Deep-Learning

I used Captum to interpret the output of a MobileNetV2, which visualized the main regions in the input image that drove the model to generate its output.

I converted a PyTorch model to Core ML and ran it on an iPhone.

I combined my previous posts on image captioning and visual question answering and extended them to a wider topic - connecting computer vision and natural language.

Recently I was working with PyTorch multi-GPU training and I came across a nightmare GPU memory problem. After some expensive trial and error, I finally found a solution for it.

I will start from the problem of semantic segmentation, introduce how to use CNNs to solve it, and talk about fully convolutional networks, a widely used framework for semantic segmentation, in great details. Moreover, I will analyze the MXNet implementation of FCNs.

I read some papers on VQA and summarized its state-of-the-art, bottlenecks and possible solutions.

I played with image captioning using neuraltalk2 written by Andrej Karpathy.