學生對 deeplearning.ai 提供的 Improving Deep Neural Networks: Hyperparameter Tuning, Regularization and Optimization 的評價和反饋

The class should include more introduction on the current ml frameworks such as tensor flow etc. Possibly it should include one more project for the ml framework. Hope to give students more experience on the ml frameworks.

Good and detailed information on how to tune parameters, optimization techniques and regularization. I'm confident that this course learning will help me in training NN to better convergence in a shorter time than earlier.

I would have liked more details on the math. Furthermore, I think that the discussion of TensorFlow was a bit too short. Although I was able to do the assignment I have not yet developed an understanding of TensorFlow.

Excellent. Mon regret est que l'exercice final ne mette pas en oeuvre le tuning des hyperparamètres sur un jeu de cross validation. Un exercice supplémentaire avec TensorFlow ou Keras sur cette notion aurait été un plus.

Very Interesting

As usual, it is very good from theory point of view. Practical examples are also really interesting.

Do not expect to be autonomous after the course, as you won't be able to use Tensorflow or Keras alone.

I'd make the tesnsorflow section a separate week with much more elaboration, the first time (in both course 1 and course 2) I felt a subject was lacking information. It's mostly noticeable in the programming assignment.

The course really takes the student by the hand through the exercises. The disadvantage is that it is not really necessary to understand what you are doing. Just follow the guidance. But on the whole really satisfactory

Very exciting and interesting course overall but the programming assignment with Tensorflow was not practical in my opinion. Instead having practical experience building NN with Keras might have been the better choice.

Generally the course is very good! I liked that I could manually implement the steps of hyperparameters tuning. I wish there was a bit less boilerplate code. Implementing everything from scratch would be more valuable!

Could be Greatly improved by having us build a NN using previous learning's with the only change being use of SoftMax for Cost. Then have us use TF to do the same and compare the code effort, and the results 1-to-1...

Muy buen curso! Me hubiese gustado que se desarrolle un poco más TensorFlow. No me quedo claro por ejemplo, cuándo hay que inicializar variables, si es realmente necesario definir las constantes con tf.constant, etc.

The explanations are amazing. I do not qualify with 5 stars as I think that practice can be structured in a better way. If the practice is done after each module in each 'week' it would help to retain main concepts.

Good course overall. but labs could be expanded. Don't know if the Coursera platform supports it, but labs between lectures about different topics would be nice instead of having all practical exercises at the end.

Nice course, as always!

But I think the hyperparameter tuning methods are hopelessly outdated / missing the most promising current developments. A pity since this is such a central part of the actual work with DNNs!

You can get a better insight as to how to improve neural networks that go beyond the fundamentals. The quizzes and assignments helps you get a hands-on experience of the theoretical material covered in the course.

Hi

I enjoy the course a lot!

for tensor flow - I am not sure if its me or the course - but I need much more training to start thinking the tensor flow way. maybe i will practice more on real work cases.

thanks !

Oriel

You've learned deep neural nets but on the first problem you apply them to they seem to not work or learn to slowly. Don't panic, all you may need is a little fine-tuning, that is what this course will teach you.

After this second course you will be able to start build things using Tensorflow. Really great to see how good this course is structured. Things from course one is comming back making it easy to grasp new content.

Overall, the course is interesting and introduces systematically technical details. There are still some confusing part in the assignment. For example, the direction in the last assignment is kind of misleading.

The explanation and examples are very informative throughout the course. The quizzes and the assignments are highly related to the topics covered in the videos which provide a solid understanding of the course.

Some very interesting consideration, though I would have liked a section about reproducibility and randomisation (including for GPU trainining), though I understand that this is framework and language dependent

Some of the programming assignments are a bit confusing, and the grader seems to suffer from bugs at the moment. Nevertheless I found especially the part on optimization algorithms very helpful and interesting.

Maybe TF should be thought a little earlier with small exercises in the weeks 1 & 2. Also the final programming assignment should be improved. The seed initialisation at the Xavier initializer is ambiguous.

The insights and intuitions Andrew communicates are good, but as he starts to point out towards the end of this course, in practice one uses a DL Framework and you don't code these things from the ground up.

The level of detail in this course really leads to a good understanding. A bit more programming exercises with TensorFlow (more than a single model) would be good to understand the intricacies a bit better.