great!

xiexie

坚实的基础

Good!

Show!

nice！

good

good

good

cool

Nice

Nice

cool

good

best

good

good

good

ty

:)

gj

​

Very good way to get familiar with Tensoflow - it's pluses as well as its minuses.

Good overview of applying tf.keras to this topic. Machine learning is clearly a practical discipline (i.e. theory alone will not get you there), so I appreciated the chance to write some code and read a decent amount of code.

Laurence Moroney is a good, upbeat instructor.

All the courses within the Tensorflow in Practice specialization on Coursera may be most beneficial after first taking Andrew Ng's course on AI (also Coursera), but if you know something about loss functions, gradient descent, and backpropagation (which can be learned quick-and-dirty online), then you should be fine to go ahead and take this specialization before Professor Ng's course.

My one persistent wish for all four of the courses in this specialization is that significantly more time be spent on understanding the shapes of tensors as they flow through the models. Invariably, the only areas that gave me real problems as I did the coding homework were those where my tensor shape did not match what the model needed to see. Documentation at Tensorflow.org was of little help with this topic. Looking at Stackoverflow, it is apparent that there are certain (unwritten?) facts about the order and count of dimensions for the tensors as they flow through, e.g. batch count is listed first, time step is second, frame is third, or something like that. What if I have twelve dimensions in my tensor? Do certain model layers require a minimum number of dimensions of input or output? etc. etc.

Finally, this specialization really teaches the tf.keras framework, not Tensorflow itself, which I do not think was explained in the course info, but maybe I missed it. Still - keras is probably a good way to enter the subject.

All in all, I do know a lot more than I did before, and have acquired new skills. Clearly, there's more to work on, which is good.

I like very match the first and second week of the course, because it contains dense new theoretical and practical things. The idea of time series forecasting and preparing windowed dataset was explained very clear and was very usefull for all next lessons. Also the difference between statistic and neural network approaches was very helpful.

The 3 and 4 week I would prefer zip in one , because the experiments with RNN, LSTM and Conv is very familiar and actually I've done them together one by one. I would pleased to learn some explanation and examples why each type of architecture follow their result. How the results depend on dataset preparation. Particulary, I did not get what architecture work better with seasonality, autocorrelations, and noise.

I think overall it is a good course, these are the things I learnt:

First-hand experience with tensorflow, but more focus on the basics of keras

Knows how to preprocess data for image, text, and times series to feed it into NN

Knows basic concepts of keras layers such as CNN, LSTM, RNN, Conv1D, DNN

Knows learning rate rough gauge techniques

Things to improve:

Fix the typos, such as window[:1], there are a few posted in the forum

Should introduce more basics of tensorflow instead of kerasShould

include more links/documentation for the side knowledge, such as paddingAdding

some layers seems magical, such as Conv1D before LSTM for time series, what is the logic behind?