Course 4, Week 2, Assignment 1 (Residual Networks) Model performance

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

So I’ve been playing around with the model in Coursera | Online Courses & Credentials From Top Educators. Join for Free | Coursera

When you complete the assignment, the model performance is accuracy 0.86 and loss 0.55

I thought I’d play around a bit with the model and see how different layers affect the performance… So I started by removing all the layers from Stage 3, 4, 5 (including the last avg pool). I basically reduced the model just to Stage 1 and 2 with one FC layer and suddenly my model performs much better (accuracy 0.933 and loss 0.18

Can anyone comment why that is?

2

If I recall correctly, the model in the assignment was not trained to full convergence, due to the need to limit the amount of processing power required of Coursera’s server.

3

@abaumhof Clear the cache EVERYTIME YOU MAKE THE CHANGES
and then try to run the program.
Check if you are getting the same results.

1 Like
4

I think that means:

  • “Kernel->Restart & Clear Output”
  • Followed by “Cell->Run All”.

(So as to avoid confusion with a clearing the browser cache).

1 Like
5

Yes @TMosh I meant to say that only :blush: Thanks

image
image273×306 7.26 KB

1 Like
6

Thanks a lot for your comments..

I restarted the kernel and the results are essentially the same… I’ll keep digging a bit…

the numbers changed slightly, but still considerably better than with stage 3-5

Loss = 0.27733924984931946
Test Accuracy = 0.9416666626930237