training: Various algorithms of the Optimizer, learning rate decay, basic_train_loop, session, checkpoint, processing derivatives and gradients, queues, distributed execution
ops: bound c++ operation
framework: bindings to c++
client: handle session
Estimator: the abstraction of the estimator, the Estimator wrapper class, the input queue cache
models: the implemented model
contrib: high-level abstraction
layers: layers
nn：Neural Network
Contrib details:
tf.contrib.bayesflow.entropy Shannon Information Theory
tf.contrib.bayesflow.monte_carlo Monte Carlo integration Monte Carlo integration
tf.contrib.bayesflow.stochastic_graph Stochastic Computation Graphs Stochastic Computation Graphs
tf.contrib.bayesflow.stochastic_tensor random tensor
tf.contrib.bayesflow.variational_inference Variational inference
tf.contrib.crf CRF layer conditional random field (conditional random field)
tf.contrib.ffmpeg ffmeg codec audio
tf.contrib.framework parameter scopes, variables, checkpoints
tf.contrib.graph_editor modifies the computational graph at runtime
tf.contrib.integrate.odeint ode solves ordinary differential equations
tf.contrib.layers build layers, regularization, initialization, optimization, Feature columns (mapping between data and models)
tf.contrib.learn Advanced Learning Library
tf.contrib.linalg linear algebra (matrix)
tf.contrib.losses loss function
tf.contrib.metrics metrics
tf.contrib.distributions probability distributions
tf.contrib.rnn rnn related
tf.contrib.seq2seq implements codec based on rnn
tf.contrib.staging.StagingArea add pipeline
tf.contrib.training mini batch and group (bucket)
tf.contrib.util
tf.contrib.nn Sampling clipping information entropy
概括：losses layers training learn rnn seq2seq