Gain insights into industry-level machine learning by applying advanced TensorFlow and Python techniques. Explore efficient methods for creating scalable models, from data pipelines to deployment and inference.

AML-DLI.tar.gz

case-study

In this course, you'll level up your skills learned in the Industry Case Study and Machine Learning for Software Engineers. You'll take the modeling and data pipeline concepts and apply them to production-level classification and regression models for industry deployment, while continuing to practice the most efficient techniques for building scalable machine learning models. After this course, you will be able to complete industry-level machine learning projects, from data pipeline creation to model deployment and inference.

The code for this course is built around the TensorFlow framework, one of the premier frameworks for industry machine learning, and the Python pandas library for data analysis. Knowledge of Python and TensorFlow are prerequisites. 

This course was created by AdaptiLab, a company specializing in evaluating, sourcing, and upskilling enterprise machine learning talent. It is built in collaboration with industry machine learning experts from Google, Microsoft, Amazon, and Apple.

Applied Machine Learning: Deep Learning for Industry

# Chapter Goals:
- Learn how to configure a dataset
- Implement a function that shuffles, repeats, and batches an input dataset

# A. Shuffling
When using a dataset to train a machine learning model, there are certain things we need to do to properly configure the dataset. When we first create a dataset from NumPy arrays or files, the observations may be ordered in a particular way. For example, many data files will sort the data observations by some particular feature, like a person's name or year.

While systematic ordering of data files makes it easier for humans to look over the data, it actually hinders the training of a machine learning model. The model will learn to make predictions based on the ordering of the observations rather than the observations themselves, which is not what we want our model to do. To avoid this, we need to randomly shuffle our dataset prior to training a model, which is done with the [shuffle](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#shuffle) function.

import numpy as np
import tensorflow as tf

data = np.random.uniform(-100, 100, (1000, 5))
original = tf.compat.v1.data.Dataset.from_tensor_slices(data)

shuffled1 = original.shuffle(100)
print(shuffled1)

shuffled2 = original.shuffle(len(data))
print(shuffled2)

In the example, `original` contains 1000 observations, with each observation consisting of 5 floating point numbers. The datasets `shuffled1` and `shuffled2` are the results of randomly shuffling `original` using different *buffer sizes*.

The buffer size (specified by the required argument of `shuffle`) essentially dictates how random our shuffling is. The larger the buffer size, the more uniformly random the shuffling will be.

If the buffer size is 1, then no shuffling occurs. If the buffer size is $\ge$ the number of total observations (the case for `shuffled2`), then the shuffling is uniformly random across the entire dataset. When the buffer size is somewhere in between (the case for `shuffled1`), shuffling will still occur but it will not be uniformly random for all the observations.

In most cases, uniform shuffling is the best option, since that's the safest bet to avoiding any systematic ordering of the initial data observations. However, for datasets that take up a lot of memory, it may be useful to initially shuffle the entire dataset once, then use a smaller buffer size for faster training.

Configure a dataset with shuffling, repetitions, and batch size.

What you'll learn in this course

Data Pipeline

Model Execution

Configuration

Chapter Goals:

A. Shuffling