In your opinion, what's a feature that Keras is missing, or otherwise something that Keras does poorly at this time? Be extremely specific (precise workflow, API signature). Thanks!
Show this thread
-
Checkpoint hacking: We had a trouble trying to do the inflation trick (converting 2d to 3d) and that ended up being the main factor pushing the project to Pytorch. I think internally at Google there is something called graphter?
-
-
It's hard to modify the shapes and values of the weights in a checkpoint, then load it. It is common for 3D conv networks to hack an existing 2D checkpoint in order to initialize. It seems like was a lot easier to do it for non-keras models. When I was at Google I used graphter.
-
+1 Even when using the tf1 trick, it's not always possible to change the checkpoint just by changing the names, especially when the source ckpt is in .h5. I ended up doing a lot of hacking including using tf1. This still keeps quite some researcher from moving to tf2 keras.
- Show replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.