[python] einops 관련 함수

numpy/pytorch vs. einops¶

• transpose : transpose vs. rearrange
• reshape : reshape/view vs. rearrange
• upsize : repeat/upsample vs. repeat
• downsize : interpolate vs. reduce
• split : split vs. rearrange
• permute : permute vs. rearrange

rearrange함수로 앵간한 건 다 다루는 듯

In [8]:

import torch
import torch.nn.functional as F
import matplotlib.pyplot as plt
import numpy as np
from torch import nn
from torch import Tensor
from PIL import Image
from torchvision.transforms import Compose, Resize, ToTensor
from einops import rearrange, reduce, repeat
from einops.layers.torch import Rearrange, Reduce
from torchsummary import summary

In [14]:

x = Image.open('./data/ac2.jpg')
x

Out[14]:

In [15]:

x = x.resize((224,224))
tf_toTensor = ToTensor() 
x = tf_toTensor(x)
x = rearrange(x,'c h w -> h w c')
x.shape

Out[15]:

torch.Size([224, 224, 3])

flatten¶

In [18]:

rearrange(x, 'h w c -> (h w c)').shape
"""
input tensor : x
input shape : height, width, channel로 명시함
output shape : fuse height, width, channel by multiplying them together 
()연산이 product operation
"""

Out[18]:

torch.Size([150528])

* flatten in numpy¶

In [20]:

np.reshape(x, (-1,)).shape

Out[20]:

torch.Size([150528])

Images to patches ***¶

split an image into 2x2 patches

In [23]:

rearrange(x, "(p1 h) (p2 w) c -> (p1 p2) h w c", p1=2, p2=2).shape
"""
input shape: (p1 h) : split height into 2, (p2 w) : split width into 2
output shape : (p1 p2) : an array of 4 image patches
"""

Out[23]:

torch.Size([4, 112, 112, 3])

Mixing 2 images¶

each image is a mix of input images

In [168]:

def preprocess(x):
    x = x.resize((224,224))
    tf_toTensor = ToTensor() 
    x = tf_toTensor(x)
    x = rearrange(x,'c h w -> h w c')
    return x

In [169]:

x = Image.open('./data/ac2.jpg')
y = Image.open('./data/ac4.jpg')
x = (preprocess(x)).numpy()
y= (preprocess(y)).numpy()
imgs = np.array([x,y])

In [170]:

imgs = rearrange(imgs, 'b (k h) w c -> k b h w c', k=2)

In [171]:

imgs.shape

Out[171]:

(2, 2, 112, 224, 3)

In [138]:

imgs[::2].shape

Out[138]:

(1, 2, 112, 224, 3)

In [172]:

imgs=np.concatenate([imgs[::2], imgs[1::2,::-1]],axis=0)
#imgs[::2] => upper image
#imgs[1::2,::-1] ==> lower image reverse
imgs = rearrange(imgs, "i j h w c -> j (i h) w c")

In [173]:

r1=(imgs[0]*256).astype(np.uint8)
Image.fromarray(r1)

Out[173]:

In [174]:

r2=(imgs[1]*256).astype(np.uint8)
Image.fromarray(r2)

Out[174]:

RGB to grayscale¶

In [190]:

x = Image.open('./data/ac2.jpg')
x = x.resize((224,224))
tf_toTensor = ToTensor() 
x = tf_toTensor(x)
x = rearrange(x,'c h w -> h w c')

In [191]:

img = reduce(x, "h w c -> h w", 'mean')

upsampling¶

In [192]:

img = repeat(x, "h w c -> (h 2) (w 2) c")

downsampling¶

In [194]:

img = reduce(img, "(h 2) (w 2) c-> h w c", 'mean')

In [ ]:

In [ ]: