I think this course is fantastic - given you come equipped with the right expectation and prerequisites. I'm new to python programming and this was definitely a step up compared to the previous courses in this certification. I have a solid background in all things matrices - so the endless discussions of dimensions did not really help me that much, but I can see how they would really help others. IMO at the end of this course, I'm not automatically equipped to build my own ConvNet, but I would definitely know where to start, and how to learn (read the right papers, get the right code off github, start from there...) if I wanted to set out to do so.

great course from andrew ng, though i would like the programming assignments to be a bit more <<hardcore>>. i mean to do the whole work from scratch and not just load some huge models, though i get it, to train those huge nets the students would need a lot of time waiting. Still some modules i got my hands on, in some utils.py files, were really interesting and i think it would be really educative to write some code on them too.As a suggestion i would say since python is used the videos should focus a bit moew on the libraries, at least on the really important stuff like import pandas.

overally, great course! i recommend it, without any doubt.!

The content is amazing and very informative. The content is clearly explained and I particularly liked the references for each paper in order to get a more thorough understanding of the models in the lecture.

On the other hand, the video editing needs some taking care of since professor Ng is often heard repeating the same sentence which is often confusing. Also the weekly programming exercises, excluding the first week, are far too easy and require mostly copying and pasting.

As a matter of personal preference I would also prefer if we used the same library for all assignments rather than having to alternate between Keras and Tensorflow

Thanks for putting together this great course. I just finished the course and my initial review of this is that it was more exhausting than the previous courses. The Course is well defined and helped me understand how CNN is used and what are some of the problems that can be addressed through CNN. However, there were moments during the exercise where I found applying the concepts challenging. I may have to relearn, apply and practise more on my own to get to really understand more of this. However, this course has given me a great start on how I can address some of the issues and how this part of deep learning is applied. Thank you.

It's an amazing topic and the course for the main part is well put together. Some of the videos haven't had the retakes removed so Andrew literally repeats himself. I thought my internet connection was playing up for a while. Would be nice if they could clean this up to not waste everyone's time. Some of the tutorials can be confusing/frustrating if they don't work first time. The style transfer one in particular, the instructions don't make that much sense. They say add 2 lines but 2 weren't required etc. Error messages are confusing, and rerunning the example is error prone but this is more Python's fault than this course.

Andrew has a very good video-lecture style.

The programming exercises can be a bit frustrating at times for the wrong reasons, but at this point the course has been available long enough that you should be able to find a thread in the Discussion forum that provides enough hints to resolve any issue you might encounter. Nonetheless I appreciate the effort that went into designing the programming assignments.

As others have noted the video editing is surprisingly poor, with brief clips that should have been cut scattered throughout, but ultimately it doesn't detract from one's ability to absorb the content, so not a huge deal.

The material presented in the class is very interesting and useful. The explanations are clear and the examples are good, especially teaching us to use transfer learning based on pretrained models. The programming was very helpful.

I couldn't give a 5 though, because I spent so much time on the programming assignments due to errors in the assignments, the grader and the hints. Additionally the mentors need to monitor the discussions in the week's section. Some students' comments were helpful, but others were wrong and completely off the rails. Some mentor feedback would really have helped all of us.

Absolutely great content - many important computer vision papers discussed in an approachable way that highlights differences in approach.

However, the Nov 2017 version of the course suffered from at least two grader errors: in once instance requiring literal following of instructions that would lead to an implementation that is faulty under real world conditions or a broader test set, in another instance deviating from the given instructions in a rather random if minor way. Hopefully these will be addressed in a later session - if not, be prepared for a bit of forum reading and trial and error to pass.

There were some problems to get the graded functions through the grader, although they were actually correct. Having the grading system a bit more flexible would save everybody's time.

I didn't understand all the operations we did in tensorflow in the neural style transfer programming assignment. How did we choose the arguments for the "assign" function calls? How did we tell tensorflow to minimize the total cost by modifying the input image, not the weights/biases of the model as it happened normally? Although the theory seemed to be clear, I didn't really get how we did it with tensorflow.

I think the assignments for this course could be structured to help students learn better. Specifically, I think there was not enough instructions on the difference between Keras backend and tensorflow. This was confusing at the beginning. Also, I think the codes snippets that are left as None sometimes help too much that there is not much thinking required, so I found that it becomes a lot harder to think what line of code is needed when there is not as much instructions. But all in all, I rate the assignments very highly and these are some of the ways that I think can improve the course.

Very informational course and the pace is reasonably challenging. Knowledge learned in this course are super practical and can be directly applicable in many areas. The Keras curriculum is especially welcoming. Do take note that course 2 and TensorFlow are sort of hard requirements to be successful in this course. I did course 1 and this course together, and it turned to be not the best choice.

Also, the risk of taking a new course is there were definitely several technical glitches which resulted in more troubleshooting than necessary. I believe this would be better with iterations.

Good introductory course for ConvNet and its trending applications such as object detection and facial recognition. Materials are presented to give students more of an intuition and process to carry out ConvNet applications rather than a rigorous mathematical understanding. Basic TensorFlow knowledge is highly recommended or one may face difficulties or confusion during assignments. I personally would like the course materials to have more depth, so to really nail in every step in building a ConvNet application (since as of now the content can be surface-deep and easy to forget).

The course instructor did a good job in presenting the basics of CNN and some of its applications in the domain of computer vision. The applications presented are 1) Image classification 2) Object detection and 3) Neural style transfer. As a researcher myself, what I most appreciate is that the instructor presented the topics to sufficient depths enabling the reader to appreciate the underlying theory and at the same time keeping it high-level. For those who would like to go more in depth, the relevant citations are presented in the lectures. Overall a very satisfying course.

Great content! Andrew really makes the concepts crystal clear. The lectures are very coherent and extremely organized, in terms of the actual contents. I've took MOOC in CNN before, and personally felt that Andrew's version is the best.

I took one star off for two reasons: 1) Sometimes Andrew would say half of the sentence and then start over -- it seems to me that those were to be taken off during editing of the video, but they are not -- hopefully they'd get fixed at some point; 2) It would be nice to upgrade the code to TF2, as TF1 has become less popular for a reason.

Andrew is great as always, but the assignments are no so great. It feels like you are doing a little piece of useless code to the full application. The order of the functions and the extensive use of Keras/TensorFlow is frustrating and very confuse. Maybe, in the future, you can add some explanations about these tools. Also, given the assignments the weeks is no longer 4/4.5h, it can take hours just to understand some silly syntatical tf related errors. But I now know how to get started with my project and where to look for hints and improvements thanks to you guys.

The content is really great. It is giving very good overview on the state of the art, and how convolutional neural networks can be useful. I think it is hard to get such a great overview in current deep learning books that usually focus on more theoretical aspects, which are covered in this course. The only negative point I would say is that is it not always easy to understand how to use some very specific python tools, and one can easily get stuck into implementing a single line of code. However, the discussion forum provides great resources to solve these issues.

Excellent content but platform was frustrating at times. The exercises still use TensorFlow 1.2.1, which created some aggravation because newer versions have rearranged many functions into tf.math.xxx, etc., so the documentation "hints" give answers that don't match what is needed for the exercises. Also, the grader is maddeningly finicky, giving no points for code that is in fact correct. It's like playing "Simon Says". It seems that a decent chunk of the required effort consists of scouring through the discussion forums for workarounds to these glitches.

This course gives a comprehensive introduction to CNN. However, I am not satisfied with the exercises designed for this course. In each assignment, you are only required to fill in a few blank spaces, leaving a lot of important parts as black boxes. Please make the assignments more like complete projects. I know it will take people much more time to complete and may turn some people away. But it is crucial to give a complete picture of each programming exercise assigned. Anyway, I still want to thank Andrew and your team for offering this series of courses.

The material was very interesting and the technologies introduced were very good. The only problem is that unlike the previous 3 courses, this one seems to have been done in a rush.

The video aren't edited that well. (there's weird sentence repetitions, and the coding parts are sometimes bugged up, very very hard to get to the right answer, and had a few typos. Hopefully those little bugs will go away, but overall it was an informative course. Some of the topics were so interesting, I felt like they would entertain an audience uninterested in deep learning,

Positives:

1) Well designed course that takes you through the concepts of CNNs step by step and introduces cutting edge state-of-art applications based on it.

2) As always well prepared lectures effectively deliver the course material.

Negatives:

1) Course lectures should have covered overviews of actual models used in assignments (YOLO for object detection, Inception network for face recognition..) and the actual cost functions that were used to train them. That would have helped a lot in getting more practical real life feel helping user community a lot.

The content is great. The best so far in the DL specialization perhaps but I lost a lot of time in the last week's assignment where the grader was prompting wrong output in spite having written a program that gave the correct output. That was very frustrating and the worst part is I still don't understand why that was happening(I got full score after submitting the same program for the 10th time) and even jupyter notebook took ages to load(my net speed is 140MBPS). Apart from this, it was a really nice course and the experience was very satisfying.

Like the others, a fantastic course. Some of the videos and exercises seem a little underprepared, and require more time examining the discussion forums than the first three courses. For example the NST tutorial appears to require using np.square rather than tf.square to obtain expected results. This is not documented, and obtaining other results may result in passing, but it is unclear the ConvNet is working as it should. However the course covers current and quite advanced topics extremely clearly, and includes great links to original papers.

The content is superb, but the realisation of the course seems a bit rushed in comparison to the previous courses in the specialisation. The editing of the videos has many issues (fragments that were meant to be cut out are left in the lectures), there are many typos in the notebooks and the references for documentation are outdated. In one case the grader of the notebook has an unexpected mistake built in (it expects one rectangle area to have a negative value and gives a 0/10 grade when you try to prepare the code for such edge case).

Everything was absolutly great and undoutdetly quite usefull , I have one complaint though , if I had a problem in assignments( mostly related to grader output like U-net segmenation which I passed all the tests successfully but somehow my grade turned out 66) and I 've put my problem in discussion no mentor would answer it and if they did it was always the same "The topic has been moved somewhere that I don't know"

anywho it was great thanks for allowing me to take this course , it really helped me I hope I could return your favor

Very good course. I only wish there were even more examples with harder homework. I did every homework in less than an hour and felt like I still couldn't reproduce much on my own after taking a class. The lectures are great and cover most of the material you need. But as programming assignments go, it is still a very introductory class and you are definitely not ready to write much on your own afterwards. But I think they can improve a lot with a few more homework jupyter notebooks that are more advanced that their current ones.