Timeline for answer to Why did my style-transfer model suddenly stop learning after 3700/20000 batches? by Susmit Agrawal
Current License: CC BY-SA 4.0
Post Revisions
3 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Mar 13, 2020 at 0:38 | comment | added | michael | [SOLVED]: it was an all black input image that caused a divide by zero error in the domain scaling code. | |
| Mar 12, 2020 at 7:55 | comment | added | michael |
it actually learns, then it suddenly fails. maybe that's a sign of exploding gradients--I'm still trying to learn the intuition. But for now, I excluded the first 4K batches and it seems to continue to learn just fine. maybe it's a problem with an input tensor with domain out of bounds, or an alpha channel? I'm using tf.image.decode_jpeg() in my preprocessing code
|
|
| Mar 11, 2020 at 8:30 | history | answered | Susmit Agrawal | CC BY-SA 4.0 |