Yann Dubois
Yann Dubois
Hi, To be really sklearn-ish I think it would be very nice to have a class `NeuralNetTransformer` which inherits from `TransformerMixin` and implements a `.transform` method. **Usecases: representation learning.** Example:...
Minor Fixes: - typo in `nift_loader`, so wasn't working. 'x.endsiwth' replaced by 'x.endswith'. - small number of changes so that `.fit_load()` works also with a `torch.utils.data.Dataset` and not only with...
Max save is off by 1, which is annoying because usually you want to store only the best model and the current code would store nothing. Note that you could...
Hi as the title suggests, The function `.fit_load` can only work with a `TensorDataset` and not with a torch `Dataset`. You only have to make 2-3 changes to make it...
Hi, Thanks for the package, it's great. But I think there are dependencies like `pandas` and `nibabel` that are very specific that there shouldn't be required to use the package....
If I understand correctly when performing linear probing you take the representations before the linear projection heads. In the code, this can be done very nicely thanks to this line:...
I'm adding the [CLIP self-supervised compressor](https://arxiv.org/abs/2106.10800) from FAIR + Vector Institute. This allows compressing any dataset orders of magnitude better than with JPEG (e.g. 1000x for ImageNet) without decreasing downstream...
## Bug I see that you are testing whether JIT of the EntropyBottleneck works (and it does), but JITing the compress/decompress method of an entropy bottleneck does not work. which...
Hi, Thanks for the library! I'm new to the JAX+LLM ecosystem and trying to understand which library I should be using. I see a lot of (very impressive) computational efficiency...