164. Combining Losses
Here is one way to combine losses using PyTorch. Let’s say we want to train a binary semantic segmentation model using Binary Cross Entropy Loss. When you look at your data, you’ve noticed that the data is quite imbalanced. So…
163. Why Normalize Inputs?
Why Do we Normalize Inputs? When the input is not normalized, the shape of the cost function can become distorted like the diagram on the left. This leads to instability when optimizing the model. The training speed decreases depending on…
162. Residual Blocks
Why are residual blocks called “residual” blocks? The reason why I was confused was that the equation in the diagram explaining the residual blocks on the research paper was f(x) + x. So I thought, “Where is the residual..?” When…
161. ESRGAN
Abstract Even though SR-GAN was able to make a huge improvement, there was still a gap between the generated image and the ground truth image. The proposed ESR-GAN further enhances the performance. Three Key Modification Components Network Remove all batch…
160. DeepPose
DeepPose DeepPose is a research done by Google for human pose estimation. Pose Vector First, the paper encodes all “k” body joints into a pose vector. To avoid using absolute coordinates for the body joints like right now, the paper…
159. M-RNN
M-RNN Multi-Modal Recurrent Neural Network is a research done by The University of California and the Baidu Research Team which generates captions for images. In this research,Deep Recurrent Neural Network is used for sentences, and Deep Convolutional Neural Network is…
158: SR-GAN
SR-GAN Today I learned about SR(Super Resolution)-GAN, so I’d like to share it here. In previous research, super-resolution tasks(Enhancing resolution) struggled when recovering finer text details at large upscale factors. SR-GAN is the first to be able to infer images…
157. CycleGAN
CycleGAN Today I’ve learned about CycleGAN, so I’d like it here. Before this research, Image-to-Image translation tasks(Learning how to map an input image to a different style image) required “PAIR” data sets for training. Unfortunately, in most cases, you don’t…
156. Highway Networks
Highway Networks Training models with DEEP networks becomes difficult, even when using variance-preserving initialization. By adding an information highway (Learning how to route information through the network), it makes it easier to train models even when it is really DEEP.…
155. S3D (Separable 3D CNN)
I’ve learned about S3D(Separable 3D CNN) today so I like to share it here. S3D helps solve three challenges for video analysis. How to understand spatial information. (Recognizing the appearance of an object) How to understand temporal information. (Such as…