TensorFLow is an open source software library for numerical computation using data flow graphs.
The overview of a similar graph of a three-layer neural network in TensorFlow shows below. Nodes in the graph represent mathermatical operations, while the graph edges represent the multidimensional data arrays(tensors) communicated between them.
static graph vs dynamic graph
- static graph: you define graph statically before a model can run. All communication with outer world is performed via
tf.Sessionobject andtf.Placeholder - dynamic graph: you can define, change and execute nodes as you go, no special session interfaces or placeholders.
For production:
- efficient
- Scalability
For research:
- Flexibility
Other reasons:
- Popularity
Basic elements:
- graph
- session
- tensor
- operation
- Phase 1: assemble a graph
- Phase 2: use a session to execute operations in the graph