequence.
assert np.all(
buffer[0].as_array() ==
np.array([[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11]], dtype=np.float32)
)
# Get a RaggedBatch with 2 randomly selected sequences.
buffer[np.random.permutation(4)[:2]]
Addition
You can add two RaggedBuffer
s with the +
operator if they have the same number of sequences, sequence lengths, and features. You can also add a RaggedBuffer
where all sequences have a length of 1 to a RaggedBuffer
with variable length sequences, broadcasting along each sequence.
import numpy as np
from ragged_buffer import RaggedBufferF32
# Create ragged buffer with dimensions (3, [1, 3, 2], 1)
rb3 = RaggedBufferI64(1)
rb3.push(np.array([[0]], dtype=np.int64))
rb3.push(np.array([[0], [1], [2]], dtype=np.int64))
rb3.push(np.array([[0], [5]], dtype=np.int64))
# Create ragged buffer with dimensions (3, [1, 1, 1], 1)
rb4 = RaggedBufferI64.from_array(np.array([0, 3, 10], dtype=np.int64).reshape(3, 1, 1))
# Add rb3 and rb4, broadcasting along the sequence dimension.
rb5 = rb3 + rb4
assert np.all(
rb5.as_array() == np.array([[0], [3], [4], [5], [10], [15]], dtype=np.int64)
)
Concatenation
The extend
method can be used to mutate a RaggedBuffer
by appending another RaggedBuffer
to it.
import numpy as np
from ragged_buffer import RaggedBufferF32
rb1 = RaggedBufferF32.from_array(np.zeros((4, 5, 3), dtype=np.float32))
rb2 = RaggedBufferF32.from_array(np.zeros((2, 5, 3), dtype=np.float32))
rb1.extend(r2)
assert rb1.size0() == 6
Clear
The clear
method removes all elements from a RaggedBuffer
without deallocating the underlying memory.
import numpy as np
from ragged_buffer import RaggedBufferF32
rb = RaggedBufferF32.from_array(np.zeros((4, 5, 3), dtype=np.float32))
rb.clear()
assert rb.size0() == 0
License
ENN Ragged Buffer dual-licensed under Apache-2.0 and MIT.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distributions
Hashes for ragged_buffer-0.3.0-cp39-cp39-manylinux_2_24_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 7492a333c8223fff078ee2d65725028b70334199b30569d5e35bb8b06c8738b0 |
|
MD5 | af3d2d82bc689bcd34c9e4a59fec56c4 |
|
BLAKE2b-256 | 815635903aab99f46b20c81b6666e06b180850f08fbe9d3441b1528013a515af |
Hashes for ragged_buffer-0.3.0-cp39-cp39-macosx_10_7_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 57f83c74395fa5e6504eee1f1b4e5e75456d3037fbfe5b9ac72c51937ab58960 |
|
MD5 | 51592711b7d0f8d1866d7303e3ceb72b |
|
BLAKE2b-256 | e8ab8d35587bc23467ddb37ed9250047726b60af3caec9c1293949e36500a13c |
Hashes for ragged_buffer-0.3.0-cp38-cp38-manylinux_2_24_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 5894f58115af6ba268b32e9369764004df1bde82c57317b3ac66359e6cabe7b3 |
|
MD5 | a2114b0449e280616a20b25c362b9ff8 |
|
BLAKE2b-256 | 1fb586265942aa1f62d44c30f1d017905a2a980d9e390cf6b798d0e49e195c3d |
Hashes for ragged_buffer-0.3.0-cp38-cp38-macosx_10_7_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | fdf1719f28ca33b0103c17656785dffd9e27e46d35ace90f86a9be63ccee368c |
|
MD5 | 1f67be09063a0f494ef2256a5857981a |
|
BLAKE2b-256 | 6da9db70c24ce9b5b29f068d760c2010876425623652700415ecbb7bfc9fd7d2 |
Hashes for ragged_buffer-0.3.0-cp37-cp37m-manylinux_2_24_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | b1fba5fb1192f92945e9b44d9cab85385a18d36b6a675658801aef9b0aa3995e |
|
MD5 | 27b6fc6bf6c6a0efdfcf695284afdba6 |
|
BLAKE2b-256 | 85269e3ecb9a6506050c22e3f77cde46f2d8da41fd3ceae7decd20d39aed4fe9 |
Hashes for ragged_buffer-0.3.0-cp37-cp37m-macosx_10_7_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | dfaea1c984ce1b0ec2ec322d7382fad649c45c675fb60b3770437029a4386cec |
|
MD5 | 172ef16f9e00b89579a98f5ae6c18df8 |
|
BLAKE2b-256 | 4f187ab28718f8e5201018530375793deadfe690edff8527b1a03f2b86f6b011 |