alexander-g

Module Slicing

17

Experimental support for slicing modules. e.g: ``` x = jnp.zeros((2, 224, 224, 3)) resnet = elegy.nets.resnet.ResNet18() submodule = elegy.module_slicing.slice_module_from_to( resnet, start_module=None, end_module=["/res_net_block_1", "/res_net_block_3", "/res_net_block_5", "/res_net_block_7" ], sample_input=x, ) outputs =...

Module Slicing 2

10

Alternative approach to #115 - This is now based on `jax.make_jaxpr` and `jax.named_call`, not on summaries. Specifically, I've extended hooks context with a new attribute, if it is active, `Module.call`...

[WIP] U-Net

2

- [x] U-Net architecture for image segmentation - [x] `MeanIoU` metric - [x] `class_weight` and `ignore_index` parameters for `SparseCategoricalCrossentropy` - [x] Backboned U-Net - [x] Training examples - [x] Pretrained...

[Bug] Size of parameters.h5 too large

1

The `parameters.h5` file saved by `Model.save` is way larger than the model parameters. The factor is not fixed and varies between 3 to 5 depending on the model. ```python model...

Verbosity options 1 and 3 slow down training

2

The print callbacks for live monitoring significantly slow down training. For example my GPU utilization drops from 99% to

Reproducing the OOD results

35

Hi, I'm interested in how you have achieved the 2.5% OOD false positive rate on CUB from your paper. I've run your code but getting different results, here is the...