AGENT SKILL

classification

Guides Attention-Gated Networks ultrasound classification workflows, including Sononet training, evaluation, HDF5 data layout, and attention overlays.

License 许可
MIT
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
classification
license
MIT
metadata
{ "disco-role": "operating" }
description
Guides Attention-Gated Networks ultrasound classification workflows, including Sononet training, evaluation, HDF5 data layout, and attention overlays.
disable-model-invocation
true

02 资源文件 6

03 SKILL.md 正文

Classification

Use this sub-skill for ultrasound scan-plane classification tasks in Attention-Gated Networks. It covers Sononet, Sononet2, Sononet Grid Attention, FeedForwardClassifier, AggregatedClassifier, the HDF5 ultrasound loader, classification training/testing scripts, and attention overlay export.

Read first

  • Read api-reference.md when you need exact model names, wrapper behavior, config fields, aggregation modes, or class and function signatures.
  • Read data-layout.md before preparing the ultrasound HDF5 file or debugging loader/sampler failures.
  • Read workflows.md for training, testing, checkpoint evaluation, model inspection, and safe attention-overlay export.
  • Read troubleshooting.md when CUDA, torchsample, Visdom, sampler, HDF5, or overlay failures appear.
  • Run run_classifier.py for skill-owned training and testing commands that replace the source repo's root entry points.
  • Run export_attention_overlay.py to replace the source repo's hard-coded attention visualization script with a parameterized, skill-owned helper.
  • For fast environment validation, run the root helper ../../scripts/check_env.py --repo-root /path/to/Attention-Gated-Networks --mode classification.

When to use this sub-skill

Use this route when the user asks to:

  • train or test ultrasound scan-plane classification models;
  • choose between sononet, sononet2, and sononet_grid_attention;
  • configure aggregated_classifier, deep supervision, or attention aggregation;
  • inspect classifier outputs, class counts, losses, or scikit-learn metrics;
  • prepare or validate the us HDF5 dataset layout;
  • export compatibility-score attention maps or overlay PNGs;
  • debug classification-specific CUDA, Visdom, sampler, HDF5, or checkpoint issues.

Route 3D segmentation, NIfTI validation, and multi-attention U-Net feature-map exports to ../segmentation/SKILL.md instead.

Quick workflow

  1. Install the repository with CUDA-enabled PyTorch and the legacy torchsample dependency. Then run:

    python ../../scripts/check_env.py --repo-root /path/to/Attention-Gated-Networks --mode classification
    
  2. Prepare an ultrasound HDF5 file with x_train, p_train, x_val, p_val, x_test, p_test, and label_names. See data-layout.md.

  3. Copy a Sononet config and replace its dataset path. The source configs use model.output_nc=14, input_nc=1, tensor_dim='2D', and gpu_ids=[0].

  4. Train with the bundled replacement when real data is available:

Relative config paths are resolved from the explicit --repo-root; relative data_path values are resolved from the config file's parent. The HDF5 data file and any checkpoint/weights must already exist outside this skill. Bundled runners and check_env.py --config ... reject missing or private /vol/... paths before constructing a dataset.

python scripts/run_classifier.py --repo-root /path/to/Attention-Gated-Networks --config path/to/config.json --mode train
  1. Evaluate a checkpoint with the bundled replacement:

    python scripts/run_classifier.py --repo-root /path/to/Attention-Gated-Networks --config path/to/config.json --mode test
    
  2. For attention models, use model_type='sononet_grid_attention' and read workflows.md before choosing aggregation_mode and aggregation.

  3. Export a safe synthetic overlay, or adapt the command for a real 2D input:

    python scripts/export_attention_overlay.py \
      --repo-root /path/to/Attention-Gated-Networks \
      --config configs/config_sononet_grid_att_8.json \
      --output-dir /tmp/ag-net-attention
    

Key decisions

  • CUDA is required for the unmodified wrappers because the model and tensor paths call .cuda().
  • The stock ultrasound configs assume 14 classes. Do not reuse them unchanged for a different HDF5 label set.
  • AggregatedClassifier is for multiple-output attention models. Plain Sononet classifiers use FeedForwardClassifier.
  • Synthetic smoke checks prove wiring only; they do not replace training on the real ultrasound HDF5 file or evaluating a real checkpoint.
  • The original attention visualization script hard-codes private paths. Use the bundled helper or parameterize your own script rather than running it as-is.

The source train_classifaction.py contains a 10-hour sleep after the final epoch's updates. It is retained in the source checkout for provenance; use the bundled runner when a bounded run without that hold is required.

Expected outputs

Training writes checkpoints and logs below the configured checkpoint directory and experiment name. Testing writes accumulated metrics and a test_result.pkl inside the model save directory. The bundled attention helper writes overlay PNGs and .npy arrays to the requested output directory.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-triton-opt

Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-solve

END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-skill

PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-pattern-fusion

Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-orchestrate

Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

passnet-feedback

Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).

★ 351