Timeline for MobileNets object keypoints localization with Keras
Current License: CC BY-SA 3.0
21 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Sep 14, 2019 at 1:02 | history | bumped | CommunityBot | This question has answers that may be good or bad; the system has marked it active so that they can be reviewed. | |
| May 13, 2019 at 9:02 | history | bumped | CommunityBot | This question has answers that may be good or bad; the system has marked it active so that they can be reviewed. | |
| Jan 12, 2019 at 19:00 | history | bumped | CommunityBot | This question has answers that may be good or bad; the system has marked it active so that they can be reviewed. | |
| Mar 26, 2018 at 23:08 | comment | added | Michael Ramos | those results are with [0,1] & sigmoid. I think now might be a good time to revert and try [-1,1] & relu | |
| Mar 26, 2018 at 23:07 | comment | added | Michael Ramos | @AlexR. it seems I have been wasting other's times... the data had an error in the normalization. After 200 epochs I hit a val_loss of 0.0012, with some overfitting, however the model predictions are much better and closer to where I need them. | |
| Mar 26, 2018 at 22:30 | comment | added | Alex R. | Also the fact that all 0s are still close to your true keypoints is strange. Is your data correctly normalized? | |
| Mar 26, 2018 at 21:48 | comment | added | Alex R. | So to be clear, you shifted your keypoints to [0,1], added the sigmoid activation, and now all your ouputs are 0? What happens if you run model.predict() on your data before starting to train? Is it all saturated at 0? | |
| Mar 26, 2018 at 21:47 | comment | added | Michael Ramos | the outputs are [0. 0. 0. 0. 0. 0. 0. 0.]] | |
| Mar 26, 2018 at 21:45 | comment | added | Michael Ramos | the test seems off... the loss is a bit weird as well, it starts around 7.xxxxe-06 and grows, after 10 epochs it lowers to 1.xxxxe-05 | |
| Mar 26, 2018 at 21:41 | comment | added | Michael Ramos | @AlexR. im seeing a dramatic improvement in loss with your suggestion... Im waiting for a few more epochs and then Ill test | |
| Mar 26, 2018 at 21:32 | comment | added | Alex R. | The default activation for "Dense" is "none". Whereas your predictions should really occur on [-1,1], relative to the scale of the image. Have you tried a sigmoid activation instead (shifting your keypoints to [0,1])? | |
| Mar 26, 2018 at 21:31 | comment | added | Michael Ramos | @MatiasValdenegro to get {-1,1} I do (y - 112) / 112 | |
| Mar 26, 2018 at 21:22 | history | edited | Michael Ramos | CC BY-SA 3.0 |
loss and normalization
|
| Mar 26, 2018 at 21:21 | comment | added | Michael Ramos | @MatiasValdenegro yes | |
| Mar 26, 2018 at 21:20 | comment | added | Dr. Snoopy | Also you didn't mention the loss, so I assume its mean squared error? | |
| Mar 26, 2018 at 21:15 | comment | added | Dr. Snoopy | Then you should use an output activation that matches that range, say tanh. | |
| Mar 26, 2018 at 21:14 | comment | added | Michael Ramos | @MatiasValdenegro {-1,1} normalization | |
| Mar 26, 2018 at 21:10 | comment | added | Dr. Snoopy | Your model does not have an activation at the output, how did you normalize the target coordinates? | |
| Mar 26, 2018 at 20:47 | answer | added | Jakub Bartczuk | timeline score: 2 | |
| Mar 26, 2018 at 19:10 | review | First posts | |||
| Mar 26, 2018 at 19:13 | |||||
| Mar 26, 2018 at 19:09 | history | asked | Michael Ramos | CC BY-SA 3.0 |