글쓴이 아카이브:jiyeonlab

건축 도면을 데이터 셋으로 활용하기 전 건축 도면과 비슷한 데이터 셋을 실험적으로 돌려보았다. Dataset: 100개(모양, 위치, 선굵기, 선개수), PPT를 사용하여 만듦 Training 사진 크기가 1024×1024 이면 Google Collaboratory RAM 부족으로 오류 발생(참고로 Colab RAM은 25.51GB) 따라서 사진 크기를 512×512 로 변경 후 training 시작 batch=10, epoch=100 batch=2, epoch = 10 두번의 training 모두 d_loss 값이“GAN tutorial #5” 계속 읽기

Question of today: Why input has to be square(same x and y) in GAN? I found good reference in Reason for square images. I tried to use rectengular input image to my GAN code. I used 432*624 image and there was no problem at all.

I draw UML of GAN algorithm. UML And I success at training 512 x 512 image. code REF http://taewan.kim/post/transpose_reshape/

Again, I try to understand the code https://colab.research.google.com/drive/1E5ZQLKux4RXqYK2-kmcxfAolEgSzxFh4 What is cross entropy loss? What is transform? transforms.ToTensor() 데이터 타입을 Tensor 형태로 변경 transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]) 이미지의 경우 픽셀 값 하나는 0 ~ 255 값을 갖는다. 하지만 ToTensor()로 타입 변경시 0 ~ 1 사이의 값으로 바뀜. transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))를 이용하여 -1“GAN tutorial in Pytorch #2” 계속 읽기

What is GAN? GAN is composed of two networks. Generator is a network that generates new data instance and Discriminator is a network that determine if the data instance is a real or fake. Generator aims to generate data instance that is hard to discriminate and Discriminator aims to have good performance at discriminating real“GAN tutorial in Pytorch” 계속 읽기

At first, I tried to test my model. Since I didn’t separate dataset into train data and test data, I just downloaded sample data while training. I added new code that saved result images in result folder. However, I tried to change images into grayscale images. But I failed….

Yesterday, the problem was that the loss value did not change at all while the model were training. So, I compared floor dataset with original pokemon dataset.I found out that this model extract the alpha value from each image.However, floor plans were composed of only black and white colors.So I made white color transparent so“How can I generate floor plan using GAN?” 계속 읽기

At first I prepared 20 floor plan images from Google search.I tried to find clear floor plan images which have less noises such as text, and color. Then, I fixed the size of images into 256px X 256px using Window 3D paint software. Then, I uploaded images into file folder named pokemon.pokemon folder is a“How can I generate floor plan using GAN?” 계속 읽기

I saved some floor plans from Google search. However, errors occured when non-trimed data went into pokemon GAN code. So, I trimed all the dataset into fixed size 256×256 which was same size with original pokemon images. The sizes of epoch was bit small but the there were no errors while training!

My problem was the tutorial was successful but I wanted to modify tutorial algorithm so that my dataset can be trained. Tensorflow algorithm was well stable and structured algorithm. But MNIST dataset was too many for me to modify and test the algorithm. Therefore I found simple and intuitive but modify-needed algorithm. Original LinkMy source“How can I modify algorithm with tutorial?” 계속 읽기