This never-ending repo includes multiple projects using tensorflow, including CNN, RNN applications. Deep learning is one of my great passions and I cannot stop learning these cutting-edge knowledge. In all the code, I put some important tips and notices as comment. It does take some time to have a deeper grasp of tensorflow. Understanding scope, graph, session, tensors, import_meta_graph, restore is very useful points to study tensorflow.

Please use pip install or conda install to prepare dependencies shown as below:

• python 3.5
• tensorflow
• sklearn
• pandas
• numpy
• pickle

To run tensorflow in gpu, you need to install cuda and cudnn separately from conda or pip install. On windows, make sure cuda pathhas correctly configured as environment variables.

In this project, the program classifies images from the CIFAR-10 dataset. The dataset consists of airplanes, dogs, cats,and other objects.

After preprocessing, the features are normalized and labels are one-hot encoded.

python CNN.py

In this project, the program use skip-gram model to process the data "text8" which can be downloaded from remote server. After two steps of de-noise, word are represented as tokens. By using embedding_look_up, huge compute cost is saved. Finally the words are translated into embedding matrix. Each word corresponds to one row in the embedding matrix. One visualization function can show the relationship between each words by projecting word in higher dimension(#embedding column/feature) into 2D space.

python app.py

In this project, the program trains a RNN neural network to learn writing article. Given training data like text files, books and articles, the program will orgnaize the text into a number of long sequences. Each long sequence will then be evenly splitted into shorter sequences with specified length of characters (also called number of steps). Thus training data will be processed into batches, and each batch size will be #number of long sequence * #number of steps. Input and output pair will be the current character and the following character, ex. "I am good" will have input pairs as ('I', ' '), (' ', 'a'), ('a', 'm')... To avoid gradient vanishing or exploding, lstm cell is used with extral help of gradienct threshold cutting.

The important idea of using RNN to process such problem is the state transportation. The determination of next character depends on all previous characters rather than the current character only. The program can learn a context of given word, sentence or even paragraph because of such feature.

python app.py

In this project, the program trains a RNN neural network to analyze sentiment. The dataset is movie reviews associated with labels(postive or negative). The most difference between other RNN projects is the cost function. Given a sequence of words, usually we take lstm outputs of all the words. However, in this project, the cost function only depends on the last word's lstm output. In this way, the model will be able to read through setence and make judge at end.

python app.py

In this project, the program trains a encoder-decoder model which can learn summarizing paragraphs.

python app.py

Before stably running these programs by gpu version of tensorflow, make sure there are no other interactive session openning like jupyter notebook. Otherwise there will be error such as Blas GEMM launch failed.

• Tianwen Chu - Initial work