Just begins.
Recently I've done a project which needs to model using Graph data(not image graph, but network graph). A big problem I encountered is how to slice subgraph into model training. An intuitive idea will be like using sparse matrix to slice out the corresponding neighbors for any given node. But due to the constraint of tf.SparseTensor, which do not support frequent and also efficient slice operation for now, I choose to use csr_matrix in scipy to take care of this part.
While I interned at my current company - Kuaishou, I was in part of the project where I need to set up a LSTM model, and I have to transfer my skill sets to Tensorflow from Pytorch in a short period of time , which was really painful experience for me. So I want to share this blog with you to ease your pain.
This passage will lead you through the whole process of the implementation of image classification tasks on CIFAR-10 and also explain some confusing parts for beginners in detail. Besides, I also put down what I've learned in the middle and hope you will at least grasp how you should build up the whole pipeline of conv net at last.
During the learning of RNN, I undergone a very tough time to deduce the BPTT for vanilla RNN. There are some pieces of work online that try to explain this process(click here or here). However, I found the derivation relevant to hidden state unsatisfying. Specificly, They are mixing up the partial derivative and total derivative, which is a total disaster for beginners to understand it. So I want to write this blog to put down some tough points in the process. And meanwhile, I will also skip the parts that's been explained thoroughly in the posts above. Hopefully, this blog will help you save a bunch of time to wrap you head around this fantastic algorithm.