@researai/deepscientist 1.5.16 → 1.6.0

package/AISB/catalog/aisb.t3.074_lsvi.yaml

@@ -0,0 +1,163 @@
+id: aisb.t3.074_lsvi
+name: Least Squares Variational Inference
+version: 0.1.0
+one_line: Gradient-free variational inference via least-squares updates on tempered
+  log-target evaluations, supporting Gaussian BBVI and exponential-family posteriors
+  without gradient estimators.
+task_description: 'This packaged benchmark implements LSVI (Least Squares Variational
+  Inference), a gradient-free, Monte Carlo-based scheme for variational inference
+  within exponential families. Instead of gradient-based optimization, LSVI performs
+  successive ordinary least squares regression on tempered log-target evaluations
+  under the variational approximation. Each iteration solves a fixed-point equation
+  via OLS, yielding a bias-corrected stochastic natural gradient descent step. The
+  approach avoids unstable gradient estimators required when gradients of log pi are
+  unavailable, such as for discrete distributions or likelihood-free inference. The
+  benchmark includes Gaussian variational family implementations that eliminate Fisher
+  information matrix inversion: O(d) complexity for mean-field Gaussians and O(d^3)
+  for full-covariance cases. Primary evaluation uses the Pima Indians diabetes logistic
+  regression task, measuring KL divergence to the true posterior via the kl_objective
+  metric.
+
+  '
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 2-6h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- kl_objective
+- variational_inference_report
+launch_profiles:
+- id: gaussian_logreg
+  label: Gaussian Logistic Regression
+  description: 'Execute the packaged least-squares variational inference experiment
+    using Gaussian variational approximation on the Pima Indians diabetes binary classification
+    task. Runs 8 random seeds with 8000-step phase-1 and 1500-step phase-2 optimization.
+    Outputs KL divergence metrics and variational parameter trajectories.
+
+    '
+dataset_download:
+  primary_method: bundled
+  sources:
+  - type: package
+    path: LSVI/
+    description: Benchmark code and experiments are bundled in the archive
+  notes:
+  - Pima Indians dataset is included in the package under experiments/logisticRegression/
+  - No external data fetching required
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - All data and code are self-contained in the package
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+    gpu_vram_gb: 0
+    notes:
+    - CPU-only execution fully supported via JAX CPU backend
+    - 8-core CPU recommended for parallel seed execution
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 0
+    gpu_vram_gb: 0
+    notes:
+    - Additional cores reduce wall-clock time for multi-seed experiments
+    - GPU is optional; JAX supports GPU acceleration if available
+environment:
+  python: '3.10'
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - jax>=0.4.0
+  - jaxlib
+  - numpy>=1.21
+  - scipy
+  notes:
+  - JAX is the primary computation library; no PyTorch dependency
+  - GPU support is optional via CUDA-enabled jaxlib
+  - CPU-only execution is the primary tested configuration
+  - See bundled requirements.txt or setup.py for full dependency list
+risk_flags: []
+risk_notes:
+- No external API calls; fully offline benchmark
+- CPU-bound execution eliminates GPU availability risk
+- Long runtime (up to 6h) but low computational cost
+recommended_when: 'Use this benchmark for evaluating probabilistic inference methods
+  that must operate without gradient access, including discrete variable models, likelihood-free
+  inference, and non-differentiable posteriors. The CPU-friendly execution profile
+  makes it suitable for resource-constrained environments or methodological research
+  on variational inference.
+
+  '
+not_recommended_when: 'Do not use this benchmark for evaluating large-scale deep learning
+  methods, GPU-accelerated neural network training, or tasks requiring automatic differentiation
+  of the target density. It does not test gradient-based VI alternatives like reparameterization
+  tricks or score-function estimators.
+
+  '
+paper:
+  title: Least squares variational inference
+  venue: NeurIPS
+  year: 2025
+  url: https://arxiv.org/abs/2502.18475
+  authors:
+  - Yvann Le Fay
+  - Nicolas Chopin
+  - Simon Barthelme
+  affiliation:
+  - ENSAE, CREST, IP Paris
+  - GIPSA-Lab, CNRS
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.074_lsvi.zip
+  archive_type: zip
+  local_dir_name: paper-74-LSVI
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.074_lsvi.zip
+  sha256: 2f33740ed8cefdcd2bfd419c0ad454037de523a1b5a515e8da465918cc54af7a
+  size_bytes: 71181
+display:
+  palette_seed: cream-slate-gaussian
+  art_style: probabilistic-notebook
+  accent_priority: medium
+image_path: ../image/074_aisb.t3.074_lsvi.jpg
+capability_tags:
+- variational_inference
+- bayesian_inference
+- gradient_free_optimization
+- probabilistic_modeling
+- research_code_optimization
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+code_anchors:
+  primary_executable:
+    path: LSVI/experiments/logisticRegression/pima_logistic_regression/gaussian.py
+    description: Gaussian variational inference experiment for Pima logistic regression
+      with BBVI baseline
+  variational_modules:
+    path: LSVI/variational/
+    description: Core variational inference implementations including LSVI, Laplace
+      approximation, and exponential family distributions
+metric_contract:
+  primary_metric: kl_objective
+  backing: LSVI/experiments/logisticRegression/pima_logistic_regression/gaussian.py
+  provenance: code-backed
+caveats:
+- No benchmark execution was performed during this packaging pass
+- Runtime metric values require validation before publication
+- Top-level README is absent in the current snapshot

package/AISB/catalog/aisb.t3.074_lsvi.zh.yaml

@@ -0,0 +1,129 @@
+id: aisb.t3.074_lsvi
+name: Least Squares Variational Inference
+version: 0.1.0
+one_line: 通过对温度化对数目标评估做最小二乘更新实现无梯度变分推断，支持高斯 BBVI 与指数族后验，无需梯度估计器。
+task_description: 该基准封装了 LSVI（最小二乘变分推断），一种在指数族中进行变分推断的无梯度蒙特卡洛方法。LSVI 不做梯度优化，而是在变分近似分布下对温度化对数目标进行连续 OLS 回归。每次迭代通过 OLS 求解不动点方程，得到偏差校正的随机自然梯度下降步。该方法适用于无法获得 log pi 梯度的场景（如离散分布或无似然推断），可避免不稳定梯度估计器。基准包含无需 Fisher 信息矩阵求逆的高斯变分族实现：均值场 O(d)，全协方差 O(d^3)。主要评测任务为 Pima Indians 糖尿病逻辑回归，使用 kl_objective 衡量与真实后验的 KL 散度。
+task_mode: evaluation_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: runnable
+support_level: turnkey
+time_band: 2-6h
+cost_band: low
+difficulty: medium
+data_access: public
+primary_outputs:
+- kl_objective
+- variational_inference_report
+launch_profiles:
+- id: gaussian_logreg
+  label: 高斯逻辑回归
+  description: 在 Pima Indians 糖尿病二分类任务上运行封装的最小二乘变分推断实验。使用 8 个随机种子，阶段一 8000 步、阶段二 1500 步优化，输出 KL 散度与变分参数轨迹。
+dataset_download:
+  primary_method: bundled
+  sources:
+  - type: package
+    path: LSVI/
+    description: Benchmark code and experiments are bundled in the archive
+  notes:
+  - Pima Indians dataset is included in the package under experiments/logisticRegression/
+  - No external data fetching required
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - All data and code are self-contained in the package
+resources:
+  minimum:
+    cpu_cores: 8
+    ram_gb: 16
+    disk_gb: 20
+    gpu_count: 0
+    gpu_vram_gb: 0
+    notes:
+    - CPU-only execution fully supported via JAX CPU backend
+    - 8-core CPU recommended for parallel seed execution
+  recommended:
+    cpu_cores: 16
+    ram_gb: 32
+    disk_gb: 50
+    gpu_count: 0
+    gpu_vram_gb: 0
+    notes:
+    - Additional cores reduce wall-clock time for multi-seed experiments
+    - GPU is optional; JAX supports GPU acceleration if available
+environment:
+  python: '3.10'
+  cuda: null
+  pytorch: null
+  flash_attn: null
+  key_packages:
+  - jax>=0.4.0
+  - jaxlib
+  - numpy>=1.21
+  - scipy
+  notes:
+  - JAX is the primary computation library; no PyTorch dependency
+  - GPU support is optional via CUDA-enabled jaxlib
+  - CPU-only execution is the primary tested configuration
+  - See bundled requirements.txt or setup.py for full dependency list
+risk_flags: []
+risk_notes:
+- No external API calls; fully offline benchmark
+- CPU-bound execution eliminates GPU availability risk
+- Long runtime (up to 6h) but low computational cost
+recommended_when: 当你需要评估在无梯度条件下工作的概率推断方法时使用本基准，包括离散变量模型、无似然推断与不可微后验。其 CPU 友好配置也适合资源受限环境中的变分推断研究。
+not_recommended_when: 不适用于评估大规模深度学习、GPU 加速神经网络训练，或需要目标密度自动微分的任务。它也不覆盖重参数化技巧或 score-function 估计等梯度型 VI 方案。
+paper:
+  title: Least squares variational inference
+  venue: NeurIPS
+  year: 2025
+  url: https://arxiv.org/abs/2502.18475
+  authors:
+  - Yvann Le Fay
+  - Nicolas Chopin
+  - Simon Barthelme
+  affiliation:
+  - ENSAE, CREST, IP Paris
+  - GIPSA-Lab, CNRS
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.074_lsvi.zip
+  archive_type: zip
+  local_dir_name: paper-74-LSVI
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.074_lsvi.zip
+  sha256: 2f33740ed8cefdcd2bfd419c0ad454037de523a1b5a515e8da465918cc54af7a
+  size_bytes: 71181
+display:
+  palette_seed: cream-slate-gaussian
+  art_style: probabilistic-notebook
+  accent_priority: medium
+image_path: ../image/074_aisb.t3.074_lsvi.jpg
+capability_tags:
+- variational_inference
+- bayesian_inference
+- gradient_free_optimization
+- probabilistic_modeling
+- research_code_optimization
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+code_anchors:
+  primary_executable:
+    path: LSVI/experiments/logisticRegression/pima_logistic_regression/gaussian.py
+    description: Gaussian variational inference experiment for Pima logistic regression with BBVI baseline
+  variational_modules:
+    path: LSVI/variational/
+    description: Core variational inference implementations including LSVI, Laplace approximation, and exponential family distributions
+metric_contract:
+  primary_metric: kl_objective
+  backing: LSVI/experiments/logisticRegression/pima_logistic_regression/gaussian.py
+  provenance: code-backed
+caveats:
+- No benchmark execution was performed during this packaging pass
+- Runtime metric values require validation before publication
+- Top-level README is absent in the current snapshot

package/AISB/catalog/aisb.t3.075_treeslicedentropy.yaml

@@ -0,0 +1,201 @@
+id: aisb.t3.075_treeslicedentropy
+name: Tree-Sliced Entropy Partial Transport
+version: 0.1.0
+one_line: GPU-accelerated unbalanced partial transport via tree-sliced entropy for
+  point clouds, image generation, and image-to-image translation.
+task_description: 'This packaged benchmark covers PartialTSW (Partial Tree-Sliced
+  Wasserstein), an extension of Tree-Sliced Wasserstein distances to unbalanced/partial
+  transport settings with closed-form formulations suitable for dynamic-support distributions
+  in generative modeling. The task requires selecting a concrete experiment family
+  (point cloud gradient flow, image generation with JAX/diffusers, or image-to-image
+  translation) and running the corresponding transport objective or ablation sweep.
+  Evaluation focuses on convergence quality (relative_bias_pct), runtime performance
+  (mean_time_s), and stability (coefficient_of_variation) rather than internal optimizer
+  accuracy targets. The benchmark is code-backed by three executable anchors: convergence
+  analysis (`analysis/convergence/simple_convergence.py`), runtime benchmarking varying
+  sample count (`analysis/runtime_plot/runtime_plot_compare_n_partial.py`), and runtime
+  benchmarking varying dimensionality (`analysis/runtime_plot/runtime_plot_compare_d_partial.py`).
+
+  '
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: recovery
+time_band: 2-4d
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- relative_bias_pct
+- mean_time_s
+- coefficient_of_variation
+launch_profiles:
+- id: convergence_eval
+  label: Convergence Evaluation
+  description: 'Run `analysis/convergence/simple_convergence.py` to assess estimator
+    bias and stability across varying tree counts (L) and lines per tree (k). Uses
+    PyTorch with GPU acceleration. Outputs relative bias percentage and coefficient
+    of variation per configuration.
+
+    '
+- id: runtime_n_scaling
+  label: Runtime Scaling (Sample Count)
+  description: 'Run `analysis/runtime_plot/runtime_plot_compare_n_partial.py` to benchmark
+    runtime and GPU memory across sample sizes N ∈ [10², 10⁵]. Compares PartialTSW
+    against UOT baselines (sopt, spot, pawl, sliced_unbalanced_ot). Includes warm-up
+    runs for stable timing.
+
+    '
+- id: runtime_d_scaling
+  label: Runtime Scaling (Dimension)
+  description: 'Run `analysis/runtime_plot/runtime_plot_compare_d_partial.py` to benchmark
+    runtime and GPU memory across dimensions d ∈ [1, 2048] with fixed N=10⁴. Compares
+    PartialTSW against UOT baselines. Includes warm-up runs for stable timing.
+
+    '
+- id: experiment_family
+  label: Experiment Family Workflow
+  description: 'Select and execute one packaged experiment family: point_cloud (gradient
+    flow), image_gen (JAX/diffusers ablation sweep across multiple GPUs), or img2img
+    (image-to-image translation with class imbalance handling). Each family has distinct
+    data requirements and runtime profiles.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - torchvision (auto-download for MNIST-based experiments)
+  - manual (pretrained checkpoints for image generation, assets for img2img)
+  notes:
+  - Convergence and runtime eval scripts use synthetic Gaussian data and require no
+    external datasets.
+  - Image generation experiments may auto-download simple datasets through torchvision.
+  - Image-to-image translation and some ablation routes require manual asset preparation
+    or pretrained checkpoints before evaluation.
+  - Treat dataset setup as experiment-family specific; there is no universal dataset
+    route.
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No explicit API or dataset credential variables were detected in the shipped source
+    snapshot.
+  - Image generation sweep uses explicit GPU slot scheduling (not cloud credential-based).
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: GPU required for meaningful runtime benchmarks; CPU-only execution is impractical
+      for N≥10⁴ or d≥64 configurations.
+  recommended:
+    cpu_cores: 32
+    ram_gb: 128
+    disk_gb: 300
+    gpu_count: 1
+    gpu_vram_gb: 48
+    notes: '48 GB VRAM accommodates full dimension sweep (up to d=2048) and larger
+      sample counts. Image generation ablation grid benefits from additional GPU slots
+      if available.
+
+      '
+environment:
+  python: '3.11'
+  cuda: '11.8'
+  pytorch: 2.1.0
+  flash_attn: null
+  key_packages:
+  - jax==0.4.23
+  - jaxlib==0.4.23
+  - transformers==4.51.3
+  - diffusers==0.33.1
+  - torch>=2.1.0
+  - ot (POT library for baseline comparisons)
+  - tqdm
+  - matplotlib
+  - pandas
+  - numpy
+  notes:
+  - The packaged repository mixes PyTorch (analysis scripts), JAX/jaxlib (image generation),
+    and diffusion-style experiment routes rather than one narrow runtime profile.
+  - Analysis scripts default to CUDA with `torch.device('cuda' if torch.cuda.is_available()
+    else 'cpu')`.
+  - Image generation sweep script (`experiments/image_gen/exp.sh`) schedules 72 jobs
+    round-robin across allocated GPU slots; wall-clock time scales with GPU count.
+  - Baseline comparisons in runtime benchmarks import sopt, spot, pawl, sliced_unbalanced_ot,
+    and unbalanced_sliced_ot from the bundled baselines module.
+risk_flags:
+- route_caveat
+- multi_route_experiment_family
+- mixed_runtime_environment
+- multi_gpu_scheduling_dependent
+risk_notes:
+- The packaged repository is structurally complete, but the optimizer target is not
+  one root command; point-cloud, image-generation, image-to-image, and analysis routes
+  differ materially in setup and execution.
+- The packaged `experiments/image_gen/exp.sh` schedules 72 jobs round-robin across
+  the listed GPUs, so a full ablation sweep on one GPU is naturally multi-day.
+- The runtime environment mixes PyTorch (convergence, runtime analysis), JAX/jaxlib
+  (image generation), and diffusion libraries; ensure both CUDA toolchains are compatible.
+- Some subroutes can auto-download simple datasets such as MNIST, while others require
+  manual assets or pretrained checkpoints before evaluation.
+- Convergence eval uses synthetic Gaussian data only; generative downstream quality
+  must be evaluated separately via experiment family routes.
+recommended_when: 'Use this benchmark when you need a geometry-heavy optimal transport
+  benchmark with support for unbalanced measures, are prepared to explicitly select
+  one experiment family, and have GPU infrastructure suitable for multi-day ablation
+  sweeps. Ideal for comparing partial transport solvers on convergence bias, runtime
+  scaling, or generative applications with class imbalance.
+
+  '
+not_recommended_when: 'Do not use this benchmark if you require a single one-command
+  execution, a CPU-first workflow, a single compact dataset with no experiment branching,
+  or pre-computed metrics without execution. Not suitable for environments without
+  CUDA-capable GPUs.
+
+  '
+paper:
+  title: Tree-Sliced Entropy Partial Transport
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=41ZbysfW4h
+  authors:
+  - Viet-Hoang Tran (National University of Singapore)
+  - Thanh Tran (VinUniversity)
+  - Thanh Chu (National University of Singapore)
+  - Tam Le (Institute of Statistical Mathematics)
+  - Tan M. Nguyen (National University of Singapore)
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.075_treeslicedentropy.zip
+  archive_type: zip
+  local_dir_name: paper-75-TreeSlicedEntropy
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.075_treeslicedentropy.zip
+  sha256: c73a8da6a3b04ef4a24eb1345993f985e84234db16636906626f07cb762fa33c
+  size_bytes: 774639
+display:
+  palette_seed: jade-bone-transport
+  art_style: geometric-minimal
+  accent_priority: high
+image_path: ../../../AISB/image/075_aisb.t3.075_treeslicedentropy.jpg
+capability_tags:
+- optimal_transport
+- tree_sliced_wasserstein
+- partial_transport
+- unbalanced_measures
+- generative_modeling
+- probabilistic_geometry
+- scientific_ml
+- research_code_optimization
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null

package/AISB/catalog/aisb.t3.075_treeslicedentropy.zh.yaml

@@ -0,0 +1,148 @@
+id: aisb.t3.075_treeslicedentropy
+name: Tree-Sliced Entropy Partial Transport
+version: 0.1.0
+one_line: 基于树切片熵的 GPU 加速非平衡部分传输方法，覆盖点云、图像生成与图像到图像转换。
+task_description: 该基准覆盖 PartialTSW（部分树切片 Wasserstein），将 Tree-Sliced Wasserstein 扩展到非平衡/部分传输设置，并给出适用于生成建模中动态支撑分布的闭式形式。任务要求选择具体实验族（点云梯度流、JAX/diffusers 图像生成、或图像到图像转换）并运行对应目标或消融。评估聚焦收敛质量（relative_bias_pct）、运行性能（mean_time_s）与稳定性（coefficient_of_variation），而非内部优化器精度目标。代码锚点包括三类可执行脚本：收敛分析、按样本数的运行时基准、按维度的运行时基准。
+task_mode: experiment_driven
+requires_execution: true
+requires_paper: true
+integrity_level: cas_plus_canary
+snapshot_status: external_eval_required
+support_level: recovery
+time_band: 2-4d
+cost_band: high
+difficulty: hard
+data_access: public
+primary_outputs:
+- relative_bias_pct
+- mean_time_s
+- coefficient_of_variation
+launch_profiles:
+- id: convergence_eval
+  label: Convergence Evaluation
+  description: 'Run `analysis/convergence/simple_convergence.py` to assess estimator bias and stability across varying tree counts (L) and lines per tree (k). Uses PyTorch with GPU acceleration. Outputs relative bias percentage and coefficient of variation per configuration.
+
+    '
+- id: runtime_n_scaling
+  label: Runtime Scaling (Sample Count)
+  description: 'Run `analysis/runtime_plot/runtime_plot_compare_n_partial.py` to benchmark runtime and GPU memory across sample sizes N ∈ [10², 10⁵]. Compares PartialTSW against UOT baselines (sopt, spot, pawl, sliced_unbalanced_ot). Includes warm-up runs for stable timing.
+
+    '
+- id: runtime_d_scaling
+  label: Runtime Scaling (Dimension)
+  description: 'Run `analysis/runtime_plot/runtime_plot_compare_d_partial.py` to benchmark runtime and GPU memory across dimensions d ∈ [1, 2048] with fixed N=10⁴. Compares PartialTSW against UOT baselines. Includes warm-up runs for stable timing.
+
+    '
+- id: experiment_family
+  label: Experiment Family Workflow
+  description: 'Select and execute one packaged experiment family: point_cloud (gradient flow), image_gen (JAX/diffusers ablation sweep across multiple GPUs), or img2img (image-to-image translation with class imbalance handling). Each family has distinct data requirements and runtime profiles.
+
+    '
+dataset_download:
+  primary_method: mixed
+  sources:
+  - torchvision (auto-download for MNIST-based experiments)
+  - manual (pretrained checkpoints for image generation, assets for img2img)
+  notes:
+  - Convergence and runtime eval scripts use synthetic Gaussian data and require no external datasets.
+  - Image generation experiments may auto-download simple datasets through torchvision.
+  - Image-to-image translation and some ablation routes require manual asset preparation or pretrained checkpoints before evaluation.
+  - Treat dataset setup as experiment-family specific; there is no universal dataset route.
+credential_requirements:
+  mode: none
+  items: []
+  notes:
+  - No explicit API or dataset credential variables were detected in the shipped source snapshot.
+  - Image generation sweep uses explicit GPU slot scheduling (not cloud credential-based).
+resources:
+  minimum:
+    cpu_cores: 16
+    ram_gb: 64
+    disk_gb: 150
+    gpu_count: 1
+    gpu_vram_gb: 24
+    notes: GPU required for meaningful runtime benchmarks; CPU-only execution is impractical for N≥10⁴ or d≥64 configurations.
+  recommended:
+    cpu_cores: 32
+    ram_gb: 128
+    disk_gb: 300
+    gpu_count: 1
+    gpu_vram_gb: 48
+    notes: '48 GB VRAM accommodates full dimension sweep (up to d=2048) and larger sample counts. Image generation ablation grid benefits from additional GPU slots if available.
+
+      '
+environment:
+  python: '3.11'
+  cuda: '11.8'
+  pytorch: 2.1.0
+  flash_attn: null
+  key_packages:
+  - jax==0.4.23
+  - jaxlib==0.4.23
+  - transformers==4.51.3
+  - diffusers==0.33.1
+  - torch>=2.1.0
+  - ot (POT library for baseline comparisons)
+  - tqdm
+  - matplotlib
+  - pandas
+  - numpy
+  notes:
+  - The packaged repository mixes PyTorch (analysis scripts), JAX/jaxlib (image generation), and diffusion-style experiment routes rather than one narrow runtime profile.
+  - Analysis scripts default to CUDA with `torch.device('cuda' if torch.cuda.is_available() else 'cpu')`.
+  - Image generation sweep script (`experiments/image_gen/exp.sh`) schedules 72 jobs round-robin across allocated GPU slots; wall-clock time scales with GPU count.
+  - Baseline comparisons in runtime benchmarks import sopt, spot, pawl, sliced_unbalanced_ot, and unbalanced_sliced_ot from the bundled baselines module.
+risk_flags:
+- route_caveat
+- multi_route_experiment_family
+- mixed_runtime_environment
+- multi_gpu_scheduling_dependent
+risk_notes:
+- The packaged repository is structurally complete, but the optimizer target is not one root command; point-cloud, image-generation, image-to-image, and analysis routes differ materially in setup and execution.
+- The packaged `experiments/image_gen/exp.sh` schedules 72 jobs round-robin across the listed GPUs, so a full ablation sweep on one GPU is naturally multi-day.
+- The runtime environment mixes PyTorch (convergence, runtime analysis), JAX/jaxlib (image generation), and diffusion libraries; ensure both CUDA toolchains are compatible.
+- Some subroutes can auto-download simple datasets such as MNIST, while others require manual assets or pretrained checkpoints before evaluation.
+- Convergence eval uses synthetic Gaussian data only; generative downstream quality must be evaluated separately via experiment family routes.
+recommended_when: 当你需要支持非平衡测度的几何最优传输基准、能够明确选择实验族、并具备可支撑多日消融的 GPU 资源时使用本基准。适合比较部分传输求解器在收敛偏差、运行时扩展性和类别不平衡生成任务中的表现。
+not_recommended_when: 如果你需要单命令执行、CPU 优先流程、无分支的紧凑数据集，或无需执行即可获得预计算指标，不建议使用本基准。在无 CUDA GPU 的环境下也不适用。
+paper:
+  title: Tree-Sliced Entropy Partial Transport
+  venue: NeurIPS 2025
+  year: 2025
+  url: https://openreview.net/forum?id=41ZbysfW4h
+  authors:
+  - Viet-Hoang Tran (National University of Singapore)
+  - Thanh Tran (VinUniversity)
+  - Thanh Chu (National University of Singapore)
+  - Tam Le (Institute of Statistical Mathematics)
+  - Tan M. Nguyen (National University of Singapore)
+download:
+  url: https://github.com/ResearAI/DeepScientist/releases/download/aisb-v0.0.1/aisb.t3.075_treeslicedentropy.zip
+  archive_type: zip
+  local_dir_name: paper-75-TreeSlicedEntropy
+  provider: github_release
+  repo: ResearAI/DeepScientist
+  tag: aisb-v0.0.1
+  asset_name: aisb.t3.075_treeslicedentropy.zip
+  sha256: c73a8da6a3b04ef4a24eb1345993f985e84234db16636906626f07cb762fa33c
+  size_bytes: 774639
+display:
+  palette_seed: jade-bone-transport
+  art_style: geometric-minimal
+  accent_priority: high
+image_path: ../../../AISB/image/075_aisb.t3.075_treeslicedentropy.jpg
+capability_tags:
+- optimal_transport
+- tree_sliced_wasserstein
+- partial_transport
+- unbalanced_measures
+- generative_modeling
+- probabilistic_geometry
+- scientific_ml
+- research_code_optimization
+aisb_direction: T3
+track_fit:
+- paper_track
+- benchmark_track
+commercial:
+  annual_fee: null