@researai/deepscientist 1.5.17 → 1.6.0

package/AISB/catalog/aisb.t3.072_fedgmt.yaml

@@ -0,0 +1,181 @@
+id: aisb.t3.072_fedgmt
+name: 'One Arrow, Two Hawks: Sharpness-aware Minimization for Federated Learning via
+  Global Model Trajectory'
+version: 0.1.0
+one_line: Two-stage federated learning benchmark combining trajectory-aware sharpness
+  control via EMA gradients and late-stage SWA stabilization for improved convergence
+  under non-IID data heterogeneity.
+task_description: 'This packaged benchmark covers federated optimization with trajectory-aware
+  sharpness control and late-stage SWA stabilization. The task is to improve federated
+  generalization and convergence under non-IID data distributions while maintaining
+  client optimization stability. FedGMT combines an EMA-based trajectory tracking
+  mechanism with Stochastic Weight Averaging in the final rounds to smooth the loss
+  landscape. The implementation extends PFLlib and supports cross-device FL scenarios
+  with configurable numbers of heterogeneous clients, participation ratios, and data
+  partitioning strategies.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- test_accuracy
+- federated_checkpoint
+- round_metrics
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: Run one packaged federated-learning evaluation route on CIFAR-10 with
+    default hyperparameters and reduced rounds for smoke testing.
+- id: full_federated_run
+  label: Full Federated Run
+  description: Run the full packaged FedGMT federated optimization workflow across
+    500 rounds with full evaluation suite.
+dataset_download:
+  primary_method: bundled
+  sources:
+  - type: cifar10
+    partition: dirichlet
+    params:
+      concentration:
+      - 1.0
+      - 0.1
+      - 0.01
+      num_clients: 100
+      join_ratio: 0.1
+  - type: cifar100
+    partition: dirichlet
+  - type: cinic10
+    partition: dirichlet
+  notes:
+  - Dataset partitions generated via dataset/generate_dataset.py
+  - Dirichlet non-IID partitioning implemented in dataset/utils/
+  - 10% active clients per round across 100 total clients in standard config
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 12
+    notes: CNN model; ~2.72s per round on RTX 4090; ~0.93GB VRAM per round
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: ResNet18 model; ~20s per round on RTX 4090; ~1.86GB VRAM per round
+environment:
+  python: '3.7'
+  cuda: '11.8'
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - torch
+  - torchvision
+  - tqdm
+  - numpy
+  notes:
+  - Install via pip install -r requirements.txt
+  - Python 3.7 tested; see bundled requirements for full dependency set
+risk_flags:
+- extended_training_time
+- multi_round_convergence
+- non_iid_data_heterogeneity
+risk_notes:
+- Full benchmark requires 500 rounds; estimated 4.5h (CNN) to 33h (ResNet18) on single
+  GPU
+- SWA stabilization begins at round 400; late-stage convergence may vary with data
+  heterogeneity
+- Client drift under extreme non-IID conditions (Dir-0.01) requires careful hyperparameter
+  tuning
+recommended_when: 'Use this benchmark when evaluating federated learning algorithms
+  under realistic data heterogeneity, studying round-level stability and client drift
+  tradeoffs, or benchmarking sharpness-aware minimization techniques in distributed
+  settings with privacy-preserving constraints.
+
+  '
+not_recommended_when: 'Do not use this if you need single-node non-federated training,
+  require personalized per-client models, or need rapid prototyping without multi-round
+  communication overhead.
+
+  '
+paper:
+  title: 'One Arrow, Two Hawks: Sharpness-aware Minimization for Federated Learning
+    via Global Model Trajectory'
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/li25bd.html
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.072_fedgmt.zip
+  archive_type: zip
+  local_dir_name: paper-72-FedGMT
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.072_fedgmt.zip
+  sha256: 469c3e18a45ac8e776a73ab672c969a944e795cf11e24c6385c37bc755a7fdcb
+  size_bytes: 84064
+display:
+  palette_seed: teal-crimson-federated
+  art_style: distributed-dashboard
+  accent_priority: high
+image_path: ../image/072_aisb.t3.072_fedgmt.jpg
+capability_tags:
+- research_code_optimization
+- federated_learning
+- sharpness_aware_minimization
+- distributed_training
+- image_classification
+- non_iid_learning
+- swa_stabilization
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+hyperparameters:
+  common:
+    optimizer: SGD
+    learning_rate: 0.01
+    momentum: 0.9
+    weight_decay: 1.0e-05
+    batch_size: 50
+    local_epochs: 5
+    total_rounds: 500
+  fedgmt_specific:
+    ema_coefficient_alpha:
+    - 0.95
+    - 0.995
+    - 0.998
+    sharpness_strength_gamma:
+    - 0.5
+    - 1.0
+    - 2.0
+    penalty_coefficient: 10
+    swa_start_round: 400
+    swa_window: 100
+algorithm_family:
+- FedAvg
+- FedDyn
+- FedSAM
+- FedSpeed
+- FedSMOO
+- FedLESAM
+- FedGMT
+code_anchors:
+  main_entry: system/main.py
+  server_base: system/flcore/servers/serverbase.py
+  fedgmt_server: system/flcore/servers/servergmt.py
+  data_partition: dataset/generate_dataset.py

package/AISB/catalog/aisb.t3.072_fedgmt.zh.yaml

@@ -0,0 +1,158 @@
+id: aisb.t3.072_fedgmt
+name: '一箭双雕：基于全局模型轨迹的联邦学习锐度感知最小化'
+version: 0.1.0
+one_line: 两阶段联邦学习基准，结合EMA梯度的轨迹感知锐度控制与后期SWA稳定化，在非独立同分布数据异质性条件下改善收敛性能。
+task_description: |
+  本基准套件涵盖基于轨迹感知锐度控制和后期SWA稳定化的联邦优化任务。任务目标是在非独立同分布数据分布下提升联邦泛化能力和收敛性，同时保持客户端优化稳定性。FedGMT结合基于EMA的轨迹跟踪机制与最终轮次的随机权重平均，以平滑损失曲面。实现基于PFLlib扩展，支持跨设备联邦学习场景，可配置异质性客户端数量、参与比例和数据分区策略。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- test_accuracy
+- federated_checkpoint
+- round_metrics
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: 使用默认超参数和减少轮次在CIFAR-10上运行一条封装的联邦学习评估路径，用于冒烟测试。
+- id: full_federated_run
+  label: 完整联邦运行
+  description: 跨500轮运行完整的封装FedGMT联邦优化工作流，包含完整评估套件。
+dataset_download:
+  primary_method: bundled
+  sources:
+  - type: cifar10
+    partition: dirichlet
+    params:
+      concentration:
+      - 1.0
+      - 0.1
+      - 0.01
+      num_clients: 100
+      join_ratio: 0.1
+  - type: cifar100
+    partition: dirichlet
+  - type: cinic10
+    partition: dirichlet
+  notes:
+  - 数据集分区通过 dataset/generate_dataset.py 生成
+  - Dirichlet非独立同分布分区实现于 dataset/utils/
+  - 标准配置中每轮10%的活跃客户端，共100个客户端
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 1
+    gpu_vram_gb: 12
+    notes: CNN模型；在RTX 4090上每轮约2.72秒；每轮约0.93GB显存
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 100
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: ResNet18模型；在RTX 4090上每轮约20秒；每轮约1.86GB显存
+environment:
+  python: '3.7'
+  cuda: '11.8'
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - torch
+  - torchvision
+  - tqdm
+  - numpy
+  notes:
+  - 通过 pip install -r requirements.txt 安装
+  - 已测试Python 3.7；完整依赖集见捆绑的requirements文件
+risk_flags:
+- extended_training_time
+- multi_round_convergence
+- non_iid_data_heterogeneity
+risk_notes:
+- 完整基准需要500轮；在单GPU上估计耗时CNN约4.5小时，ResNet18约33小时
+- SWA稳定化从第400轮开始；后期收敛可能因数据异质性而有所差异
+- 在极端非独立同分布条件下（Dir-0.01）的客户端漂移需要仔细调整超参数
+recommended_when: |
+  在评估真实数据异质性下的联邦学习算法、研究轮级稳定性和客户端漂移权衡，或在具有隐私保护约束的分布式设置中对锐度感知最小化技术进行基准测试时使用本基准。
+not_recommended_when: |
+  如需单节点非联邦训练、需要个性化每客户端模型，或需要在多轮通信开销下进行快速原型开发时，不建议使用本基准。
+paper:
+  title: 'One Arrow, Two Hawks: Sharpness-aware Minimization for Federated Learning
+    via Global Model Trajectory'
+  venue: ICML 2025
+  year: 2025
+  url: https://proceedings.mlr.press/v267/li25bd.html
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.072_fedgmt.zip
+  archive_type: zip
+  local_dir_name: paper-72-FedGMT
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.072_fedgmt.zip
+  sha256: 469c3e18a45ac8e776a73ab672c969a944e795cf11e24c6385c37bc755a7fdcb
+  size_bytes: 84064
+display:
+  palette_seed: teal-crimson-federated
+  art_style: distributed-dashboard
+  accent_priority: high
+image_path: ../image/072_aisb.t3.072_fedgmt.jpg
+capability_tags:
+- research_code_optimization
+- federated_learning
+- sharpness_aware_minimization
+- distributed_training
+- image_classification
+- non_iid_learning
+- swa_stabilization
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+hyperparameters:
+  common:
+    optimizer: SGD
+    learning_rate: 0.01
+    momentum: 0.9
+    weight_decay: 1.0e-05
+    batch_size: 50
+    local_epochs: 5
+    total_rounds: 500
+  fedgmt_specific:
+    ema_coefficient_alpha:
+    - 0.95
+    - 0.995
+    - 0.998
+    sharpness_strength_gamma:
+    - 0.5
+    - 1.0
+    - 2.0
+    penalty_coefficient: 10
+    swa_start_round: 400
+    swa_window: 100
+algorithm_family:
+- FedAvg
+- FedDyn
+- FedSAM
+- FedSpeed
+- FedSMOO
+- FedLESAM
+- FedGMT
+code_anchors:
+  main_entry: system/main.py
+  server_base: system/flcore/servers/serverbase.py
+  fedgmt_server: system/flcore/servers/servergmt.py
+  data_partition: dataset/generate_dataset.py

package/AISB/catalog/aisb.t3.073_rld.yaml

@@ -0,0 +1,161 @@
+id: aisb.t3.073_rld
+name: Regularized Langevin Dynamics for Combinatorial Optimization
+version: 0.1.0
+one_line: Optimize combinatorial optimization via regularized Langevin dynamics with
+  Hamming-distance regularization for simulated annealing and neural solvers.
+task_description: 'This packaged benchmark covers combinatorial optimization via Regularized
+  Langevin Dynamics (RLD), which enforces an expected Hamming distance between sampled
+  and current solutions to escape local optima in discrete domains. The benchmark
+  includes two solvers: RLSA (Regularized Langevin Simulated Annealing) based on gradient-guided
+  MCMC sampling, and RLNN (Regularized Langevin Neural Network) supporting both differentiable
+  (erdoes) and fully black-box (reinforce) training losses. Problems include Maximum
+  Independent Set (MIS), Maximum Clique (MCL), and Maximum Cut (MCut) on Erdős–Rényi,
+  Revised Model B, and Barabási–Albert random graphs. RLSA achieves comparable or
+  superior performance to SOTA SA baselines at 20% of their runtime. RLNN trains efficiently
+  via local objectives without requiring labeled data or long-horizon return estimation.
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- rlsa_size
+- rlsa_time
+- optimization_trace
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run one packaged RLSA evaluation script on a supported combinatorial
+    problem. Uses accelerate with configurable precision (fp16/no). Set num_processes=1
+    for accurate timing. Tests on small graph instances (RB-[200-300], ER-[700-800],
+    BA-[200-300]) with 500-1000 test samples.
+
+    '
+- id: full_train_eval
+  label: Full Train + Eval
+  description: 'Run the full packaged regularized Langevin dynamics and neural baseline
+    workflow. Includes RLNN training with erdoes loss (differentiable) or reinforce
+    loss (black-box), followed by evaluation across small and large graph problem
+    instances.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - type: huggingface
+    url: https://huggingface.co/datasets/shengyuf/RLD4CO
+    description: Pre-generated test graphs including ER-[9000-11000] for scale evaluation
+  - type: generator
+    description: Random graph generators included in data/ folder (er_generator.py,
+      rb_generator.py, ba_generator.py)
+  notes:
+  - ER test graphs are from DIMES repository
+  - RB and BA graphs generated using DIffUCO code
+  - ER-[9000-11000] has no training data; evaluate solvers trained on ER-[700-800]
+    directly
+  - 1000 instances per problem (2 groups of 500-size test sets)
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.8'
+  cuda: '12.4'
+  pytorch: 2.4.1
+  flash_attn: null
+  key_packages:
+  - accelerate
+  - torch
+  - torch_geometric
+  - torch_scatter
+  - scipy
+  - numpy
+  - tqdm
+  notes:
+  - Install dependencies via conda environment.yml: conda env create -f environment.yml
+      && conda activate RLD4CO
+  - Framework built on accelerate for distributed evaluation
+  - Use --mixed_precision fp16 for faster inference
+risk_flags:
+- gpu_memory_bound
+- disk_io_bound
+- reproducibility_requires_seed
+risk_notes:
+- RLSA timing requires num_processes=1; multi-GPU parallelization affects wall-clock
+  measurements
+- Large graph evaluation (ER-[9000-11000]) requires significant GPU memory; use mixed
+  precision
+- RLNN training with reinforce loss is computationally intensive due to full black-box
+  gradient estimation
+- Random graph generation with multiprocessing (num_processes parameter) affects reproducibility
+recommended_when: 'Use this benchmark when you want a stochastic combinatorial-optimization
+  task with controllable sampling and neural baselines, need to evaluate simulated
+  annealing variants on random graphs, require unsupervised or black-box optimization
+  capability, or want to benchmark against ICML 2025 SOTA SA and NN-based CO solvers.
+
+  '
+not_recommended_when: 'Do not use this if you need a pure supervised-learning benchmark,
+  require exact/optimal solutions as ground truth, work with structured real-world
+  graph datasets rather than random generators, or need a deterministic reproducible
+  baseline without seed control.
+
+  '
+paper:
+  title: Regularized Langevin Dynamics for Combinatorial Optimization
+  venue: ICML 2025
+  year: 2025
+  url: https://arxiv.org/abs/2502.00277
+  authors:
+  - Shengyu Feng
+  - Yiming Yang
+  affiliation: Carnegie Mellon University
+  contact: shengyuf@cs.cmu.edu
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.073_rld.zip
+  archive_type: zip
+  local_dir_name: paper-73-RLD
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.073_rld.zip
+  sha256: 0cbaa2a2ca423ac065c17e38dc60fed2d00bf4a92a39cd1eab7c0a6decb7a2b9
+  size_bytes: 53784
+display:
+  palette_seed: brass-midnight-langevin
+  art_style: optimization-physics
+  accent_priority: high
+image_path: ../image/073_aisb.t3.073_rld.jpg
+capability_tags:
+- research_code_optimization
+- combinatorial_optimization
+- langevin_dynamics
+- graph_optimization
+- sampling
+- simulated_annealing
+- neural_heuristics
+- unsupervised_learning
+- black_box_optimization
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track

package/AISB/catalog/aisb.t3.073_rld.zh.yaml

@@ -0,0 +1,129 @@
+id: aisb.t3.073_rld
+name: 组合优化的正则化朗之万动力学
+version: 0.1.0
+one_line: 通过Hamming距离正则化的正则化朗之万动力学优化组合问题，用于模拟退火和神经求解器。
+task_description: >-
+  本基准测试包涵盖基于正则化朗之万动力学（RLD）的组合优化问题。RLD通过在采样解和当前解之间施加期望Hamming距离约束来逃离离散域中的局部最优。该基准测试包含两个求解器：基于梯度引导MCMC采样的RLSA（正则化朗之万模拟退火），以及支持可微分损失（erdoes）和完全黑盒损失（reinforce）的RLNN（正则化朗之万神经网络）。问题涵盖Erdős–Rényi、修订模型B和Barabási–Albert随机图上的最大独立集（MIS）、最大团（MCL）和最大割（MCut）。RLSA在仅需SOTA SA基线20%运行时间的情况下达到相当或更优的性能。RLNN通过局部目标高效训练，无需标注数据或长时域回报估计。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: advanced
+time_band: 6-24h
+cost_band: medium
+difficulty: hard
+data_access: public
+primary_outputs:
+- rlsa_size
+- rlsa_time
+- optimization_trace
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: >-
+    在支持的组合问题上运行一个封装的RLSA评估脚本。使用accelerate，支持可配置精度（fp16/无）。设置num_processes=1以获得精确计时。在小型图实例（RB-[200-300]、ER-[700-800]、BA-[200-300]）上测试，包含500-1000个测试样本。
+- id: full_train_eval
+  label: 完整训练+评估
+  description: >-
+    运行完整的封装正则化朗之万动力学和神经基线工作流程。包括使用erdoes损失（可微分）或reinforce损失（黑盒）训练RLNN，然后在小型和大型图问题实例上进行评估。
+dataset_download:
+  primary_method: mixed
+  sources:
+  - type: huggingface
+    url: https://huggingface.co/datasets/shengyuf/RLD4CO
+    description: 预生成的测试图，包括ER-[9000-11000]用于规模化评估
+  - type: generator
+    description: data/文件夹中包含的随机图生成器（er_generator.py、rb_generator.py、ba_generator.py）
+  notes:
+  - ER测试图来自DIMES仓库
+  - RB和BA图使用DIffUCO代码生成
+  - ER-[9000-11000]无训练数据；直接在ER-[700-800]上训练的求解器进行评估
+  - 每个问题1000个实例（2组500大小的测试集）
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 16
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.8'
+  cuda: '12.4'
+  pytorch: 2.4.1
+  flash_attn: null
+  key_packages:
+  - accelerate
+  - torch
+  - torch_geometric
+  - torch_scatter
+  - scipy
+  - numpy
+  - tqdm
+  notes:
+  - 通过conda environment.yml安装依赖：conda env create -f environment.yml && conda activate
+    RLD4CO
+  - 基于accelerate构建的分布式评估框架
+  - 使用--mixed_precision fp16加速推理
+risk_flags:
+- gpu_memory_bound
+- disk_io_bound
+- reproducibility_requires_seed
+risk_notes:
+- RLSA计时需要num_processes=1；多GPU并行化会影响实际运行时间测量
+- 大型图评估（ER-[9000-11000]）需要大量GPU内存；使用混合精度
+- 使用reinforce损失的RLNN训练由于完全黑盒梯度估计而计算密集
+- 使用多进程（num_processes参数）进行随机图生成会影响可重复性
+recommended_when: >-
+  当您需要一个具有可控采样和神经基线的随机组合优化任务时使用本基准测试；需要评估随机图上的模拟退火变体时；需要无监督或黑盒优化能力时；或者想要与ICML 2025 SOTA SA和基于NN的CO求解器进行基准对比时。
+not_recommended_when: >-
+  如果您需要一个纯监督学习基准测试时不要使用；需要精确/最优解作为真值时不要使用；处理结构化真实世界图数据集而非随机生成器时不要使用；或者需要一个没有种子控制的确定性可重复基线时不要使用。
+paper:
+  title: Regularized Langevin Dynamics for Combinatorial Optimization
+  venue: ICML 2025
+  year: 2025
+  url: https://arxiv.org/abs/2502.00277
+  authors:
+  - Shengyu Feng
+  - Yiming Yang
+  affiliation: Carnegie Mellon University
+  contact: shengyuf@cs.cmu.edu
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.073_rld.zip
+  archive_type: zip
+  local_dir_name: paper-73-RLD
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.073_rld.zip
+  sha256: 0cbaa2a2ca423ac065c17e38dc60fed2d00bf4a92a39cd1eab7c0a6decb7a2b9
+  size_bytes: 53784
+display:
+  palette_seed: brass-midnight-langevin
+  art_style: optimization-physics
+  accent_priority: high
+image_path: ../image/073_aisb.t3.073_rld.jpg
+capability_tags:
+- research_code_optimization
+- combinatorial_optimization
+- langevin_dynamics
+- graph_optimization
+- sampling
+- simulated_annealing
+- neural_heuristics
+- unsupervised_learning
+- black_box_optimization
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track