galli-leo
galli-leo
@iwantooxxoox That sounds very cool (cannot seem to find the model though :/). Could you also include the changes you made to the coreml source code? I want to try...
Regarding the `nan` issue. I think it has to do with a lot of values becoming 0. This seems to happen especially after ReLU layers. ~~After digging through the coreml...
After debugging some more, it seems that the very first batch norm layer already has different results than the pytorch version. Not sure if this is relevant? Also the results...
Thanks for the suggestion, that function seems really useful! So if I only use the first layer, the assertion still fails, but the arrays aren't visibly different: ``` Arrays are...
Interesting, so the first nan values "appear" after the 18 layer (a DepthConcat one). They appear in the filters 384 to 479 and all of them in the "first" pixel:...
I finally got it to work! Problem was that I had to run a random tensor through the model in torch, to get it to load in pytorch (see https://github.com/pytorch/pytorch/issues/956)....
I didn't, I converted it into a coreml model.
I have the same issue with the following xml snippet: ``` 1 a ASDF 123 ```
I think this is what would allows this to work very well: https://avaidyam.github.io/2018/02/17/CAPluginLayer_CABackdropLayer.html
@olunx Interested in using this for a project, how did you build the library for android?