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 two RaggedBuffer
variants, RaggedBufferF32
(storing float32 values) and RaggedBufferI64
(storing int64 values).
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))
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-0.2.9-cp310-none-win_amd64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 3989459ee9ceb12f7d9788df41bd38eb0b189506d4218d259760cfe64efff60d |
|
MD5 | 1da19bdae3014eb011d2cf20960b38c5 |
|
BLAKE2b-256 | 6fb69990213e55e1cb3ef637be2d86a65fd33418121e60bc624cd240598e2747 |
Hashes for ragged_buffer-0.2.9-cp310-cp310-macosx_10_7_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | feed256754219caca467afbc02b016fc8eff8153da215465d686f2faa48b963f |
|
MD5 | 5fa28cab43de65c8c78c25f226eb7b49 |
|
BLAKE2b-256 | fbce7426410e098e0105df4b5338536c70991efc0b6a90f82561040a027b8f95 |
Hashes for ragged_buffer-0.2.9-cp39-none-win_amd64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 62efbd049154aa61395fd97a65cb4740cb340d7bd0d9278861422617dcebb258 |
|
MD5 | 0ada8c432ae1f264338592133dfe4f11 |
|
BLAKE2b-256 | dee26ddb80c2523cddf6805c6a16125f37ac4cfd05897e1152106117340ef20e |
Hashes for ragged_buffer-0.2.9-cp39-cp39-macosx_10_7_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | ebae8931fab1e3fe8e9d35fd0a673eb1c76e26f55c061c00978360ef226c7c2a |
|
MD5 | 8e25a64802240e8e0cb904b0343777f1 |
|
BLAKE2b-256 | c379281b9328e8bf55a7d2d883fa6c56984f9343128f03cc55bae0f7cdbc9d03 |
Hashes for ragged_buffer-0.2.9-cp38-none-win_amd64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | e211bc02a27d398bd1b5f43c001a116e8abed75abbfbf73859a9830c40eaac3d |
|
MD5 | b95746213b6b899fd4bcd3c98eaa4840 |
|
BLAKE2b-256 | 00198d03927f8b93a774aafeff08965df4f05536464d67c5e3ff0b70029308b5 |
Hashes for ragged_buffer-0.2.9-cp38-cp38-manylinux_2_24_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | de693ffc8f7227830662f830eb06930ac6a12721d6829eb0fd5eb3b8e8cc29de |
|
MD5 | f629f7d28f921adfa001d2ec31d7e5f8 |
|
BLAKE2b-256 | 8605c2b5a5047bfe53bafc6514719f9ec53d98e2ee6f592497a7be27b5c4dd0e |
Hashes for ragged_buffer-0.2.9-cp38-cp38-macosx_10_7_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 57b29c301aed5c53d93d0015c8d770307bd2607abb94fdab24638388cf1752b1 |
|
MD5 | 3c20d568cd234832336622a35b30ff5c |
|
BLAKE2b-256 | a946e86a398f8ac3970e60cd0ebb45ac5d824f714fcea5cd2b0be6d23322eb99 |
Hashes for ragged_buffer-0.2.9-cp37-none-win_amd64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | b314a6b9c18599f27523ebf89d5fef1424da84bb091b3a7d2130131023ce95e3 |
|
MD5 | 3ccd1eeb0d03adde4268d4104b1edcee |
|
BLAKE2b-256 | d3e1f6d80454a845c81ce83e6bf221035748370328fa6dfcafbc5c3e9f583b63 |
Hashes for ragged_buffer-0.2.9-cp37-cp37m-macosx_10_7_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 0990d7f9112f6b5300a28ad7f6a8aa7904999282033cf139f5fbb54c83b8d402 |
|
MD5 | 8d0fbd6d43a77f8e89229968485227a2 |
|
BLAKE2b-256 | a1e1f758ce308439557404310aab0c4b417d44e6bd6e6182fdef96ae94c41df0 |
Hashes for ragged_buffer-0.2.9-cp36-none-win_amd64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | d637252dd345df2962656097c490fe3d6ff5ef37d119bb82fe2335dea582eaa0 |
|
MD5 | 4082725d6d8853105ba51b3e8ab30876 |
|
BLAKE2b-256 | 35e04acc817c0d8799bd601576e19000b67316517c5a09ca0f5b6cfe50106090 |