Do pre-trained model weights e.g. ResNet50 get updated?

My team uses the ResNet50 backbone for some of our models. Lately we’ve observed small but consistent performance discrepancies between model versions trained a year ago and those we train now using the same backbone. As far as we can tell, all other elements of our training environment e.g. package versioning, training data, etc. are constant in our A/B tests. This leads me to wonder if the backbone model weights queried through the torch API have changed – hence my question. If they are not static, is there somewhere to view their version control history?

Any insight here would be greatly appreciated!

You can use specific versions and the DEFAULT option that is updated with the best results. Here is some information https://pytorch.org/vision/0.20/models.html