Open maxmatical opened 3 years ago
First, cool library ... wasn't aware of that one!
Second, it would be helpful you can post a gist I can run so I can see full stack trace here. The inp
is expecting a tensor, but in the case of Blurr (and HuggingFace) what is output at this point is an object with a bunch of info such as loss, etc.... With callbacks, I'm sure this can be altered to work with Blurr/HF.
Btw, any particular reason you're using BatchLoss
? Just curious :)
Here is a gist using the BatchLossFilter
callback with the standard blurr training script. I'm currently experimenting with BatchLossFilter
since it had some traction on twitter a while back, plus intuitively, focusing on the harder examples could potentially improve performance, so extra tools in the toolbox never hurts 😄
i was able using the example implementation with other forms of data (images, timeseries etc.), so it looks to be an issue specific to huggingface if that helps
I've been trying to experiment with using tsai's BatchLossFilter callback. If I try to run the training with this callback
I get the following error, which is due to the
SequenceClassifierOutput
object from huggingfaceIs there any way I can adapt
BatchLossFilter
to be functional with blurr? I haven't had any issues using other callbacks with blurr