Eldar Kurtić

Bar plot with XData as categorical array

3

Hi, I encountered a problem while I was trying to use `matlab2tikz` to convert bar plot with `XData` as categorical array. Simple `matlab` code for bar plot that illustrates this...

Implement mask-transfer between models at init

2

* masks are transferred from the given model at initialization * to keep masks on, one should add ConstantPruningModifier in the yaml recipe * add apply_weight_mask function to transfer masks...

Add args for black/isort/flake in the Makefile

2

**Is your feature request related to a problem? Please describe.** It could be useful to have an option to specify some arguments to black/isort/flake8 when running `make quality` and `make...

**Is your feature request related to a problem? Please describe.** It would be a good idea to mention here https://github.com/neuralmagic/sparseml/blob/main/README.md#:~:text=Instead%20of%20training,manager.apply(model) that one-shot pruning runs with pruners that use gradients require...

**Describe the question** Why https://github.com/neuralmagic/sparseml/blob/a6477f900b55afd555734fe6cf784e0137d815a5/src/sparseml/pytorch/sparsification/modifier.py#L525 and https://github.com/neuralmagic/sparseml/blob/a6477f900b55afd555734fe6cf784e0137d815a5/src/sparseml/pytorch/sparsification/modifier.py#L548 accept `steps_per_epoch`, given that it's unused param in these methods?

Weights in the `Conv2d` layer are stored as tensors with shape `(out_channels, in_channels, kernel_size[0], kernel_size[1])`, while weights in the `ConvTranspose2d` are stored as tensors with shape `(in_channels, out_channels, kernel_size[0], kernel_size[1])`.

Hi, I am not able to run the code with the specified pytorch version (0.4.0). Could you please post more details on the exact cuda version, torchvision version, or for...

This feature improves OBSPruningModifier by allowing to recompute inverse Hessian `num_recomputations` times when doing a pruning step. For example, in the standard pipeline we would ask OBSPruningModifier to prune a...

Fixes https://github.com/neuralmagic/sparseml/issues/676

Some Mappings missed when checking for Dict type only

1

**Describe the bug** In the case of masked-language-modeling experiment, inputs to the model (outputs from the tokenizer) are instances of the `BatchEncoding` class ([HF source code](https://github.com/huggingface/transformers/blob/198c335d219a5eb4d3f124fdd1ce1a9cd9f78a9b/src/transformers/tokenization_utils_base.py#L167)) which is a subclass...