Bruce Kim

ML Engineer

[ML] Numpy & PyTorch Summary

4 minute read

numpy-version-1.18.1 torch-version-1.4.0

Numpy & PyTorch Summary

import numpy as np
import torch


arr = np.random.randn(1, 2, 3, 4)
tensor = torch.randn(1, 2, 3, 4)



Shape 확인 및 변환

Numpy

  • shape, ndim, size, reshape()

Note: numpy 의 reshape 의 반환값은 view() 가 적용된다.
즉 copy() 가 되는 형태가 아니기 때문에 base array 를 참조하여 메모리를 공유한다.

print(arr.shape)
print(arr.ndim)
print(arr.size)

>>> (1, 2, 3, 4)
>>> 4
>>> 24

new_arr = arr.reshape(-1, 3)
print(new_arr.shape)

>>> (8, 3)


PyTorch

  • size(), dim(), numel(), view(), ndim, shape, reshape()
print(tensor.size())
print(tensor.size(0))
print(tensor.size(1))
print(tensor.size(2))
print(tensor.size(3))
print(tensor.dim())
print(tensor.numel())

>>> torch.Size([1, 2, 3, 4])
>>> 1
>>> 2
>>> 3
>>> 4
>>> 4
>>> 24

new_tensor = tensor.view(-1, 3)
print(new_tensor.size())

>>> torch.Size([8, 3])



차원 변환

Numpy

# 차원 전치
new_arr = arr.transpose()  # == arr.T
print(new_arr.shape)

>>> (4, 3, 2, 1)

# 차원 순서 변환
new_arr = arr.transpose(2, 0, 1, 3)
print(new_arr.shape)

>>> (3, 1, 2, 4)

# 차원 수가 1인 차원 제거
new_arr = arr.squeeze()
print(new_arr.shape)

>>> (2, 3, 4)

# 차원 추가
new_arr = np.expand_dims(arr, axis=-1)
print(new_arr.shape)

>>> (1, 2, 3, 4, 1)

# 차원 추가 2
new_arr = arr[:, :, :, :, np.newaxis]
print(new_arr.shape)

>>> (1, 2, 3, 4, 1)


PyTorch

# 차원 순서 변환
new_tensor = tensor.permute(2, 0, 1, 3)
print(new_tensor.size())

>>> torch.Size([3, 1, 2, 4])

# 차원 순서 변환 (선택 2개)
new_tensor = tensor.transpose(dim0=1, dim1=2)  # numpy 와 다르게 2개의 차원만 변경가능
print(new_tensor.size())

>>> torch.Size([1, 3, 2, 4])

# 차원 수가 1인 차원 제거
new_tensor = tensor.squeeze()
print(new_tensor.size())

>>> torch.Size([2, 3, 4])

# 차원 추가
new_tensor = tensor.unsqueeze(dim=-1)
print(new_tensor.size())

>>> torch.Size([1, 2, 3, 4, 1])

# tensor 가 2차원일 경우 t() 를 통해 전치가능
tensor = torch.randn(3, 5)
print(tensor.size())

new_tensor = tensor.t()
print(new_tensor.size())

>>> torch.Size([3, 5])
>>> torch.Size([5, 3])



데이터 타입 확인 및 변환

Numpy

  • astype(), dtype
new_arr = arr.astype(np.int32)
print(new_arr.dtype)

>>> int32


PyTorch

  • type(), dtype, short(), int(), long(), half(), float(), double(), bool()
new_tensor = tensor.type(dtype=torch.int32)
print(new_tensor.type())
print(new_tensor.dtype)

>>> torch.IntTensor
>>> torch.int32

new_tensor1 = tensor.short()
new_tensor2 = tensor.int()
new_tensor3 = tensor.long()
new_tensor4 = tensor.half()
new_tensor5 = tensor.float()
new_tensor6 = tensor.double()
new_tensor7 = tensor.bool()
print(new_tensor1.type())
print(new_tensor1.dtype)
print(new_tensor2.type())
print(new_tensor2.dtype)
print(new_tensor3.type())
print(new_tensor3.dtype)
print(new_tensor4.type())
print(new_tensor4.dtype)
print(new_tensor5.type())
print(new_tensor5.dtype)
print(new_tensor6.type())
print(new_tensor6.dtype)
print(new_tensor7.type())
print(new_tensor7.dtype)

>>> torch.ShortTensor
>>> torch.int16
>>> torch.IntTensor
>>> torch.int32
>>> torch.LongTensor
>>> torch.int64
>>> torch.HalfTensor
>>> torch.float16
>>> torch.FloatTensor
>>> torch.float32
>>> torch.DoubleTensor
>>> torch.float64
>>> torch.BoolTensor
>>> torch.bool



PyTorch 유사 메서드 비교

  • torch.Tensor == torch.FloatTensor


  • torch.tensor vs torch.as_tensor
    as_tensor: copy 를 피해 참조 (data type 이 호환되고, device 가 동일한 경우).
    즉 numpy.ndarray 인 경우 copy 한다.
    이런 경우를 방지하기 위해 .from_numpy() 메서드를 사용한다.
    numpy.ndarray 의 dtype 을 그대로 가져오며 copy 를 피할 수 있다.


  • torch.chunk vs torch.split
    • torch.chunk: view of the input tensor
    • torch.split: view of the original tensor


  • torch.nn.relu vs torch.nn.functional.relu --> Just code style
    • torch.nn.relu: modules for the activations
    • torch.nn.functional.relu: functional calls


  • torch.nn.ConvTranspose2d vs torch.nn.Upsample --> Have Learing Parameters or not
    • torch.nn.ConvTranspose2d: Have Learning Parameters
    • torch.nn.Upsample: No Learning Parameters


  • view() vs reshape() (Documentation Link)
    • tensor.contiguous().view(${some shape}) == tensor.reshape
    • reshape(): tensor 가 contiguous 하지 못하면 copy 를 해서라도 동작



Numpy 유사 메서드 비교

  • np.array vs np.asarray
    • np.array: copy array
    • np.asarray: Reference data if data is compatible with ndarray
  • arr.flatten() vs arr.ravel()
    • arr.flatten(): Return copy of original array
    • arr.ravel(): Return only reference/view of original array



clip (clamp)

Numpy

new_arr = np.clip(arr, 0, 10)
# --> new_arr = np.clip(a=arr, a_min=0, a_max=10, out=None)

PyTorch

new_tensor = torch.clamp(tensor, 0, 10)
# --> new_tensor = torch.clamp(input=tensor, min=0, max=10, out=None)



where (get indices on specific conditions)

np.where, torch.where 사용법 동일

조건에 해당하는 값의 Index 를 반환한다.

  • ndim 만큼의 길이를 가지는 tuple 반환

  • tuple 의 첫번째 요소인 array 는 첫번째 차원의 Index
  • tuple 의 두번째 요소인 array 는 두번째 차원의 Index
  • tuple 의 세번째 요소인 array 는 세번째 차원의 Index
  • . . .

다차원의 경우, 해당값의 정확한 좌표를 구하기 위해서는
zip() 함수를 활용할 수 있다.


Example 1d array

import numpy as np


arr_1d = np.random.rand(10)
print(arr_1d)

# Out
[0.49931615 0.72081879 0.33509536 0.13864185 0.48212933 0.34469938
 0.29396133 0.19329217 0.71502539 0.6183096 ]

result = np.where(arr_1d < 0.5)
print(type(result))
print(len(result))
print(result[0])  # 조건에 해당하는 값의 index

>>> <class 'tuple'>
>>> 1
>>> [0 2 3 4 5 6 7]


Example 3d array

import numpy as np

arr_3d = np.random.rand(1, 2, 3)
print(arr_3d)

# Out
[[[0.65868097 0.70970589 0.9627755 ]
  [0.16858234 0.08135782 0.29145379]]]

result = np.where(arr_3d < 0.5)
print(type(result))
print(len(result))

>>> <class 'tuple'>
>>> 3

print_form = 'result[0]: {} | type(result[0]): {} | result[0].dtype: {}'
print(print_form.format(result[0], type(result[0]), result[0].dtype))
print(print_form.format(result[1], type(result[1]), result[1].dtype))
print(print_form.format(result[2], type(result[2]), result[2].dtype))

# Out
result[0]: [0 0 0] | type(result[0]): <class 'numpy.ndarray'> | result[0].dtype: int64
result[0]: [1 1 1] | type(result[0]): <class 'numpy.ndarray'> | result[0].dtype: int64
result[0]: [0 1 2] | type(result[0]): <class 'numpy.ndarray'> | result[0].dtype: int64


  • 조건을 만족하는 값의 좌표 구하기
pts = list(zip(*map(lambda arr : list(arr), result)))
print(pts)

>>> [(0, 1, 0), (0, 1, 1), (0, 1, 2)]



Pytorch Method

index_select

  • torch.index_select(input, dim, index, out=None)
x = torch.rand(5, 3)
print(x)

# Out
tensor([[0.4063, 0.2598, 0.2011],
        [0.2355, 0.2475, 0.2283],
        [0.1244, 0.3219, 0.2943],
        [0.0240, 0.4658, 0.9061],
        [0.6218, 0.1086, 0.2883]])


Example 1

a = torch.index_select(x, 0, torch.LongTensor([1, 3]))
print(a)

# Out
tensor([[0.2355, 0.2475, 0.2283],
        [0.0240, 0.4658, 0.9061]])


Example 2

b = torch.index_select(x, 1, torch.LongTensor([0, 2]))
print(b)

# Out
tensor([[0.4063, 0.2011],
        [0.2355, 0.2283],
        [0.1244, 0.2943],
        [0.0240, 0.9061],
        [0.6218, 0.2883]])

Leave a comment

Related posts