CNN Convolution and Pooling

Posted on June 2, 2019
Tags: machinelearning

1 Imports

from pathlib import Path
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import torch
print("Using torch", torch.__version__)
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
from torchvision import transforms
from PIL import Image
trans_GreyScale = transforms.Compose([
                    transforms.Grayscale()
                    ])
imgRaw = Image.open("sample.png")
img = trans_GreyScale(imgRaw)
from torchvision import datasets
dataset = datasets.MNIST(root=Path.cwd(), download=True)
img, label = dataset[10]
plt.imshow(img,cmap='gray')

2 Kernel(Filter) Convolution

2.1 Kernel(Filter) matrix

kernel = torch.tensor([[[[ -1, -1, -1],
                         [  2,  2,  2],
                         [ -1, -1, -1]]]],dtype=torch.float)

2.2 Image to Tensor

trans_Img2Tensor = transforms.Compose([
                    transforms.ToTensor()
                    ])
demo = trans_Img2Tensor(img)

2.3 Convolve

convImg_tensor = torch.nn.functional.conv2d(demo,kernel,stride=1,padding=0)

2.4 Tensor to Image

trans_Tensor2Img = transforms.Compose([
                    transforms.ToPILImage()
                    ])
newImg = trans_Tensor2Img(convImg_tensor)
plt.imshow(newImg,cmap='gray')

3 Pooling

poolImg_tensor = torch.nn.functional.max_pool2d(convImg_tensor,kernel_size=2)
newpoolImg = trans_Tensor2Img(poolImg_tensor)
plt.imshow(newpoolImg,cmap='gray')

4 MultiChannel Convolution

raw image with 3 channels AKA RGB color 
kernel3 = torch.tensor([[[[ -1, -1, -1],
                          [  2,  2,  2],
                          [ -1, -1, -1]],
                         [[ -1, -1, -1],
                          [  2,  2,  2],
                          [ -1, -1, -1]],
                         [[ -1, -1, -1],
                          [  2,  2,  2],
                          [ -1, -1, -1]]]]
                          ,dtype=torch.float)
#ideally if you built the tensor correctly above, the below reshaping should do nothing
kernel3c = torch.reshape(kernel3,(1,3,3,3)) 
# Shape = (Num of Samples,Channels,Height,Width) = (1,3,3,3); 
demo = trans_Img2Tensor(imgRaw)
convImg_tensor = torch.nn.functional.conv2d(demo,kernel3c)
Img3c = trans_Tensor2Img(convImg_tensor)

const kernFilter = tf.tensor4d([-1,-1,-1,2,2,2,-1,-1,-1, -1,-1,-1,2,2,2,-1,-1,-1, -1,-1,-1,2,2,2,-1,-1,-1],[3,3,3,1],"float32") 
#gets converted into 
const kernFilter =      tf.tensor4d([[[[-1],[-1],[-1]],
                                      [[2],[2],[2]],
                                      [[-1],[-1],[-1]]],
                                     [[[-1],[-1 ],[-1]],
                                      [[2],[2 ],[2]],
                                      [[-1],[-1 ],[-1]]],
                                     [[[-1],[-1],[-1]],
                                      [[2],[2],[2]],
                                      [[-1],[-1],[-1]]]])