This was an interesting course. It provides a high level look at face recognition/verification and various state-of-the-art aspects of convolutional neural networks. The one thing I found frustrating in this course was the grader. It was very particular for at least one homework assignment on the order in which you entered your variables. I spent way too much time on debugging for simple things like that.

The content was great, and is probably the best available. However, the grader was so flaky it really shook my confidence in the material. I'm the type of person who will try and try until I'm literally about to give up before I look for help in the forums, so I lost a LOT of time on these exercises. This was by far the WORST of the five courses in the specialization. Sorry to yell, but YOU CAN DO BETTER!

The Keras and TensorFlow versions used in this course are by now to a large degree outdated. The Newest TF version is at the date of writing 2.3, while the course uses <2.0, so many of the functions used are deprecated in the newer versions

However, Andrew's explanations are great as always except for the convolutional implementation of sliding windows in the 3rd Week. (therefore only 3 stars this time)

Depplearning.AI: Please do not release content unless it is ready. The content is fine, but the assignments were clearly hastily put together and had basic bugs discussed all over in the forums. In particular, week 4 is a complete mess. Boiler-plate code that doesn't even relate student-content (to load a dataset) doesn't even run for many people. This wastes everyone's time. Really disappointing.

Overall a great intro to CNNs. But the last part of the course on object detection and facial recognition is very superficial. It explains the logistics of the disciplines (how to keep track of bounding boxes, etc), but it doesn't teach how to actually build such a system. The exercises make you fill in a bunch of indices and do vector math, but deliver the actual hard parts inside black boxes.

Very good covarage of the algorithms when it comes to analyzing pictures, and a good intro to the theory behind the models. But it is too little emphasis on other uses of convolutional networks like 1d convolutions, causal convolutions and similar. Maybe there are some coverage of these topics in the sequence course in the series, but it should be covered here to a larger extent either way.

All video content of this course where great, but i can't say it about programing assignments. YOLO and Neural Style transfer are by my opion advanced topics. I would more appreciate longer programming excersice, not only something where i only add some piece of code and i hardly understand what is going about. For example, convnets were clear, i could implement it by myself, but yolo no.

Course lectures and questions are very good. The programming assignments are also good questions wise, but the grading mechanism is quite annoying. We had to find out clumsy workarounds to get the correct grading, in one case, the code change needed for getting the correct grade did not result in the expected output. Coursera needs to change the method of grading programming assignments.

In this course the material/resources regarding Keras is highly insufficient. So far I have made little progress in the programming assignments related to Keras and I achieved this by figuring out stuff by googling which took a long time. If the instructors could extend the course with stuff related to Keras it would be great. At least there should be references to sufficient resources.

I can't believe the number of technical problems I've had with notebooks not saving my work on homework assignments. It's very frustrating. The content is good and I will continue with the course, but this is the first Coursera course I've had (actually, the whole series in this topic) where I have experienced the lost of work - I just lost about an hour on a homework assignment.

I am very annoyed with the evaluation of the notebooks. Not with the content itself, but with the support from instructors, which is non existent.

It took me two days and 10 re-submitions to solve a problem that was unrelated to the code, but to the behavior of the grader system. No one replied my cries for help in the discussions.

Very disappointed with the lack of support.

The content of the course is very good, as with all the Andrew Ng / deeplearning.ai material. However production standards seem to have slipped for this one. Repeated sections in video material and a final notebook exercise that contained errors and required finding a work around that was posted in a discussion forum. I hope these things can be corrected for others.

I have mixed feelings about this course in particular, although one learns many interesting and useful concepts, I did little implementation on my own. Also the involvement of Keras I found annoying, yes it eases the implementation of ConvNets, but while learning I would have preferred to use tensor flow instead, or even implement a simple NumPy ConvNet on my own.

The content is very good. The exercises are a bit useless. Don' expect to be able to use tensorflow after this course. Furthermore, they teach the syntax of tensorflow 1 rather than the new 2. Therefore, when you try to solve the exercises you don't understand the discrepancies between the online documentations and what they want for these exercises

Excellent course! I have learned the skills to combine image processing with machine learning.

However, the assignment of the Week 3 has a problem that you have not fixed for a long time, and thus it wasted some time. And the assignment in Week 4 has problems of submitting, and that's why I only rated three stars. Hope you can fix the problems soon.

First, there should be an upgrade to TF 2.0. In at least one instance the documentation for a function was non-existent. Second there are many places the videos can be cleaned up eg. transcriptions are just wrong like a machine did it, or the speaker repeats the same thing twice in rapid succession. Overall the course felt unpolished and dated.

The Face_recognition assignment was a tough one to solve, i only got grader problems but still i wasnt able to figure out how to pass grader even though my code yieded right answers. That's my only complaint otherwise the videos from Andrew were really easy to understand and the programming assignments were very well documented. Thanks :)

The assignments are less copy paste and some allow the student to explorer different NN architectures. However, most of the videos are still a waste of time. And the methods needed to complete the assignments aren't taught to the student. Instead you have to spend a lot of time searching and hoping you find the right method.

video courses were really good, but the programming assignments drove me nuts. I am a senior software developer and I am writing software for more than 10 years now. I had a really hard time understanding the Tensorflow code. I think it is better to suggest a student to learn the basics of Tensorflow before doing this course

This course is well organized with CNN knowledge. However, it seems like the team is overwhelmed to maintain both a big Tensoflow tech stack in programming assignments while keeping academic science core concepts. Some code comments are outdated and the mentors definitely can not follow up all the issues raised by students.

Couple of suggestions: 1) fix the artwork via neural networks assignment as there is a bug in your code 2) With the lectures I would suggest that you do a summary explanation of how the whole process works (all steps and motivation - a review) at the end of each group of lectures (one for artwork one for face recognition)

I think this course lacks hands-on experience, and for that I think it should improve the labs, people would learn much more from a youtube-kind tutorial than from these "#your code starts here / #your code ends here"-labs.

The quiz interface is horrible, i need to refresh my browser many times before it loads properly.

The theory is very good but the exercise part is not good enough for me (For example in the Face Recognition exercise, I'd like to build (even a simple model) and train the triplet loss function... However, all that I can do is only loaded a trained model and then apply some simple similarity measure on encoding vector)

There is major room for improvement on the automatic grader, under some particular cases, the answers are correct but the grader will give you zero with no feedback (in my case, I made a mistake with declaring local variables as global in the code). This issue is quite frustrating, the forums helped a lot though.

Great class and amazing assignments. I really enjoyed learning about CNNs, YOLO, and Neural Style Transfer.

Errors with submitting the assignments, particularly weeks 2 & 4 took away considerably from the overall satisfaction with the course.

Thank you once again for providing a rich learning environment. :)