Timeline for Why is my training loss not changing?
Current License: CC BY-SA 4.0
6 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Dec 13, 2021 at 19:24 | vote | accept | Ad Ve | ||
| Dec 13, 2021 at 19:24 | comment | added | Ad Ve | Got it! Turns out, the normalization wasn't the issue. I didn't pass the sample weights while training, which meant the class labels were very unbalanced. Passing them solved the issue. Since it was related to your answer, I'll accept it. Thanks for the help! | |
| Dec 13, 2021 at 18:21 | comment | added | Enes Kuz | There are different kind of normalizations. If you are certain this is not the problem then i would suggest try to recreate the problem with a more compact code. The code you shared is hard to understand and not ready to run. | |
| Dec 13, 2021 at 13:55 | comment | added | Ad Ve | I have normalized the data such that each pixel value is in the range [0,1]. I'm not sure if I testing the model on the data mentioned in the original paper will be possible? The base model used in the architecture is VGG16, while I'm trying it with the Dilated ReLayNet model as base. Funnily enough, the Dilated ReLayNet doesn't have this issue, even though the same dataset is being used. What could be the problem then? | |
| S Dec 12, 2021 at 21:17 | review | First answers | |||
| Dec 12, 2021 at 23:22 | |||||
| S Dec 12, 2021 at 21:17 | history | answered | Enes Kuz | CC BY-SA 4.0 |