paganpasta / eqxvision Goto Github PK

Super cool to see this project get off the ground!

I just thought I'd give you a heads-up if you're using my doc-building stack, that the mkdocs command needs running twice in order to build the documentation correctly. See for example https://github.com/patrick-kidger/diffrax/blob/77bd28b03aafdebcd985d900118ffb900be66d65/.github/workflows/build_docs.yml#L33

The TL;DR is that the first call is used to populate a cache of "all documented objects", which is used by the second pass to display their names in the documentation in the desired way.

Whether this is possible with ReadTheDocs I don't know.

FWIW I know the underlying mkdocstrings/pytkdocs stack is being reworked quite substantially, so another approach that may be possible is to rewrite my custom hacks tweaks to fit the new world order. (I've not looked into this at all -- right now I'm just using pinned versions of mkdocstrings+pytkdocs and was planning to figure this out only if something eventually breaks.)

This is to mirror the meta_data which torchvision is using for providing additional info on the pre-trained weights.

Missing models listed in no particular order.

• VGGs
• Mobilenet-v2/3
• Swin
• Efficientnet-v1/2
• SqueezeNet
• ConvNext
• DenseNet

Is there an example how to switch to inference mode for evaluation?

An initial idea for the design is:

1. Download weights from pre-hosted torchvision or timms URLs
2. Have common converters which maps layers between the two frameworks. eg: torch.nn.Linear to eqx.nn.Linear
3. Lastly, assembly of the weights can be performed by individual architectures.

Open to discussion!

Currently these are only used to load weights. Is it possible to move the model loading bit into the package so that we dont have to install torch and torchvision. ?

Hi everyone!

I installed the latest EqxVision and got an import error when importing equinox.experimental. This seems to be because Equinox graduated BatchNorm to stable on v0.10.3.

I'll create a short PR for this.

hi there, since a recent change in eqx and split of functionality to google/jaxtyping, the import of PyTree and others from equinox.custom_types breaks and causes error. i can confirm i resolved locallish by running sed on all the files that import from that, changing to jaxtyping, and resolved the import errors, e.g.
sed -i 's/equinox.custom_types/jaxtyping/' .local/lib/python3.7/site-packages/eqxvision/models/segmentation/_utils.py

Currently, Imagenet pre-trained VGGs show a large gap in top-1 accuracy. The difference could be due to rigidity of the classifier layer. To see if the generated features are comparable, the experiment will be to finetune and compare Vgg-11 say on a small dataset like STL-10.

Currently, gradient computation is not supported with BatchNorm models. With the next release of Equinox, it would be nice to have test cases to check that networks are passing gradient computation.

Hey @paganpasta! I often point people at Eqxvision, as it's a great model zoo. (And one that you've clearly put a lot of work into.)

I was curious if this project was still maintained? I notice things don't seem to have been updated to eqx.nn.BatchNorm (which has been stable for some months now).

If it is maintained, I'd like to see if we could promote this project a little more. But if it's not, then no worries -- I understand maintenance is a burden, and there's libraries I've previously decided to stop supporting as well.

Adding a sample training notebook (preferably on each network) on Imagenette demonstrating some different techniques. For example, different optimizers, transfer learning, distillation etc.
Open to discussions!

• Image Classification
• Transfer Learning
• Generating Adversarial Examples
• Visualising Attentions (Vision Transformer)
• Class activation maps
• Training a GAN
• Training a Diffusion Model