From bf04b102d097c40bb621af5ad461a6a717de0c8b Mon Sep 17 00:00:00 2001 From: Panjaksli <82727531+Panjaksli@users.noreply.github.com> Date: Thu, 31 Aug 2023 08:08:54 +0200 Subject: [PATCH 1/3] Update README.md --- README.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/README.md b/README.md index b57368c..0696878 100644 --- a/README.md +++ b/README.md @@ -42,6 +42,8 @@ Quality-wise cubl and swish provide smoother upscaled image than relu. This repository comes with custom pre-trained model for high quality image upscaling, that achieves far better results than any simple upscaling algorithm (bicubic, bilinear). ### Comparison: bicubic vs model vs reference ![image](https://github.com/Panjaksli/BNN/assets/82727531/fb3a9592-5987-4eb9-bde0-dccecb1c459e) +### Lowres anime vs upscale +![image](https://github.com/Panjaksli/BNN/assets/82727531/388cda89-ed14-4f15-b1ae-efc009cd7a42) ### How does it work ? The model is trained on the error of reference image and low res image upscaled with bicubic interpolation:\ d(x) = f(x) - g(x),\ From bec7aa38e50c508ac61ba856d1a11b6e1ecb374d Mon Sep 17 00:00:00 2001 From: Panjaksli <82727531+Panjaksli@users.noreply.github.com> Date: Thu, 31 Aug 2023 08:11:27 +0200 Subject: [PATCH 2/3] Update README.md --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 0696878..07dfe28 100644 --- a/README.md +++ b/README.md @@ -1,7 +1,7 @@ # BNN Basic Neural Net - Neural network built on top of Eigen/Tensor library, offering support for various layers and parallel CPU learning. The library delivers exceptional performance in spatial convolutions, achieving up to 100x performance of Convolve method from Eigen/Tensor along with much better CPU clock scaling ## What is this library ? -This is an Eigen based sequential neural network library with goal of implementing various layer types, activation functions and optimizers whilst achieving superior perfomance in CPU training **and I really mean superior**. +This is an Eigen based sequential neural network library with goal of implementing various layer types, activation functions and optimizers whilst achieving superior perfomance in CPU training **and I really mean superior**. Almost everything is implemented from scratch ! ## Performance optimizations Computations use the bare minimum of temporaries and leverage column major layout of tensors as much as possible (especially the custom spatial convolution algorithm). The number of virtual calls is kept to the absolute minimum too.The library vectorizes quite well and can even leverage multiple CPU cores for learning and inference. ## Features From e82561670d3a5f775a34a64ee17d2a9cacc3b78b Mon Sep 17 00:00:00 2001 From: Panjaksli <82727531+Panjaksli@users.noreply.github.com> Date: Thu, 31 Aug 2023 09:28:15 +0200 Subject: [PATCH 3/3] Update README.md --- README.md | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/README.md b/README.md index 07dfe28..33a736e 100644 --- a/README.md +++ b/README.md @@ -39,11 +39,11 @@ Loss-wise relu is the best, cubl slightly outperforms swish. ![image](https://github.com/Panjaksli/BNN/assets/82727531/e066678c-629e-4c8d-99d4-abff40ee6de3) Quality-wise cubl and swish provide smoother upscaled image than relu. ## Custom pre-trained CNN for image upscaling -This repository comes with custom pre-trained model for high quality image upscaling, that achieves far better results than any simple upscaling algorithm (bicubic, bilinear). +This repository comes with custom pre-trained model for high quality image upscaling, that achieves far better results than any simple upscaling algorithm (bicubic, bilinear). And even though the network was trained only on real-life images, it performs on upscaling anime art too ! ### Comparison: bicubic vs model vs reference ![image](https://github.com/Panjaksli/BNN/assets/82727531/fb3a9592-5987-4eb9-bde0-dccecb1c459e) -### Lowres anime vs upscale -![image](https://github.com/Panjaksli/BNN/assets/82727531/388cda89-ed14-4f15-b1ae-efc009cd7a42) +### Upscaling anime art +![image](https://github.com/Panjaksli/BNN/assets/82727531/718568a6-111a-4436-870b-c206874185eb) ### How does it work ? The model is trained on the error of reference image and low res image upscaled with bicubic interpolation:\ d(x) = f(x) - g(x),\