Routes 3D ResNets PyTorch video action-recognition workflows across training, inference, and data preparation.
日本語の概要は準備中です。原文の説明を表示しています。
Use the Folding@home coronavirus repository conventions to curate molecular structures, prepare bounded OpenMM systems, and preserve target, publication, and provenance notes.
インストールする前に、エージェントに与えられる指示の中身を確認できます。
Use this repo-specific operating skill for reusable work derived from the Folding@home coronavirus preparation repository. It covers structure curation, OpenMM preparation/equilibration mechanics, and project-context notes. It does not run Folding@home clients, network acquisition, long production trajectories, or medical/therapeutic interpretation.
For a new simulation input, establish source-structure identity and note provenance first; curate and validate the coordinate topology; choose and document the force-field/ligand parameter path; prepare a small CPU OpenMM system; run only a bounded minimization/equilibration smoke check; inspect energy, atom count, box, and output bundle; then record exact parameters, versions, warnings, and limitations. A later Researcher may expand the run only with an explicit scientific and compute budget.
Do not treat a filename, a successful PDB parse, a short CPU step, or XML serialization as proof of biological identity, force-field suitability, stable long-timestep dynamics, binding, efficacy, or reproducibility across platforms.
The source project documents rapidly prepared inputs and commonly uses equilibrated_4fs.pdb as the Folding@home handoff. Reusable generated helpers use safer explicit arguments and short defaults instead of historical hard-coded paths or multi-nanosecond defaults. A 4 amu hydrogen mass and 4/5 fs continuation are historical workflow conventions that must be disclosed and validated for the actual system, not universal recommendations.
The required backend for this selected scope is CPU OpenMM. CUDA is optional: platform enumeration succeeded in inspection, but context creation was limited by CUDA_ERROR_UNSUPPORTED_PTX_VERSION. Do not claim CUDA verification or require it for a CPU result.
Use repo-provenance.md to anchor claims to the source revision and evidence paths. This generated skill is intentionally left in the repository-local output tree and is not imported into the managed DisCo skill library.
まだレビューはありません。使ってみた感想をお寄せください。
概要と使いどころ
Routes 3D ResNets PyTorch video action-recognition workflows across training, inference, and data preparation.
日本語の概要は準備中です。原文の説明を表示しています。
Guide 3DDFA Python inference, geometry rendering, training/evaluation, and optional C++ ONNX workflows for 3D dense face alignment.
日本語の概要は準備中です。原文の説明を表示しています。
Routes 3DDFA_V2 face-alignment setup, still-image demos, video tracking, and ONNX benchmarking workflows.
日本語の概要は準備中です。原文の説明を表示しています。
Operate AB3DMOT 3D multi-object tracking workflows for KITTI and nuScenes data, tracking, evaluation, and visualization.
日本語の概要は準備中です。原文の説明を表示しています。
Use Hugging Face Accelerate for PyTorch training-loop migration, distributed launch/configuration, DeepSpeed/FSDP/TPU backend setup, big-model inference/offload, checkpointing, tracking, and troubleshooting.
日本語の概要は準備中です。原文の説明を表示しています。
Route Acme reinforcement-learning framework tasks across core loops, replay/data, JAX agents, and TensorFlow/Sonnet agents.
日本語の概要は準備中です。原文の説明を表示しています。