Tutorial 3: Operator¶

Overview¶

In this tutorial we will introduce the Operator - a fundamental building block within FastEstimator. This tutorial is structured as follows:

• Operator Definition
• Operator Structure
• Operator Expression
• Deep Learning Examples using Operators

Operator Definition¶

From Tutorial 1, we know that the preprocessing in Pipeline and the training in Network can be divided into several sub-tasks:

• Pipeline: Expand_dim -> Minmax
• Network: ModelOp -> CrossEntropy -> UpdateOp

Each sub-task is a modular unit that takes inputs, performs an operation, and then produces outputs. We therefore call these sub-tasks Operators, and they form the building blocks of the FastEstimator Pipeline and Network APIs.

Operator Structure¶

An Operator has 3 main components:

• inputs: the key(s) of input data
• outputs: the key(s) of output data
• forward function: the transformation to be applied

The base class constructor also takes a mode argument, but for now we will ignore it since mode will be discussed extensively in Tutorial 9.

In [1]:

class Op:
    def __init__(self, inputs=None, outputs=None, mode=None):
        self.inputs = inputs
        self.outputs = outputs
        self.mode = mode
    
    def forward(self, data, state):
        return data

class Op: def __init__(self, inputs=None, outputs=None, mode=None): self.inputs = inputs self.outputs = outputs self.mode = mode def forward(self, data, state): return data

Operator Expression¶

In this section, we will demonstrate how different tasks can be concisely expressed in operators.

Single Operator¶

If the task only requires taking one feature as input and transforming it to overwrite the old feature (e.g, Minmax), it can be expressed as:

If the task involves taking multiple features and overwriting them respectively (e.g, rotation of both an image and its mask), it can be expressed as:

Multiple Operators¶

If there are two Operators executing in a sequential manner (e.g, Minmax followed by Transpose), it can be expressed as:

`Operator`s can also easily handle more complicated data flows:

Deep Learning Examples using Operators¶

In this section, we will show you how deep learning tasks can be modularized into combinations of Operators. Please note that the Operator expressions we provide in this section are essentially pseudo-code. Links to full python examples are also provided.

Image Classification:¶

MNIST

DC-GAN:¶

DC-GAN

Adversarial Hardening:¶

FGSM