Hi! I’d say a data vizualisation library that can hook into Visdom (& tensorboardX I guess…). Visdom is great and all in that it gives you the freedom to plot anything but I believe that features on top of it would be great for convenience. It could help with gradient vizualisation, model explainability, model architecture, deeper look into losses which consists of different summations etc.
Another one that I feel is missing is some sort of metric library. Precision, recall, AP & mAP for bounding boxes, segmentation metrics and god knows what.
Edit: I’m not entirely sure about what content exists other frameworks such as ignite & fastai. Do they already cover this?
Yes, I agree on easily integrated model explainability will be interesting.
@Oli training viz is something can be more or less simply inserted into ignite, there are already provided tensorboard, visdom etc loggers and we can plot out-of-box losses, scalar/vector metrics, model weights norms, grads etc: ignite.contrib.handlers — PyTorch-Ignite v0.4.13 Documentation
Metric library for all cases (as you site detection tasks) will be awesome. Today in ignite, we have a lot of basic metrics for classification, segmentation tasks and a bunch of regression metrics. Detection mAP, AP will be a of next steps…
Edit: I’m not entirely sure about what content exists other frameworks such as ignite & fastai. Do they already cover this?