In this example, you will learn the basic flow of building model and training it. It supports two data type, namely dense data and libsvm data. On both data type, you will learn how to load training data using numpy and parse the data line by line. On dense data, line will be parsed into an label and a numpy array which is feeded to the model using Tensor in tensorflow. While on libsvm data, line will be parsed into an label and two numpy arrays where one holds the sparse id list and another holds the value list. The libsvm data will then be feeded to the model using Sparse Tensor in tensorflow. You will also learn how to carry out math operation using Tensor and Sparse Tensor. It should be emphasied the difference between those two types. For example, Tensor should use embedding_lookup method while Sparse Tensor should use embedding_lookup_sparse instead. Detail
In this example, you will learn how to build a svd model and train it. In logistic regression part, the data is readed and parsed by yourself. While in this part, you will learn the use of tfrecords. It is a serialized data format using protobuf. The data should be transformed into tfrecords firstly and tensorflow has some inner designs to read and parse it. Those designs simplify the input process to tensorflow via Queue and Threads. It also suppliments other convenience such as batch read and shuffle. Detail
This example is similar to singular value decomposition in many parts except model building. And, this example use libsvm data. For sparse data, the process of transforming data to tfrecord and reading from it is something different from the dense one. Detail