Unlock the full potential of AI with Building LLMs for Production—our 470+ page guide to mastering LLMs with practical projects and expert insights!


My Journey To Becoming a Tensorflow Certified Developer

My Journey To Becoming a Tensorflow Certified Developer

Last Updated on November 3, 2022 by Editorial Team

Author(s): Pere Martra

Originally published on Towards AI the World’s Leading AI and Technology News and Media Company. If you are building an AI-related product or service, we invite you to consider becoming an AI sponsor. At Towards AI, we help scale AI and technology startups. Let us help you unleash your technology to the masses.

Passing the TensorFlow Developer Certificate exam is a goal for many people who want to get into deep learning. I would like to explain how I managed to pass without having previous experience in Machine or Deep Learning, much less with TensorFlow.

My starting point.

I’m an engineer with a good background in C++, and I’ve been a developer for a long time, but some years ago, at the beginning of my career. I've spent the last 15 years working with ATMs, the last five leading a cybersecurity team. As you can see, nothing related to AL or Machine Learning.

But I have been studying and training throughout my career. The last years studying video game development. In this way, I knew the development of characters moved by artificial intelligence using reinforcement learning. This was my whole experience with AI before deciding to make a change in my career.

It is true that I knew the concepts of Training, epochs, steps, and hyperparameters. But my experience in TensorFlow and in solving problems of NLP (Natural Language Processing), Convolutional networks, or series forecasting was totally null.

Picture Samuel Bourke on Unsplash

Why the TensorFlow Developer certification?

TensorFlow is Google’s library and, in fact, is almost a standard in the world of Machine Learning.

It can be used from any of the frameworks of the big manufacturers, be it IBM, Microsoft, or Amazon.

The certification, it seems to me, has a good reputation. It’s not a test where they ask you questions and you give them answers. It’s a test where you have to make models to solve the problems they give you with the information they give you.

Besides, it is a long exam, lasting five hours at most. There are people who finish early and people like me who reach the end with 10 minutes to spare. For the last half hour, I’ve been working to improve a couple of models to increase the note.

How I prepared for the TensorFlow exam.

I would like to say that although I have followed what could be called a central training itinerary, I have also been carrying out secondary activities. Occasionally, it’s good to take a break from studying the course material and spend time watching YouTube videos, studying Kaggle code, or reading related articles. These things have helped me a lot with passing the exam.

I strongly recommend that you take courses from different instructors. Each person has their own techniques and way of explaining things, and I’ve been able to pick up some little tricks from different people. Staying with the ones I liked the most or perhaps what I understood the most.

Courses Taken.

Deeplearning.AI TensorFlow Developer Certificate on Coursera.

The DeepLearning.ai specialization is the official course to prepare for the TensorFlow Developer Certificate exam. The course is given by Laurence Moroney. If you look at the certificate that I put at the beginning, it is signed by him. So, it’s clear that he knows what he’s talking about.

The specialization consists of four courses:

  • An introduction to TensorFlow.
  • Convolutional networks for image classification.
  • Natural Language processing, where both sentiment analysis in the text and predictive text generation will be discussed.
  • Series Prediction.

They match up with the parts of the TensorFlow exam that need to be looked at.

Each course is divided into four “weeks”. In each of them, we find videos, some light reading, a couple of exercises, which are done in Google Colab, and a final assignment.

That is, for each course, you will have about 8 / 10 exercises and 4 assignments that you must do in Google Colab.

Although they recommend one month per course, that is, four months for the entire specialty, it can be finished much earlier. I think that you can finish the specialty in a couple of months if you spend about five hours a week on it.

Verdict: Essential.

Machine Learning Crash Course on Google.

I guess most people would recommend taking this course at first. I do not. It’s excellent but more boring than Coursera’s specialization. I did it just after finishing the Coursera specialization.

The good thing is that everything sounds familiar to you, you already know it, and from time to time, they explain something to you differently, and your head clicks… and you think, oh dear… OK!

The course is made up of videos, texts, questionnaires, and projects in Colab. The content is really splendid. The format is not as nice as Coursera’s. Every so often, it was hard for me to know where I was left. The videos are a little older. But for me, doing it right after specialization was perfect!

Verdict: highly recommended.

Intro to TensorFlow for Deep Learning on Udacity.

I took this course more or less the last 10 days before the exam, and I think it was a great contribution to the result.

Not only because of the content but because it gave me the confidence that I was capable of solving problems and that I fully understood everything that was explained to me.

It is a quick course to do, one or two weeks. But what explains the concepts in a clear way even gives you time to deal with topics such as Transfer Learning. That is also covered in the Coursera specialization, but this course explains a simpler way to work this technique.

Convolutional networks, NLP, and series are touched on. The Series part is given by Tony Moses, and I thought it was excellent. Not only because of the explanation in the series forecast but because you know his little tricks or his way of working. The course, on the other hand, follows the same format as Coursera, a mix of videos with notebooks on Google Colab.

The course is done very quickly. Especially if you do it at the end since you will be able to solve the problems raised, but you will discover some new way to do it.

Alternative training.

The first recommendation is Machine Learning Foundations by Laurence Moroney. It is a must-have YouTube list. It contains much of the first two courses in Coursera’s TensorFlow specialization.

It is a good idea to go through this YouTube list beforehand so that you can face the specialization with some prior knowledge that will help you advance more quickly.

Other YouTube channels I liked were Greg Hogh and Nicolas Rennote. In both, I found videos that explained how to solve problems of any kind with Tensorflow.

My recommendation is that if you are weaker at some point, such as predictive text generation, look for videos on YouTube and see how they solve them.

Possibly, they give you a different vision than the one explained in the course you are doing. Or maybe they just explain it in a way that you understand better.

The best thing to do in Kaggle is to look at the competitions under Getting started and study how the notebooks are solved. I would focus special attention on Digit Recognizer and Natural Language Processing with disaster tweets.

The environment and the exam day.

One of the things that worried me the most was that the exam was done in the Pycharm development environment. An environment that I have never used and that I can now say I continue without having used. Don’t worry about Pycharm!

We need Pycharm to install the exam plugin. That will create a virtual environment with the exercises to be solved. But I did 100% of the work in my Jupyter local and Google Colab.

You must save the models in .h5 format and copy them to one of the directories created by the exam plug-in.

Then, you can hit the evaluate button, and Pycharm uploads the model and evaluates it.

The name of the model is not relevant, but you must have only one model in the directory. You can send the model for evaluate as many times as you want, so you can carry out tests and keep the best model.

The models score from 0 to 5. I would recommend passing all the problems quickly with a 4 and then, if you have time left over, dedicate yourself to improving your grade.

You have to be careful when sending the exam to be evaluated. We must be sure that we have the desired models in each directory. Those who are currently in the directory are evaluated, and not those previously evaluated, even if they had a better grade.

The exam lasts five hours, but it doesn’t start counting until you have everything ready. Don’t worry, it will start counting when you have the plugin correctly installed and click on the start exam button.

I used two computers: a Mac with an i5, where I configured Pycharm, with the recommended libraries, and another MAC, but Silicon, with a Jupyter environment created with Conda with the recommended libraries.

I do not recommend installing Pycharm and trying to create the exam environment on a Mac Silicon. I was unable to install TensorFlow 2.9 on Pycharm in the MAC Silicon.

But I worked on the MAC with Silicon and passed the .h5 model to the Intel MAC, where the exam environment was correctly configured with Pycharm.

Taking The Exam

I fixed the first three problems very quickly and the fourth, too, although it took me a little longer. The truth is that I had a mark of 5, 4, 5, 4, with 5 being the maximum possible mark. I still had two and a half hours to go, and I set out to tackle the fifth problem.

Well, in this fifth, I ran into an issue that took me about two hours to solve. I couldn’t get the model to pass the exam validation correctly. Finally, when I managed to understand what I was doing wrong, I passed it with a 3. Very Fair, but something is something. I had half an hour left and decided to try to improve the two fours. I was already exhausted from fighting with the fifth problem.

Starting with the first problem, I was quite relaxed. My biggest concern was that some exam setup was not working. Everything seems to be working fine.

As I progressed, I noticed that I would really be able to pass, but the fifth model… I even thought about leaving it at 0. It was five hours in which I hardly ate. I didn’t stop to eat. I had something to snack on on the table, my bottle of water. My two computers are turned on and ready. Five quite intense hours after some intense days giving the last reviews.

I suggest that you prepare the environment, ensuring that it is a comfortable one. I chose a holiday when I was home alone, and I was really nervous.

The exam is not complicated, but you do have to show that you know how to solve problems with TensorFlow in the fields of image classification, language, and series. In some problems, they can incorporate some extra difficulty, as happened to me in the fifth.

Do not worry. If you are prepared, you will pass. You don’t pass the exam with luck, but I wish you all the luck in the world!

I will be waiting for you in the Directory of TensorFlow Certified Developers.

Pere Martra TensorFlow Developer Certificate

My Journey To Becoming a Tensorflow Certified Developer was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.

Join thousands of data leaders on the AI newsletter. It’s free, we don’t spam, and we never share your email address. Keep up to date with the latest work in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI

Feedback ↓