We got @jay_mangi’s issue sorted out on dm, but for anyone in the future who faces a similar issue - please check Exercises 6.2 & 6.3: Restore the checkpoint very carefully. Restoring the checkpoint properly will give you a huge headstart on the weights. Without that, it would take way too long to get the loss down to a reasonable value.
Make sure you are assigning the right variables in the right places. It can be easy to overlook something, but if you go back and review the examples and carefully check the instructions and the comments above each line of code you need to write, it should help you find it.
Hi @Wendy hope you are well. This assignment is the most complex of all and of course the mistakes are usually because of typos or simple mistakes in the code of the learner. So my advice for any learner having problems with this assignment is to recheck all exercises and if needed to reset the notebook and redo the assignment.
Hi all!
some insights regarding the C3W2 assignment:
After dealing with quite a few issues with the colab interface, the take-home massage was, if something that should work doesn’t (like importing util packages) - restart the colab kernel (Runtime), still doesn’t work? close the colab notebook and reopen? still doesn’t work?? re-log to coursera, move a country and so on…
Exercise 2 - Make sure that you append the training images in sorted order! because of the colab issues described above, I had to comment out the plot_detections function (which calls functions from viz_utils), leading to me not able to see that the gt boxes are not in the same order as the images - this eventually caused wrong labeling and severe underfitting.
Exercise 9 - double check that you fine tune the necessary variable (duh…). I casually used a shorter prefix to search the variables and wound up training 92 (!!) vars, resulting in a really flawed and slow training process (loss > 10^5)