@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.045_pslstm.zh.yaml

@@ -0,0 +1,179 @@
+id: aisb.t3.045_pslstm
+name: P-sLSTM 长期时间序列预测
+version: 0.1.0
+one_line: >
+  通过通道独立性优化补丁化 sLSTM，在公开基准数据集上实现具有竞争力的
+  长期时间序列预测性能。
+task_description: |
+  P-sLSTM 通过两项改进将自然语言处理中的指数门控 sLSTM 架构适配到多变量时间序列预测：(1) 补丁化——将输入分割为局部时间补丁，以克服 sLSTM 的短记忆限制；以及 (2) 通道独立性——使用共享骨干网络独立处理每个变量，以防止过拟合。本基准测试评估标准公开数据集（包括 ETTh1/ETTm1、Weather、Electricity 等，来自 Time-Series-Library）的预测精度（MSE、MAE）。实验工作流使用每个数据集的 shell 脚本和集中式的 run_longExp.py 入口点。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: 高级
+time_band: 6-24h
+cost_band: medium
+difficulty: medium
+data_access: public
+primary_outputs:
+- mse
+- mae
+- forecast_checkpoint
+- rse
+- corr
+launch_profiles:
+- id: quick_eval
+  label: 快速评估
+  description: >
+    使用 scripts/EXP-LongForecasting/P_sLSTM/ 中提供的 shell 脚本，在单个准备好的公开数据集
+    （如 Weather）上运行 P-sLSTM 预测。
+- id: full_benchmark
+  label: 完整基准测试
+  description: >
+    在所有支持的数据集上运行完整的打包长期预测工作流，记录每个数据集的 MSE/MAE。
+dataset_download:
+  primary_method: external
+  sources:
+  - label: Time-Series-Library
+    url: https://github.com/thuml/Time-Series-Library
+    license: null
+  - label: iTransformer
+    url: https://github.com/thuml/iTransformer
+    license: null
+  notes:
+  - 数据集（ETTh1.csv、ETTm1.csv、weather.csv、electricity.csv 等）必须放置在 ./dataset/ 目录中。
+  - 预期结构：P-sLSTM/dataset/*.csv，并配合相应的 root_path 和 data_path 参数。
+credential_requirements:
+  mode: none
+  items: []
+  notes: []
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+environment:
+  python: '3.10'
+  cuda: '12.1'
+  pytorch: 2.1.0
+  flash_attn: null
+  key_packages:
+  - name: torch
+    version: '>=2.1.0'
+  - name: ninja
+    version: null
+  - name: matplotlib
+    version: null
+  notes:
+  - 经验证的可选方案：Python 3.12 配合 PyTorch 2.3.0，CUDA 12.1。
+  - 完整依赖列表见 requirements.txt。
+  - ninja 包需单独安装（pip install ninja）。
+risk_flags:
+- data_alignment
+- memory_bounded
+risk_notes:
+  - 确保数据集与预测范围参数（--label_len、--pred_len）对齐，以避免静默错配。
+  - GPU 显存随序列长度和通道数量增长；建议使用 24GB 显存运行完整基准测试。
+recommended_when: |
+  当您需要评估高效循环序列预测——避免 Transformer 的扩展成本，同时在长期多变量 TSF 任务上
+  获得具有竞争力的精度时，可使用此基准测试。
+not_recommended_when: |
+  如果您需要多模态输入（如视觉-语言）、超大规模语言模型预训练或非预测类序列任务，
+  请勿使用此基准测试。
+paper:
+  title: Unlocking the Power of LSTM for Long Term Time Series Forecasting
+  venue: AAAI
+  year: 2025
+  url: https://ojs.aaai.org/index.php/AAAI/article/view/33303
+  authors:
+  - Kong, Yaxuan
+  - Wang, Zepu
+  - Nie, Yuqi
+  - Zhou, Tian
+  - Zohren, Stefan
+  - Liang, Yuxuan
+  - Sun, Peng
+  - Wen, Qingsong
+  bibtex: "@inproceedings{kong2025unlocking,\n  title={Unlocking the power of lstm for long term time series forecasting},\n  author={Kong, Yaxuan and Wang, Zepu and Nie, Yuqi and Zhou, Tian and Zohren, Stefan and Liang, Yuxuan and Sun, Peng and Wen, Qingsong},\n  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},\n  volume={39},\n  number={11},\n  pages={11968--11976},\n  year={2025}\n}"
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.045_pslstm.zip
+  archive_type: zip
+  local_dir_name: paper-45-PsLSTM
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.045_pslstm.zip
+  sha256: 27a81f0bd5307f83bd05c52a8332d2d71c1ff6de8eb7c3e04826a0337c201fe5
+  size_bytes: 1001820
+display:
+  palette_seed: olive-sky-lstm
+  art_style: 循环信号风格
+  accent_priority: medium
+image_path: ../image/045_aisb.t3.045_pslstm.jpg
+capability_tags:
+- 科研代码优化
+- 时间序列预测
+- 循环模型
+- 预测任务
+- 科学机器学习
+track_fit:
+- paper_track
+- benchmark_track
+aisb_direction: T3
+executable_anchors:
+- path: run_longExp.py
+  role: main entry point
+  description: 解析模型、数据集和训练配置的 CLI 参数；编排 Exp_Main。
+- path: exp/exp_main.py
+  role: experiment driver
+  description: Exp_Main 类构建模型（P_sLSTM、LSTM、sLSTM、Informer、Autoformer、DLinear、Transformer），
+    加载数据，进行训练和评估。
+- path: utils/metrics.py
+  role: metric computation
+  description: 实现 MAE、MSE、RMSE、MAPE、MSPE、RSE、CORR 指标；metric() 返回 (mae, mse, rmse, mape, mspe, rse, corr)。
+code_roots:
+- data_provider/
+- exp/
+- layers/
+- models/
+- scripts/
+- utils/
+metric_contract:
+  mse:
+    origin_path: utils/metrics.py
+    function: MSE
+    source_ref: code-backed
+    status: executable
+  mae:
+    origin_path: utils/metrics.py
+    function: MAE
+    source_ref: code-backed
+    status: executable
+  rse:
+    origin_path: utils/metrics.py
+    function: RSE
+    source_ref: code-backed
+    status: executable
+  corr:
+    origin_path: utils/metrics.py
+    function: CORR
+    source_ref: code-backed
+    status: executable
+recovery_notes: |
+  静态代码审计发现所有阶段性指标都有可执行锚点。
+  未执行运行时测试；指标值尚未经过验证。在选择 origin_path、source_ref 和
+  evaluation_protocol.code_paths 时，应优先恢复现有代码文件而非仅依赖 README 的声明。
+caveat_snapshot: |
+  打包快照可通过提供的 shell 脚本运行。
+  在指标值可用于对比报告之前，需要在所有数据集上执行完整的基准测试。
+commercial:
+  annual_fee: null

package/AISB/catalog/aisb.t3.046_nonstatts.yaml

@@ -0,0 +1,208 @@
+id: aisb.t3.046_nonstatts
+name: TAFAS
+version: 0.1.0
+one_line: Test-time adaptation framework for non-stationary time series forecasting
+  using periodicity-aware scheduling and gated calibration modules.
+task_description: 'This packaged benchmark implements TAFAS (Test-time Adaptive ForecAsting
+  for non-stationary time Series), a pioneering test-time adaptation framework for
+  time series forecasting. The core problem addressed is that pre-trained forecasters
+  degrade under non-stationary test distributions where temporal dynamics evolve continuously.
+  Unlike traditional test-time adaptation in vision (which assumes unlabeled IID inputs),
+  TSF-TTA exploits the sequential nature of time series where partially-observed ground
+  truth becomes available during inference. TAFAS consists of two technical components:
+  (1) PAAS (Periodicity-Aware Adaptation Scheduling) that adaptively selects sufficient
+  partial ground truth windows to capture meaningful periodic patterns, and (2) GCM
+  (Gated Calibration Module) that calibrates test-time inputs to distributions the
+  source forecaster handles, with gating controlling the degree of calibration based
+  on global distribution shifts. The source forecaster remains frozen during adaptation
+  to preserve pre-trained semantic knowledge. The benchmark supports multiple base
+  forecaster architectures and benchmarks on ETTh1, ETTm1, ETTh2, ETTm2, exchange_rate,
+  and weather datasets. Particularly effective for long-term forecasting scenarios
+  with pronounced distribution shifts.
+
+  '
+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_mse
+- test_mae
+- adaptation_report
+launch_profiles:
+- id: quick_eval
+  label: Quick Eval
+  description: 'Run a single model-dataset pair through the TTA workflow with prediction
+    length 720. Executes scripts/{model}/{dataset}_720/run.sh. Suitable for validating
+    one adaptation scenario.
+
+    '
+- id: full_adaptation
+  label: Full Adaptation
+  description: 'Run the complete test-time adaptation workflow across all supported
+    model-dataset combinations. Evaluates PatchTST, DLinear, OLS, and FreTS on ETTh1,
+    ETTm1, ETTh2, ETTm2, exchange_rate, and weather datasets at various prediction
+    lengths.
+
+    '
+dataset_download:
+  primary_method: external
+  sources:
+  - name: Time-Series-Library
+    url: https://github.com/thuml/Time-Series-Library
+    type: dataset_repository
+    notes: Benchmark datasets sourced from the Time-Series-Library repository
+  - name: Google Drive
+    url: https://drive.google.com/drive/folders/16UdpRYf3BLRfem2LXLspN78dW4MOJkbn
+    type: model_checkpoints
+    notes: 'Pre-trained checkpoints provided for pred_len=720 only. Models: PatchTST,
+      DLinear, OLS, FreTS. Datasets: ETTh1, ETTm1, ETTh2, ETTm2, exchange_rate, weather.
+
+      '
+  - name: DeepScientist Archive
+    url: https://deepscientist.cc/AISB/046_nonstatts
+    type: packaged_code
+    archive_type: zip
+  placement:
+    datasets: data/{dataset}/{dataset}.csv
+    checkpoints: checkpoints/{model}/{dataset}_{pred_len}/checkpoint_best.pth
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - Public access only; no authentication required
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+    notes: Single GPU sufficient for evaluation; training requires longer time
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: '24GB VRAM recommended for batch_size 256 with longer sequence lengths.
+      Multi-GPU support available via NUM_GPUS config (tested up to 8).
+
+      '
+environment:
+  python: '3.8'
+  cuda: '11.8'
+  pytorch: 2.0.0
+  flash_attn: null
+  key_packages:
+  - wandb
+  - yacs
+  - pandas
+  - numpy
+  - torch
+  - tqdm
+  notes:
+  - See bundled requirements.txt or README for full dependency set
+  - Core model implementations reference the Time-Series-Library
+  - GCM adaptation uses PyTorch autograd; no special CUDA extensions required
+risk_flags:
+- no_runtime_execution_verified
+- checkpoint_availability_limited
+risk_notes:
+- Static code audit confirmed executable anchors for all staged metrics
+- No benchmark execution performed during packaging pass
+- Pre-trained checkpoints only available for pred_len=720; other lengths require training
+- Training can be enabled by setting TRAIN.ENABLE=True in run.sh
+recommended_when: '- Evaluating forecasting models under distribution-shifted test
+  conditions - Studying test-time adaptation techniques for sequential data - Benchmarking
+  long-term forecasting with non-stationary patterns - Comparing pre-training-only
+  approaches against adaptive methods - Researching model-agnostic adaptation for
+  time series foundation models
+
+  '
+not_recommended_when: '- Only needing static in-distribution forecasting baselines
+  - Requiring pre-trained checkpoints for prediction lengths other than 720 - Working
+  with datasets not in the supported benchmark set - Needing immediate production
+  deployment without adaptation tuning
+
+  '
+paper:
+  title: Battling the Non-stationarity in Time Series Forecasting via Test-time Adaptation
+  venue: AAAI 2025
+  year: 2025
+  authors:
+  - HyunGi Kim
+  - Siwon Kim
+  - Jisoo Mok
+  - Sungroh Yoon
+  institution: Seoul National University
+  url: https://arxiv.org/abs/2501.04970
+  github: https://github.com/kimanki/TAFAS
+  bibtex: '@article{kim2025battling, title={Battling the Non-stationarity in Time
+    Series Forecasting via Test-time Adaptation}, author={Kim, HyunGi and Kim, Siwon
+    and Mok, Jisoo and Yoon, Sungroh}, journal={arXiv preprint arXiv:2501.04970},
+    year={2025}}'
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.046_nonstatts.zip
+  archive_type: zip
+  local_dir_name: paper-46-NonStatTS
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.046_nonstatts.zip
+  sha256: 58816149c85cd2e8433b166dec9b2336ec49f0cc1a9c4fb24b986022663d3dee
+  size_bytes: 341067
+display:
+  palette_seed: blue-sand-adapt
+  art_style: adaptive-signal
+  accent_priority: medium
+  tags:
+  - time-series-forecasting
+  - test-time-adaptation
+  - distribution-shift
+  - non-stationarity
+  - tta
+  - tsf-tta
+image_path: ../image/046_aisb.t3.046_nonstatts.jpg
+execution_notes:
+  entry_point: main.py
+  config_file: config.py
+  predictor_file: predictor.py
+  script_pattern: scripts/{model}/{dataset}_{pred_len}/run.sh
+  supported_models:
+  - PatchTST
+  - DLinear
+  - OLS
+  - FreTS
+  - Informer
+  - Autoformer
+  - FEDformer
+  - ETSformer
+  - MICN
+  - Koopa
+  - TimesNet
+  - iTransformer
+  supported_datasets:
+  - ETTh1
+  - ETTm1
+  - ETTh2
+  - ETTm2
+  - exchange_rate
+  - weather
+  prediction_lengths:
+  - 96
+  - 192
+  - 336
+  - 720
+  code_structure:
+    layers: Transformer and convolution layer implementations
+    models: Forecaster architectures including TAFAS components
+    tta: tafas.py contains PAAS and GCM implementations
+    scripts: Model-specific evaluation and training scripts
+    utils: Logging, parsing, masking, and meter utilities

package/AISB/catalog/aisb.t3.046_nonstatts.zh.yaml

@@ -0,0 +1,194 @@
+id: aisb.t3.046_nonstatts
+name: TAFAS
+version: 0.1.0
+one_line: >
+  针对非平稳时间序列预测的测试时自适应框架，采用周期感知调度和门控校准模块。
+task_description: >
+  本基准测试实现了 TAFAS（针对非平稳时间序列的测试时自适应预测），
+  这是一个开创性的时间序列预测测试时自适应框架。
+  核心问题在于：预训练预测模型在非平稳测试分布下会退化，因为时间动态不断演变。
+  不同于传统视觉领域的测试时自适应（假设无标签的独立同分布输入），
+  TSF-TTA 利用时间序列的序列特性，在推理过程中可获取部分观测到的真实值。
+  TAFAS 由两个技术组件构成：（1）PAAS（周期感知自适应调度），
+  自适应选择足够的部分真实值窗口以捕捉有意义的周期模式，
+  （2）GCM（门控校准模块），将测试时输入校准到源预测器能够处理分布，
+  门控机制根据全局分布偏移程度控制校准强度。
+  源预测器在自适应过程中保持冻结，以保留预训练的语义知识。
+  该基准测试支持多种基础预测器架构，在 ETTh1、ETTm1、ETTh2、ETTm2、
+  exchange_rate 和 weather 数据集上进行基准测试。在分布偏移明显的长时预测场景中表现尤为出色。
+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_mse
+  - test_mae
+  - adaptation_report
+launch_profiles:
+  - id: quick_eval
+    label: 快速评估
+    description: >
+      在单个模型-数据集对上运行 TTA 工作流，预测长度为 720。
+      执行 scripts/{model}/{dataset}_720/run.sh。
+      适用于验证一种自适应场景。
+  - id: full_adaptation
+    label: 完整自适应
+    description: >
+      在所有支持的模型-数据集组合上运行完整的测试时自适应工作流。
+      在 ETTh1、ETTm1、ETTh2、ETTm2、exchange_rate 和 weather 数据集上
+      评估 PatchTST、DLinear、OLS 和 FreTS 模型的多种预测长度。
+dataset_download:
+  primary_method: external
+  sources:
+    - name: Time-Series-Library
+      url: https://github.com/thuml/Time-Series-Library
+      type: dataset_repository
+      notes: 基准测试数据集来源于 Time-Series-Library 仓库
+    - name: Google Drive
+      url: https://drive.google.com/drive/folders/16UdpRYf3BLRfem2LXLspN78dW4MOJkbn
+      type: model_checkpoints
+      notes: >
+        预训练检查点仅提供 pred_len=720 版本。
+        模型包括：PatchTST、DLinear、OLS、FreTS。
+        数据集包括：ETTh1、ETTm1、ETTh2、ETTm2、exchange_rate、weather。
+    - name: DeepScientist Archive
+      url: https://deepscientist.cc/AISB/046_nonstatts
+      type: packaged_code
+      archive_type: zip
+  placement:
+    datasets: data/{dataset}/{dataset}.csv
+    checkpoints: checkpoints/{model}/{dataset}_{pred_len}/checkpoint_best.pth
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+    - 仅需公开访问权限，无需身份验证
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 32
+    disk_gb: 80
+    gpu_count: 1
+    gpu_vram_gb: 12
+    notes: 单 GPU 足以进行评估；训练需要更长时间
+  recommended:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: >
+      建议使用 24GB 显存以支持批量大小 256 和更长序列长度。
+      可通过 NUM_GPUS 配置支持多 GPU（已测试至 8 卡）。
+environment:
+  python: '3.8'
+  cuda: '11.8'
+  pytorch: 2.0.0
+  flash_attn: null
+  key_packages:
+    - wandb
+    - yacs
+    - pandas
+    - numpy
+    - torch
+    - tqdm
+  notes:
+    - 详见打包的 requirements.txt 或 README 获取完整依赖列表
+    - 核心模型实现参考 Time-Series-Library
+    - GCM 自适应使用 PyTorch autograd，无需特殊 CUDA 扩展
+risk_flags:
+  - no_runtime_execution_verified
+  - checkpoint_availability_limited
+risk_notes:
+  - 静态代码审计确认所有分阶段指标的可执行锚点
+  - 打包过程中未执行基准测试
+  - 预训练检查点仅在 pred_len=720 时可用；其他长度需要训练
+  - 可通过在 run.sh 中设置 TRAIN.ENABLE=True 来启用训练
+recommended_when: >
+  - 在分布偏移的测试条件下评估预测模型
+  - 研究针对序列数据的测试时自适应技术
+  - 对具有非平稳模式的长时预测进行基准测试
+  - 比较纯预训练方法与自适应方法
+  - 研究时间序列基础模型的通用自适应方法
+not_recommended_when: >
+  - 仅需要静态分布内预测基线
+  - 需要预测长度 720 以外的预训练检查点
+  - 使用不在支持基准集合中的数据集
+  - 需要在未进行自适应调优的情况下立即投入生产部署
+paper:
+  title: Battling the Non-stationarity in Time Series Forecasting via Test-time Adaptation
+  venue: AAAI 2025
+  year: 2025
+  authors:
+    - HyunGi Kim
+    - Siwon Kim
+    - Jisoo Mok
+    - Sungroh Yoon
+  institution: Seoul National University
+  url: https://arxiv.org/abs/2501.04970
+  github: https://github.com/kimanki/TAFAS
+  bibtex: '@article{kim2025battling, title={Battling the Non-stationarity in Time Series Forecasting via Test-time Adaptation}, author={Kim, HyunGi and Kim, Siwon and Mok, Jisoo and Yoon, Sungroh}, journal={arXiv preprint arXiv:2501.04970}, year={2025}}'
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.046_nonstatts.zip
+  archive_type: zip
+  local_dir_name: paper-46-NonStatTS
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.046_nonstatts.zip
+  sha256: 58816149c85cd2e8433b166dec9b2336ec49f0cc1a9c4fb24b986022663d3dee
+  size_bytes: 341067
+display:
+  palette_seed: blue-sand-adapt
+  art_style: adaptive-signal
+  accent_priority: medium
+  tags:
+    - time-series-forecasting
+    - test-time-adaptation
+    - distribution-shift
+    - non-stationarity
+    - tta
+    - tsf-tta
+image_path: ../image/046_aisb.t3.046_nonstatts.jpg
+execution_notes:
+  entry_point: main.py
+  config_file: config.py
+  predictor_file: predictor.py
+  script_pattern: scripts/{model}/{dataset}_{pred_len}/run.sh
+  supported_models:
+    - PatchTST
+    - DLinear
+    - OLS
+    - FreTS
+    - Informer
+    - Autoformer
+    - FEDformer
+    - ETSformer
+    - MICN
+    - Koopa
+    - TimesNet
+    - iTransformer
+  supported_datasets:
+    - ETTh1
+    - ETTm1
+    - ETTh2
+    - ETTm2
+    - exchange_rate
+    - weather
+  prediction_lengths:
+    - 96
+    - 192
+    - 336
+    - 720
+  code_structure:
+    layers: Transformer 和卷积层实现
+    models: 包括 TAFAS 组件的预测器架构
+    tta: tafas.py 包含 PAAS 和 GCM 实现
+    scripts: 模型特定的评估和训练脚本
+    utils: 日志、解析、掩码和计量工具