I believe that Tensorflow certification in 2022 is still worth it. As you can see, there is no other certification in Machine Learning / Deep Learning which tests your practical knowledge in building the Deep Learning models. Again one more thing to note down about this certification is that especially once you start preparing for it, you will also gain a good amount of knowledge, which will eventually help you in your career.
Let’s now see how I prepared for the certification and how it can also help you in terms of preparation. I took the Tensorflow Developer Certificate in 2022 a course in Udemy by Daniel Bourke. A very well-designed course. I see that in many articles, people recommend learning from several courses such as Coursera, some books etc. Yes, they would be helpful, but I think the above Udemy course will provide more than enough knowledge to clear the certification. I took that only course, but I would encourage you to look at other resources.
One drawback of the course is that it’s too comprehensive and overwhelming as the course is around 64 hours, but trust me, guy’s, if you can complete the whole course, keeping the certification aside, you will gain a lot in a sense a lot of knowledge in TensorFlow, the course also includes three capstone projects in different domains, these projects can also help enhance your resume.
The immediate question is, do we need to complete the above Udemy course to take the certification?
Absolutely not. Suppose you are in a hurry to take the examination. In that case, certain sections such as (Transfer Learning, and Milestone projects are a bit more for this certification and can be ignored). Still, I would highly suggest you complete these sections after you clear certification.
Usually, you might think you can register for certification and then start giving it right away. You can do that ( Except in some cases, they need to verify your identity manually, and it takes around two days, but usually it’s done in like 5- 10 min) but believe me. I would suggest you register and pay for the certification at least three days prior to your planned certification date(As early as possible). Even though you register and pay in advance, don’t worry. You will still have time till six months to give your examination. The reason I am suggesting you do it is that once you register for the exam, TensorFlow will provide you with a confidential document which has information on how you can set up your environment and many more things, this document is around 40 pages long, and I would suggest you register early and go through it twice or thrice as it would help you well in term’s debugging the environment if you encounter an issue while giving your exam.
I can say that exam is neither complicated nor easy, but if you took the above Udemy course, I am sure you can ace it. Now let’s come to the point and see the pattern of the exam.
You are expected to build five machine learning models in five hours comprising of :
1. Regression model
2. Classification model
3. Image classification model
4. NLP model
5. Time series model
Questions here are arranged in such as way that they start easy and go hard as you progress. So you are expected to spend less time on the first three questions.
For each machine learning model you built, it’s tested and evaluated for five marks, and there is no substantial pass percentage here, but one thing is sure if you are not able to build one complete model out of 5, then there are high chances that you will not be able to pass the cert. Building a partial one and scoring 3 or 4 in one model might be ok.
So what does that mean? ‘be prepared not to score 5/5’ does it mean you are all going to fail? Nope, the primary goal of telling it is that don’t keep your hopes high that once you build your model and test it, there is a high possibility that initially, you will not be able to score 5/5. You need to tune your model and tune the parameters to increase the score, and that’s the main idea of the cert; it’s not like you build/upload your model and score 5/5.
Some people might again get the wrong idea here that what if I build a model in a very heavy way (Like maybe 5–7 dense layers, with 256, 128 neurons)? Nope, that doesn’t work either because it might lead to overfitting, and still, your model will score 3 or 4 out of 5. So make sure to build a base model with maybe 2–3 dense layers, then increase them slowly. You will have more than enough time to experiment. The most important thing is don’t lose your confidence. Keep trying, keep modifying the parameters, keep tuning, and I am sure you will reach the 5/5 mark.
Initially, no one told me that we would score less. I panicked and was confused since I was only getting 3/5 and 4/5, and this score here means that your model has been properly taken as input, but it’s either underfitting or overfitting. It took me quite a while to figure that out, so after that, it was just fun. I was adding the layer’s kept experimenting, and that’s it. I got the 5/5 in the first four. But the last one was still tricky, and I had to experiment quite a bit. Finally, I was able to get 5/5 for that as well.
No, you are not expected to train the large models (Pretrained). Instead, you need to build your layers, keep experimenting by building multiple layers, adding neurons, etc. You won’t probably need a Pretrained model or use transfer learning.
But I would highly suggest you use a laptop/ PC with GPU or get comfortable with Google colab (Especially uploading the data, getting to know the directory path etc.). You can download the model and put that into the Pycharm directory for testing
I couldn’t disclose more information on these models, but this is how the pattern of the TensorFlow certification would look like
Keras Callbacks :
I felt that specific callbacks are very important, and there is a high chance you wouldn’t be able to clear the cert without using callbacks.
So which callbacks are essential, and in which scenarios should we use them?
1. Early stopping in classification and image classification ( So the model doesn’t overfit)
2. Model checkpoint in regression models ( Since there would be a massive fluctuation in the MAE model, checkpoint helps you pull out the best model with the least MAE)
I think these two callbacks would be more than enough. But feel free to check other callbacks.
Make sure you use a sufficient amount of dropout layer’s to avoid overfitting the model.
Data augmentation :
I would highly suggest you use the augmentation techniques in the image classification, and I firmly believe that without using the data augmentation, one won’t be able to get 5/5 in the image classification question.
I wouldn’t stress you a lot as in the other articles to get very comfortable with Pycharm, but yes make sure you install it two to three days before your examination, run some TensorFlow code, and make sure everything is working.
One most important thing you need to get familiar with in Pycharm is to figure out the path (directory) in which your Pycharm runs, so in case you train your model outside the Pycharm (Google colab etc.), you need to have a clear idea of how to check the directory path and where to place the downloaded model from Google colab. So make sure you research a little on that and get comfortable pulling some models from outside the Pycharm environment, placing them in the Pycharm directory and testing them
This is the Link (Click here)
- 1. Using JSON files is not pretty much taught over the above Udemy course, so make sure you have enough knowledge on working with JSON files (For NLP) — In my advice, take it seriously and work on JSON files
1. I have prepared templates for all the above machine learning models that Daniel taught in one place. Feel free to check it once before the exam. Maybe they can get in handy during your exam —
https://github.com/RupendraRaavi/Tensorflow_templates ( Check these templates)
2. Last but not least all the best for your exam and here is my LinkedIn, please feel free to connect with me.