Björn Barz

icon company Carl Zeiss AG, Corporate Research & Technology icon location Jena, Germany icon location Machine Learning Scientist at Carl Zeiss AG, Corporate Research and Technology. Previously: PostDoc at University of Jena (Germany), Computer Vision Group.

Results 4 issues of Björn Barz

Fix restoring trainer after `lr_find()`

## What does this PR do? Fixes #12670: After using the LR finder and then training the model with the same trainer instance (which is the standard use-case), an incorrect...

pl

Return correct output shape for MultiHeadAttention

In contrast to `MultiHeadSelfAttention`, `MultiHeadAttention` has two inputs but only one input. The current implementation does not override `compute_output_shape`, which by default returns the input shapes unmodified. Instead, only the...

What happened to the ignore label?

7 comment

On February 10th, 2023, documentation for how to use data with sparse annotations via an "ignore label" was added: 9181f2cbd93d9b29efb9e634cf6981ad5176c30c On March 16th, 2023, this documentation was removed again: 8014c7be68df3da32eeebecb1d0ea42c9dfa7cab...

Ignore parameters causing ValueError when dumping to YAML

2 comment

## What does this PR do? Fixes #19730 Unserializable parameters such as nn.Modules or Tensors often fail with a `ValueError` when calling `yaml.dump` on them and using PyTorch 2.x. Lightning...

pl