Classes

A dynamic graph is a workspace for computations. All tensor variables can be tracked from a dynamic graph.

A model is a base class for stateful operations on a dynamic graph. It can be use to construct computations statically, thus, more efficient.

A model builder is a more generic type of model. A model can be quite static, thus, you have to be quite careful to have a model work with dynamic inputs. You cannot use reshape, or anything that can generate fixed tensor outputs from a fixed inputs.

A model builder on the other hand doesn’t have that restriction. When input changes, it simply calls the given builder closure to construct a new model. In such way, you can continue to use reshape etc to assume fixed inputs and outputs, it will just work for dynamic inputs. The newly built model will carry over stateful components (parameters) from the old models, thus, it doesn’t reset your training. This also means you need to make sure parameter shape won’t change when input changes, otherwise we will fatal.

Sum inputs.

Add two inputs together. It will do broadcast if needed.

Multiply two inputs together. It will do broadcast if needed.

Matrix-multiplication over two inputs.

A linear layer model.

A reshape model.

A ReLU activation model.

A softmax activation model.

A sigmoid activation model.

A tanh activation model.

A swish activation model.

The masked fill model. If the value equal to a given constant, fill with another constant.

The dropout model.

Multiply all values with a constant.

Batch normalization model.

Layer normalization model.

Make the input tensor to be 1-D tensor (respecting N).

Convolution model.

max pooling model.

average pooling model.

reduce sum model.

reduce max model.

reduce norm2 model.

min model.

max model.

Concatenate model.

LSTM model.

Embedding model.

IndexSelect model.

A type-erased model builder.

Model Inputs for Functional Model

A stream context is an object that an execution can be performed upon.