Philipp Wirth
Philipp Wirth
# DRAFT: Add SMoG Draft implementation of SMoG: https://arxiv.org/pdf/2207.06167.pdf Currently, the CIFAR10 benchmark reaches approximately `65%` accuracy after 200 epochs (depending on the exact set of hyperparameters). There are some...
# Use structured configs to enforce types in command-line tool Using structured configs is a better solution for #575. However, it requires `dataclasses` which is only available from Python 3.7....
# Command line tool not found CLI commands did not work in the AWS environment (after `pip install lightly`). But somehow the CLI commands were not registered. ``` lightly-magic: command...
# Update docs and tutorials after refactoring Depends on #378 and #379. The refactoring will likely introduce breaking changes to our current setup. This means that the documentation and tutorials...
# Write benchmark code for ImageNet Similar to the benchmarks we have for [Cifar10 and Imagenette](https://docs.lightly.ai/getting_started/benchmarks.html), we should write the benchmark code for `ImageNet` with the parameters from the respective...
# Update documentation images To Do: - [ ] Update the images in "The Lightly Platform" - [ ] Update the images in "Active Learning"
# Add examples for all ApiWorkflow functions The `ApiWorkflow` class provides a lot of usability. You can e.g. create datasets, upload samples, upload embeddings, upload custom metadata, delete datasets, and...
# CLI can't process B/W videos Running the `lightly-embed` command on black and white videos leads to this error: ``` RuntimeError: output with shape [1, 64, 64] doesn't match the...
# Add multi-modal loss As discussed in #249, adding the loss function as described in [CLIP](https://arxiv.org/pdf/2103.00020.pdf) would enable users to work with multi-modal datasets. The pseudo-code (from the paper) is:...
# Add default parameters for all projection heads It's helpful to know what the default parameters were in the papers to get started. We should add the default projection head...