TensorFlow Developer certification was my third certification after Google Cloud Associate Cloud Engineer and Professional Data Engineer certifications. Tensorflow is one of the most important Machine Learning ecosystems and its developer certification helps showcase expertise in Machine Learning and Tensorflow, of course, in particular.
If you have built basic models on Tensorflow, then cracking this exam, which tests foundational knowledge on building and integrating models with Tensorflow, and getting added to Google’s Certificate Network will be easy for you.
Here is a quick examination guide.
The certificate exam tests your ability to build models in the following domains:
- Simple regression models
- Complex models with DNNs
- Image recognition with CNNs
- Natural Language Processing with Text Preprocessing, LSTMs or CNNs
- Time Series prediction
Course for the Certification
As recommended in the Candidate Handbook for the exam, it’s better to take the Coursera Specialization course: Tensorflow in Practice. The course covers all the domains required for the exam. In addition to covering the topics for the exam, it also teaches newer concepts in Tensorflow 2.0 like Keras preprocessing utilities and API.
If you are coming back to Tensorflow after some time, this course teaches newer and simpler ways of doing things. And if you are experienced in Tensorflow, I recommend that you take the course exercises, at least.
Since I work with Tensorflow on a daily basis as a Machine Learning Engineer, I took the programming exercises at the end of each week in the specialization. Whenever I would get stuck, I would go through the Tensorflow documentation and API reference to finish those exercises. Doing those exercises prepared me well for the exam.
The examination costs $100. It is online only. You can take exams from anywhere with these prerequisites:
- A stable internet connection
- PyCharm IDE installed on your system
The exam duration is five hours and it is more than sufficient time to complete the exam.
When you start the test, you will have to install a PyCharm plugin that provides the programming questions for the examination. The plugin also creates a python virtual environment and installs Tensorflow version 2.0 and the necessary packages required for the test.
The question has the skeleton of the program, such as downloading the dataset and basic preprocessing steps. You will have to complete the script to build a Tensorflow model and train the model on the dataset. The model will be saved as an h5 file.
The datasets will not be huge enough to require a lot of time for downloading nor will they be too complicated to require a lot of time for training. You will be able to easily build and train the model in five hours.
Now for the evaluation — it is not based on the code that you write in the script. It is completely based on the models that are trained and saved. When you submit a solution, the test plugin will upload the saved model file and then evaluate based on some test set inputs hidden from the user.
For each model, a score out of 5 is presented, like 3/5. Based on the result, you can try again with a different model architecture or try training with different hyperparameters, use data augmentation, and more. You can submit the results for each category any number of times. Once you will have achieved satisfactory results in all the question categories, you can submit the whole exam. At the end of five hours, the exam will be submitted automatically.
I completed the exam in 3 hours after I got 5/5 on all 5 categories. Initially, I had scores like 2/5, 3/5, etc. But I kept trying different approaches till I got 5/5 on all of the categories. Since I got 25/25, I was confident and went ahead and submitted the exam.
As mentioned above, the evaluation is not made on the basis of the code that you write, it is based on the models you build. So, it is important that the model generalizes and does not overfit the training data. You can use data augmentation and regularization techniques to avoid overfitting and generalize the model better.
It will help a lot if you get familiar with the PyCharm IDE. Getting to know the shortcuts for quick documentation, function call signature will save a lot of time during the exam. It is also good to have the Tensorflow API reference open by the side to refer to when needed.
Since this is an online test, you can use AI explainers like LIME, Shapley to debug your model and retrain it based on the explanation. This will also save a lot of time in understanding the model if you keep on getting lower scores. But if you plan to use it, you need to get the explainer setup ready before the exam, so that you don’t end up losing time setting up.
So, this is my 2 cents from my experience of preparing for this certification with like-minded engineers driven by #LifelongLearning.