# Dependencies to run the ingestion + scoring programs (Codabench worker / # Docker image). The ingestion program runs the bundled benchopt benchmark # inference-only on the submission, so it needs benchopt plus everything the # tracks import at run time (the benchmark and its benchmark_utils ship in # the phase bundle, not in the image). # 1.10+ for `benchopt prepare` and file-path solver/dataset selection (the # ingestion program selects the submission with `-s /submission.py`). benchopt>=1.10 pandas scikit-learn # unpinned since neuralbench 0.3.x (only needs torch>=2.5.1). Portable on # purpose (participants install whatever suits their machine); the docker # image pins the CUDA wheel channel instead (tools/Dockerfile TORCH_CHANNEL). torch # Data stack for the real tasks (Simulated needs none of these). The four # neuroai packages release in lockstep (neuralbench only pins its siblings to # ~=0.4.0), so all four are pinned exactly for a reproducible image. neuralset==0.4.0 neuralfetch==0.4.0 neuraltrain==0.4.0 neuralbench==0.4.0 exca==0.5.29 moabb mne # Used by the EEGNet baselines / many torch submissions: braindecode # NeuralFetch downloads the PROTEUS and Muse releases from NEMAR through it, # importing it on demand (the neuroai packages do not depend on it). nemar-py>=0.3.1