Data Anonymous

Description

This research package accompanies the manuscript “Dynamic Linear Dimensionality Reduction-Based Optimizer Switching for NeuralNetworks in Robotic Sorting.” The package contains the data, source code, selected model checkpoints, frozen experimental configurations, and machine-readable results used in the five experiments reported in the manuscript. The classification dataset comprises 1,600 training images, 400 validation images, and 800 independent test images covering four simulated object classes: milk, bread, cereal, and can. FourHDF5 datasets provide simulator-state inputs and Cartesian waypoint targets for class-specific grasp prediction. The packagealso includes one selected image-classification checkpoint and four selected waypoint-prediction checkpoints. The engineering workflow is organized as image-classification training with SGD followed by trajectory-PCA-based P-BFGS continuation, and waypoint-prediction training with AdamW followed by P-BFGS continuation. Final validation jointly evaluates the classifier and the four waypoint regressors. Manuscript-specific comparison logic is separated into five experiment modules.These modules include the P-SGD baseline, multi-seed comparisons, pure-AdamW baselines, frozen P-BFGS update budgets,convergence-confirmation rules, component ablations, and the comparison between class-specific waypoint prediction and a shared geometric prior. The archive provides JSON, JSONL, CSV, NPZ, HDF5, PNG, and PyTorch checkpoint files, together with a SHA-256 file manifest and a reproducibility guide. Third-party software, including robosuite, MuJoCo, and PyTorch, is not redistributed and must be installed separately. Intermediate checkpoints generated during hyperparameter searches are excluded because of their size; therefore, the package supports result auditing, selected-model evaluation, and reconstruction of the reported tables and figures, while a complete replay of all configuration searches requires additional computation. The execution experiment uses simulator-state inputs and an assisted post-grasp backend. Its terminal-completion results should not be interpreted as autonomous physical grasp success.

Steps to reproduce

1. Download and extract the package without changing its directory structure. File sizes and SHA-256 hashes are listed in FILE_MANIFEST_SHA256.csv. 2. Create the environment: conda env create -f environment.yml conda activate dldr-pbfgs Install a platform-compatible PyTorch build if required. Simulator experiments additionally require robosuite, MuJoCo, and their original assets. 3. Enter the code directory and verify the frozen configurations: cd code python -m experiments.verify_config The expected message is: “Frozen configuration and convergence checks passed.” 4. Validate the supplied classifier and four waypoint models: python -m validation.run_pipeline --classifier-checkpoint ../models/classifier_pbfgs_selected.pt --waypoint-checkpoint bread=../models/bread_waypoint_pbfgs.pt --waypoint-checkpoint milk=../models/milk_waypoint_pbfgs.pt --waypoint-checkpoint can=../models/can_waypoint_pbfgs.pt --waypoint-checkpoint cereal=../models/cereal_waypoint_pbfgs.pt --output ../selected_model_validation.json--device cpu The released classifier obtains 100% accuracy and approximately 0.00821 cross-entropy on the 400-image validation set. Small numerical differences may occur across platforms. 5. Inspect or execute the engineering workflow: python run_pipeline.py --dry-run python run_pipeline.py --device cuda The workflow is image SGD -> PCA -> P-BFGS -> waypoint AdamW -> PCA -> P-BFGS -> joint validation. 6. Run the manuscript experiments separately: python -m experiments.experiment_1 --stage all --device cuda --subgoal-device cuda python -m experiments.experiment_2 --seeds 1 --device cuda python -m experiments.experiment_3 --tasks bread milk can cereal --mode both --device cuda python -m experiments.experiment_4 --help python -m experiments.experiment_5 --dry-run Use --dry-run before Experiments 2 and 3 to inspect their complete commands without training. Experiment-specific convergence rules and frozen budgets are defined only under code/experiments/. 7. Reported numerical results can be audited directly from the JSON and CSV files under results/experiment_1 to results/experiment_5. Full hyperparameter-search replay requires additional computation because intermediate search checkpoints were excluded. Experiment 1 uses simulator-state inputs and an assisted post-grasp backend; its completion rate is not autonomous physical grasp success.

Institutions

• Hunan University

  Hunan, Changsha

Categories

Licence