Day 3 - PIL, pytorch-torchvision, logging-wandb

PIL(Python Imaging Library)

: 이미지 처리에 중점을 둔 라이브러리

- 픽셀 단위의 이미지 조작이나, 마스킹, 투명도 제어, 윤곽 보정 밒 검출 등의 다양한 이미지 조작을 할 수 있다.

pip install pillow

Image read and Visualize

from PIL import Image
import matplotlib.pyplot as plt

# image read
img=Image.open('../data/image.jpeg')

# image visualize
plt.imshow(img)
plt.show()
img

Image resize

w,h=img.size
img_resize=img.resize((w///2,h//2))
plt.imshow(img_resize)
plt.show()

img_resize=img.resize((w*2,h*2))
plt.imshow(img_resize)
plt.show()

Image crop

def two_image_show(im1,im2):
	plt.subplot(1,2,1)
    plt.imshow(im1)
    plt.subplot(1,2,2)
    plt.imshow(im2)
    plt.show()
    
## (lift, upper,right,lower)
img_crop=img.crop((w//4,h//4,w*3//4,h*3//4))

plt.imshow(img_crop)

two_image_show(img,img_crop)

Image rotate/flip

img_rotate=img.rotate(45) # degree
# img_rotate.show()
plt.imshow(img_rotate)

img_flip=img.transpose(Image.FLIP_LEFT_RIGHT)
two_image_show(img,img_flip)

img_flip_and_rotate=img.rotate(45).transpose(Image.FLIP_LEFT_RIGHT)
img_flip_and_rotate.save('save_img.png')

Scikit-image

pil과 마찬가지로 이미지 필터링이 가능하다.

PyTorch, TorchVision

gpu와 cpu를 통해 딥러닝을 가능하게 하는 tensor library.

 

추가 라이브러리 - Albumentation

https://albumentations.ai/

Albumentations: fast and flexible image augmentations

Tensor 생성하기

import torch
import numpy as np

data=[[1,2],3,4]]
# 2x2 tensor
x_data=torch.tensor(data)

# scalar tensor
x_data=torch.tensor(3.14149)

# empty tensor
x_data=torch.tensor([])

# numpy array에서 tensor 만들기
np_array=np.array(data)
x_np=torch.from_numpy(np_array)

# API를 이용해 tensor 만들기
shape=(2,3)
torch.rand(shape)
torch.ones(shape)
torch.zeros(shape)

# 다른 tensor를 이용해 tensor 만들기
x_ones=torch.ones_like(x_data) # x_data의 property만 갖음
x_ones=torch.rand_like(x_data,dtype=torch.flaot) # x_data의 datatype을 덮어쓴다.

 

Tensor attribute와 GPU

# Tensor attributes
tensor=torch.rand(3,4)
tensor.shape
tensor.dtype
tensor.device

# GPU에서 tensor 실행하기
tensor=tensor.to('cuda')
tensor.device

Tensor operations

Loading data

데이터셋을 읽고 multi-process mini-batch 데이터 로딩을 해결해준다.

# getting FashionMNIST datasets from torchvision
from torch.utils.data import Dataset
from torchvison import datasets
from torchvision.transforms import ToTensor
import matplotlib.pyplot as plt

training_data=datasets.FashionMNIST(
	root="data",
    train=True,
    download=True,
    transform=ToTensor())
    
test_data=datasets.FashionMNIST(
	root="data",
    train=False,
    download=True,
    transform=ToTensor())

Custom dataset 만들때 

•  __init__(self,[]): 데이터셋 루트 디렉토리/transform/이미지 path...
• __len__(self): 데이터셋의 length(크기) 반환
• __getitem__(self,idx): (idx)th data를 return. *Dataloader calls this method

Neural Network 만들기

__init__함수와 forward 함수 필

class NeuralNetwork(nn.Module):
	def __init__(self):
    	super(NeuralNetwork, self).__init__()
        self.flatten=nn.Flatten()
        self.linear_relu_stack=nn.Sequential(
        	nn.Linear(28*28,512),
            nn.ReLU(),
            nn.Linear(512,512),
            nn.ReLU(),
            nn.Linear(512,10),
		)
	def forward(self,x):
    	x=self.flatten(x)
        logits=selg.linear_relu_stack(x)
        return logits

Training

def train_loop(dataloader,model,loss_fn,optimizer):
	size=len(dataloader.dataset)
    for batch,(X,y) in enumerate(dataloader):
    	# Compute prediction and loss
        pred=model(X)
        loss=loss_fn(pred,y)
        
        # Backpropagation
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        if batch%100==0:
        	loss,current=loss.item(),batch*len(X)
            print(f"loss: {loss:>7f} [{current:>5d}/{size:>5d}]")

Transfer learning

pre-trained model을 사용하기

model_ft=models.resnet18(pretrained=True)
num_ftrs=model_ft.fc.in_features

# Here the size of each output sample is set to 2
# Alternatively, it can be generalized to nn.Linear(num_ftrs,len(class_names))
model_ft.fc=nn.Linear(num_ftrs,2)

model_ft=model_ft.to(device)

criterion=nn.CrossEntropyLoss()

# Observe that all parameters are being optimized
optimizer_ft=optim.SGD(model_ft.parameters(),lr=0.001,momentum=0.9)

# Decay LR a factor of 0.1 every 7 epochs
exp_lr_scheduler=lr_scheduler.StepLR(optimizer_ft,step_size=7,gamma=0.1)

Logging, Wandb, Kornia

Logging

어떤 소프트웨어가 실행될 때 발생하는 이벤트를 추적하는 수단.

• print()
• logging.info()/logging.warn()
• ...

import logging

logging.warning('watch out!') # will print out a message to the console
logging.info('I told you so') # will not print anything

logging.basicConfig(filename='example.log', encoding='utf-8', level=logging.DEBUG)
logging.debug('This message should  go to the log file')
logging.info('So should this')
logging.warning('And this, too')
logging.error('And non-ASCII stuff, too..')

Wandb

학습 모델 개발 시에 하이퍼파라미터나, 시스템메트릭, 결과를 로깅해주는 패키지

tensorboard/tensorboardX나 hydra 같은 패키지들이 있다.

pip install wandb

Kornia

AI를 위한 컴퓨터비전 알고리즘의 구현체를 모아놓은 라이브러리