This repository is Pablo's development workspace for contributing to torchmil.
It keeps the devcontainer configuration, experiments, useful scripts, documentation, and useful links together.
- #29 - [BUG] MultiheadSelfAttention still uses dropout in eval mode.
- #31 - [BUG] ProbSmoothAttentionPool regularization includes padded instances.
- #33 - [BUG] SmoothTop1SVM mean loss changes when identical samples are repeated.
- #36 - [BUG] DTFDMIL prediction changes when a bag is zero-padded and masked.
- #38 - [FEATURE] Support batches of variable-size bags in TransMIL.
- #40 - [BUG] CAMIL prediction changes when a bag is padded in a batch.
- #28 - Fix MultiheadSelfAttention still using dropout in eval mode: fixes #29.
- #30 - Fix padding in ProbSmoothAttentionPool regularization: fixes #31.
- #32 - Fix SmoothTop1SVM target broadcasting: fixes #33.
- #34 - Fix documentation build command in PR checklist.
- #35 - Fix DTFDMIL predictions changing with masked padding: fixes #36.
- #37 - Support masked batches of variable-size bags in TransMIL: fixes #38.
- #39 - Fix CAMIL predictions changing with masked padding: fixes #40.
Set up the host NVIDIA driver and NVIDIA Container Toolkit so Docker can expose the GPU.
Clone the workspace and run its setup script:
mkdir torchmil
git clone https://github.com/pupeno/torchmil-workspace.git torchmil/workspace
torchmil/workspace/setup.shOpen the outer torchmil/ directory in an editor with devcontainer support, then reopen it in its devcontainer.
Run tests inside the devcontainer:
cd /workspaces/torchmil/torchmil
pytestSync dependencies:
uv sync --project /workspaces/torchmil/workspace --locked