Remove historical implementation from active tree

AGENTS.md

@@ -31,13 +31,12 @@ docs/migration.md
 
 and the relevant existing source and tests.
 
-The `legacy/` tree is a behavioral reference for the rewrite.
+The `root-gnn-parity-baseline` tag and committed reference fixtures preserve
+the behavioral baseline for the rewrite. Historical implementation code is
+not part of the active source tree.
 
-Unless explicitly instructed otherwise:
-
-* do not modify legacy code
-* do not reorganize legacy code
-* do not mechanically copy legacy architecture into the new package
+Do not reintroduce historical implementation code or imports into the active
+package. Preserve compatibility through explicit fixtures and adapters.
 
 When legacy behavior and documentation disagree, identify the discrepancy rather than silently choosing one.
 
@@ -195,7 +194,7 @@ Do not attempt to rewrite the entire legacy repository in one task.
 
 For substantial migrations:
 
-1. inspect the relevant legacy implementation
+1. inspect the relevant reference fixture and compatibility contract
 2. identify externally observable behavior
 3. inspect existing characterization/parity tests
 4. state or infer the intended new interface

README_PROJECT.md

@@ -14,8 +14,10 @@ ROOT files -> EventSample -> shared collider features
 ```
 
 The new implementation lives under [`src/gnn4colliders`](src/gnn4colliders/).
-[`legacy/`](legacy/) is a frozen behavioral reference for parity work and
-historical checkpoint investigation, not a supported runtime backend.
+Historical behavior is preserved by the
+[`root-gnn-parity-baseline`](https://huggingface.co/HWresearch/GNN4Colliders/tree/root-gnn-parity-baseline)
+tag and committed reference fixtures, not by a supported historical runtime
+backend.
 
 ## Installation
 
@@ -31,13 +33,13 @@ uv sync --dev --extra root-gnn
 ```
 
 The core package can be installed without DGL when only shared data or task
-code is needed. ROOT-GNN models, graph construction, and ROOT-GNN parity tests
+code is needed. ROOT-GNN models, graph construction, and ROOT-GNN reference tests
 require the `root-gnn` extra. On Linux x86_64, it uses the validated CUDA 12.1
 wheels configured in `pyproject.toml`; a compatible NVIDIA driver is still
 required. On Apple Silicon macOS, it installs the CPU DGL wheel, supporting
 local graph/cache development. The default ROOT-GNN backend performs training
 with native PyTorch graph tensors, so it runs on Apple MPS, NVIDIA CUDA, and
-CPU; DGL remains a cache and legacy-compatibility adapter. Do not add
+CPU; DGL remains a cache and graph compatibility adapter. Do not add
 site-specific CUDA, Slurm, or filesystem paths to model or task configuration.
 
 Use the MPS profile on an Apple Silicon Mac:

docs/agent-workflows/gnn4colliders-workflow.md

@@ -15,8 +15,9 @@ docs/migration.md
 
 Production Python belongs under `src/gnn4colliders/`. Keep collider physics
 features in `features/`, graph topology in `graphs/`, architecture code in
-`models/`, and lifecycle code in `training/`. Do not modify `legacy/`; use it
-as a behavioral reference for parity work.
+`models/`, and lifecycle code in `training/`. Use the frozen parity fixtures
+and compatibility adapters for historical behavior; do not add historical
+implementation imports.
 
 ## Environment
 

docs/architecture.md

@@ -1,378 +1,86 @@
 # GNN4Colliders architecture
 
-## Current v1 architecture
+## Current architecture
 
-The supported rewrite is layered around an architecture-neutral event boundary:
+GNN4Colliders is organized around an architecture-neutral event boundary:
 
 ```text
 ROOT/Awkward
     ↓
 EventSample + EventMetadata
     ↓
-shared collider feature construction
-    ├── GraphSample -> versioned graph cache -> GraphBatch
-    └── future SequenceSample -> ROOT-Transformer (not implemented)
-                                      ↓
-                               ROOT-GNN EdgeNetwork
-                                      ↓
-                               raw logits -> Task
-                                  ├── loss
-                                  ├── predictions
-                                  └── full-split metrics
-                                      ↓
-                          Trainer / Predictor / outputs
+shared features and representation adapters
+    ├── GraphSample -> GraphSampleCache -> GraphBatch
+    └── future SequenceSample -> transformer/token models
+    ↓
+model family
+    ↓
+Task -> Trainer/Predictor -> named outputs
 ```
 
 | Layer | Responsibility |
-| --- | --- |
-| `data` | ROOT/Awkward ingestion, event samples, metadata, graph caches, folds, and batching |
-| `features` | Shared collider-object features and derived physics quantities |
-| `graphs` | Topology, edge features, and the DGL representation adapter |
-| `models/root_gnn` | ROOT-GNN encoders, portable tensor message passing, classifier, and transfer boundary |
-| `tasks` | Loss, score/prediction, labels, weights, and metrics |
-| `training` | Optimizer lifecycle, validation, early stopping, checkpointing, and reproducibility |
-| `inference` | Ordered prediction/evaluation and NPZ/ROOT output adapters |
-| `distributed` | Rank-local devices, sharding, DDP, and cross-rank collection |
-| `config` / `cli` | Semantic Hydra composition and thin user-facing commands |
-
-`EventSample` is shared infrastructure, not a ROOT-GNN object. `GraphSample`
-is the current representation-specific adapter. At batching, the default
-ROOT-GNN backend converts cached DGL graphs into `TensorGraph` values with
-explicit node/edge indices and graph membership. Native message passing then
-uses only PyTorch tensor operations, allowing the same model to train on CPU,
-CUDA, and Apple MPS. This separation is the extension point for a future
-sequence/token representation.
-
-The new public metadata contract is named `EventMetadata(fold, weight,
-sample_id, extra)`. The legacy positional tracking tensor is accepted only by
-compatibility-facing ingestion code. A `GraphSampleCache` is deliberately a
-Level-2 graph cache; replacing it with a universal cache would couple future
-model families to DGL.
-
-For deployment, a prepared `GraphBatch` can pass through the isolated
-`RootGNNExportAdapter` into an ONNX model. ONNX is still an inference boundary:
-it does not read ROOT or construct collider features.
-
-## Current public workflow
-
-`prepare` reads ROOT through `RootEventDataset`, builds shared features and
-DGL graphs, and saves a schema-checked cache. `train` creates a model/task and
-`Trainer`; validation is the model-selection split and test is held out.
-`evaluate` computes metrics after collecting the complete split. `predict`
-returns detached CPU tensors in loader order and writes named NPZ fields.
-`write_root_scores` is an optional Python adapter with explicit entry alignment;
-the CLI currently exposes NPZ output.
-
-Compatibility responsibilities are isolated in `gnn4colliders.compat`.
-Supported historical checkpoint prefixes, classifier names, and the two-column
-tracking conversion are listed in [`compatibility.md`](compatibility.md).
-The modern pipeline does not propagate positional tracking or historical NPZ
-fields.
-
-Checkpoints are independent of the model implementation: they carry model and
-task metadata, lifecycle state, schema versions, and optional RNG state.
-Prefix normalization supports DDP `module.` and compiled `_orig_mod.` weights,
-plus the active ROOT-GNN historical classifier-name compatibility path.
-
----
-
-## Historical behavioral reference
-
-This document covers the target system in `legacy/root_gnn_dgl/`. The sibling
-`legacy/physicsnemo/` tree is a prior rewrite attempt and is not a behavioral
-target.
-
-## 1. High-level system description
-
-The active rewrite exposes `gnn4colliders.models.root_gnn.EdgeNetwork`. Its
-encoders and message-passing blocks form a reusable backbone whose decoded
-graph representation is passed to an explicit classifier. `FineTunedEdgeNetwork`
-reuses that backbone and replaces only the task-specific classifier, with
-explicit frozen or trainable-backbone control.
-
-`root_gnn_dgl` is a ROOT-to-DGL graph classification system. YAML selects
-dataset, model, loss, and finish-function classes by import path. The dataset
-reads ROOT trees, converts collider objects to fully connected DGL graphs, and
-saves graph chunks. Training loads those chunks, applies fold selection and
-optional pre-batching/padding, trains a graph network, writes one PyTorch
-checkpoint per epoch, and reports weighted loss, accuracy, and ROC AUC.
-
-The primary model is `models.GCN.Edge_Network` ([`GCN.py:182-251`](../legacy/root_gnn_dgl/models/GCN.py)).
-It encodes node, edge, and global features, repeats edge -> node -> global
-message passing `n_proc_steps` times, decodes the global state, and applies
-`classify`. Fine-tuning uses `models.GCN.Transferred_Learning_Finetuning`
-([`GCN.py:884-997`](../legacy/root_gnn_dgl/models/GCN.py)), which loads a
-pretrained `Edge_Network`, removes its final classifier, and applies a new one.
-The active configs use output size 12 for multiclass pretraining and output
-size 1 for binary tasks ([`configs/stats_100K/pretraining_multiclass.yaml:1-45`](../legacy/root_gnn_dgl/configs/stats_100K/pretraining_multiclass.yaml),
-[`configs/stats_100K/finetuning_ttH_CP_even_vs_odd.yaml:1-45`](../legacy/root_gnn_dgl/configs/stats_100K/finetuning_ttH_CP_even_vs_odd.yaml)).
-
-### Entry points and flows
-
-- `scripts/training_script.py:main` and its CLI parser load YAML, create
-  loaders, construct the model, and call `train`; `--evaluate` calls
-  `evaluate` ([`training_script.py:638-843`](../legacy/root_gnn_dgl/scripts/training_script.py)).
-- `scripts/prep_data.py:main` creates configured graph caches
-  ([`prep_data.py:68-110`](../legacy/root_gnn_dgl/scripts/prep_data.py)).
-- `scripts/inference.py:main` reconstructs an unlazy dataset, loads one or
-  more checkpoints, and writes `.npz` or ROOT scores
-  ([`inference.py:163-387`](../legacy/root_gnn_dgl/scripts/inference.py)).
-- `scripts/export_onnx.py:main` exports an ONNX-friendly model
-  ([`export_onnx.py:979-1035`](../legacy/root_gnn_dgl/scripts/export_onnx.py)).
-- `selections.py:main`, `check_dataset_files.py:main`, and
-  `plot_config_distributions.py:main` are diagnostic entry points. `run_demo.sh`
-  sequences pretraining, binary training, fine-tuning, and inference
-  ([`run_demo.sh:3-59`](../legacy/root_gnn_dgl/run_demo.sh)).
-
-The training flow is:
-
-```text
-YAML -> load_config/buildFromConfig -> RootDataset/LazyDataset
-     -> ROOT/Awkward -> DGL graph + labels/tracking/globals -> .bin cache
-     -> fold_selection -> prebatch/padding -> GraphDataLoader
-     -> Edge_Network or transfer model -> weighted loss/metrics
-     -> model_epoch_N.pt, logs, evaluation/inference output
-```
-
-`training_script.train` is the lifecycle implementation
-([`training_script.py:143-614`](../legacy/root_gnn_dgl/scripts/training_script.py));
-distributed paths use NCCL/DDP ([`training_script.py:616-839`](../legacy/root_gnn_dgl/scripts/training_script.py)).
-Inference uses `CustomPreBatchedDataset`, applies a configured finish function,
-and collects `scores`, `labels`, and `tracking_info`
-([`inference.py:20-76`](../legacy/root_gnn_dgl/scripts/inference.py),
-[`inference.py:223-325`](../legacy/root_gnn_dgl/scripts/inference.py)).
-
-## 2. Dependency and data-flow map
+|---|---|
+| `data` | ROOT/Awkward ingestion, metadata, datasets, folds, batching, and caches |
+| `features` | Collider-object features and derived physics quantities |
+| `graphs` | Topology, edge construction, and graph-specific adapters |
+| `models` | Architecture-specific neural networks |
+| `tasks` | Labels, weights, losses, scores, predictions, and metrics |
+| `training` | Optimizers, lifecycle, reproducibility, checkpointing, and distributed utilities |
+| `inference` | Ordered prediction, evaluation, and named output writing |
+| `cli` / `config` | Thin semantic entry points and configuration composition |
+
+Production code must depend on these package boundaries rather than on
+historical implementation paths. The canonical ROOT-GNN implementation is
+`gnn4colliders.models.root_gnn.EdgeNetwork` with
+`FineTunedEdgeNetwork` as its transfer boundary. New architecture families
+must reuse shared data, feature, task, training, and inference interfaces where
+their representation permits it.
+
+## Data and representation boundaries
+
+`EventSample` and named `EventMetadata(fold, weight, sample_id, extra)` are
+representation-independent. `GraphSample` and `GraphBatch` are the current
+graph representation boundary. `TensorGraph` provides a native tensor path
+for ROOT-GNN execution on CPU, CUDA, and MPS; DGL remains an optional graph
+adapter and cache dependency.
+
+The active collider node schema has seven columns:
+`[pt, eta, phi, energy, btag, charge, node_type]`. Graph edges are directed,
+source-major, fully connected without self-loops except for the one-node
+case, and carry `[deta, dphi, dR]` features. These are compatibility contracts
+captured by deterministic tests and the frozen `root-gnn-parity-baseline`
+reference fixture.
+
+## Configuration and lifecycle
+
+Experiments use semantic configuration such as `model.type: root_gnn`; model
+module paths are not part of the new public configuration contract. The CLI
+delegates to tested application factories and does not contain model or data
+processing logic.
+
+Checkpoints carry model/task metadata, lifecycle state, schema versions, and
+optional RNG state. The compatibility package accepts historical checkpoint
+prefixes and classifier names as a one-way input adapter. No executable
+historical model code is required at runtime.
+
+## Extending the model families
+
+The next model family should introduce only its representation-specific
+boundary and model implementation, for example:
 
 ```text
-ROOT files (raw_dir/file_names, tree_name)
-  -> gnn4colliders.data.RootEventDataset (Uproot/Awkward)
-  -> node feature construction (dataset.py:15-50)
-  -> full_connected_graph (dataset.py:52-59)
-  -> EdgeDataset.make_graph: [deta, dphi, dR] (dataset.py:471-482)
-  -> DGL .bin cache / LazyDataset / PreBatchedDataset
-  -> GraphDataLoader -> models.GCN -> loss/metrics -> outputs
+data.EventSample -> features -> SequenceSample -> models.root_transformer
 ```
 
-`RootDataset` provides `process`, `save`, `load`, `__getitem__`, and `__len__`
-([`dataset.py:160-469`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py));
-`LazyDataset` loads one chunk through a ring buffer
-([`dataset.py:525-578`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py));
-`PreBatchedDataset.process` selects, shuffles, batches, pads, and caches
-([`batched_dataset.py:34-146`](../legacy/root_gnn_dgl/root_gnn_base/batched_dataset.py)).
-
-`load_config` uses PyYAML `FullLoader` and shallow `include` merging, while
-`buildFromConfig` dynamically imports `module`, resolves `class`, merges extra
-keys into `args`, converts list-valued weights to tensors, and injects runtime
-arguments ([`utils.py:10-43`](../legacy/root_gnn_dgl/root_gnn_base/utils.py)).
-This reflection shape is a de facto interface for configured components.
-
-### Data and preprocessing
-
-The active node schema is seven columns: `pt`, `eta`, `phi`, `energy`, `btag`,
-`charge`, and `node_type`. `CALC_E` is `pt*cosh(eta)`, constants are broadcast
-per object type, `NODE_TYPE` is an integer type code, and feature scales are
-applied columnwise ([`dataset.py:15-50`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py)).
-`full_connected_graph` makes directed all-pairs edges; `EdgeDataset` requests
-no self-loops and stores `[deta, dphi, dR]`
-([`dataset.py:52-59`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py),
-[`dataset.py:471-482`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py)).
-
-Selections are strings evaluated with builtins disabled or
-`(variable, cut, operator)` triples (`check_selection`, `selection_mask`;
-[`dataset.py:75-145`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py)). Fold
-selection uses `tracking[:,0] % n_folds`; tracking column 0 is fold and column 1
-is weight ([`utils.py:121-143`](../legacy/root_gnn_dgl/root_gnn_base/utils.py),
-[`dataset.py:176-182`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py)).
-`hash_partition` and a seeded Torch generator control pre-batch order
-([`batched_dataset.py:27-99`](../legacy/root_gnn_dgl/root_gnn_base/batched_dataset.py)).
-Padding modes are `NONE`, `STEPS`, `FIXED`, and `NODE`; `FIXED` is hardcoded to
-16,000 nodes and 104,000 edges ([`batched_dataset.py:100-125`](../legacy/root_gnn_dgl/root_gnn_base/batched_dataset.py)).
-
-### Model, losses, and metrics
-
-`Make_MLP` builds linear/ReLU/dropout blocks followed by LayerNorm
-([`GCN.py:18-35`](../legacy/root_gnn_dgl/models/GCN.py)). Each `Edge_Network`
-step encodes inputs, copies source/destination states to edges, updates edges,
-sums edge messages into nodes, updates nodes, then mean-pools nodes/edges to
-update globals ([`GCN.py:195-249`](../legacy/root_gnn_dgl/models/GCN.py)). It
-returns logits `[graphs, out_size]` without sigmoid/softmax.
-
-The default objective is elementwise `BCEWithLogitsLoss`, multiplied by
-`tracking[:,1]`, averaged separately per unique label, then averaged across
-labels ([`training_script.py:143-185`](../legacy/root_gnn_dgl/scripts/training_script.py),
-[`training_script.py:320-359`](../legacy/root_gnn_dgl/scripts/training_script.py)).
-`--abs` makes weights positive. Binary metrics use sigmoid threshold 0.5 and
-weighted ROC AUC; multiclass metrics use argmax and one-vs-rest ROC AUC
-([`training_script.py:438-510`](../legacy/root_gnn_dgl/scripts/training_script.py)).
-Additional configurable losses/finishers live in `models/loss.py`
-([`loss.py:6-310`](../legacy/root_gnn_dgl/models/loss.py)).
-
-### Checkpoints and outputs
-
-Training writes `Training_Directory/model_epoch_<epoch>.pt` containing `epoch`,
-`model_state_dict`, `optimizer_state_dict`, and serialized `early_stop`
-([`training_script.py:565-604`](../legacy/root_gnn_dgl/scripts/training_script.py)).
-Keys strip `module.` and compiled models save the underlying `_orig_mod` state;
-`get_last_epoch`, `get_specific_epoch`, and `get_best_epoch` load the files
-([`utils.py:145-248`](../legacy/root_gnn_dgl/root_gnn_base/utils.py)).
-`evaluate` writes `evaluation_<epoch>.npz`; inference writes `.npz` fields
-`scores`, `labels`, `tracking_info`, or adds score branches and `selection_pass`
-to a cloned ROOT tree ([`training_script.py:57-140`](../legacy/root_gnn_dgl/scripts/training_script.py),
-[`inference.py:328-385`](../legacy/root_gnn_dgl/scripts/inference.py)).
-
-### Training lifecycle boundary
-
-The active rewrite keeps lifecycle orchestration architecture-independent:
-
-```text
-GraphDataLoader -> GraphBatch -> Model -> Task -> Trainer
-                                      loss/metrics
-```
-
-`gnn4colliders.training.Trainer` owns device placement, train/evaluation mode,
-gradient and optimizer steps, epoch aggregation, optional scheduler stepping,
-early stopping, and in-memory history. Tasks own loss and metric semantics;
-the trainer does not inspect positional tracking columns or collider-specific
-features. Evaluation concatenates detached outputs across the complete split
-before calling task metrics, so ROC AUC is not computed per mini-batch.
-
-Checkpoint persistence and the Python inference/output layer are implemented
-as separate adapters. The semantic CLI and distributed application boundary
-are implemented in the current stack. `gnn4colliders.inference.Predictor` accumulates detached CPU
-logits, task-defined scores/predictions, labels, and named event metadata in
-loader order; `write_npz` is the primary named-field format and ROOT score
-writing is an optional alignment-aware adapter.
-
-The new lifecycle uses conventional split semantics. `train` updates model
-parameters, `validation` is evaluated after every epoch and drives scheduler,
-early-stopping, and later model selection, and `test` is held out. The trainer
-does not accept a test loader in `fit`; callers evaluate the held-out test set
-separately after training. This deliberately corrects the legacy convention
-where a loader named `test` was used for model selection and `val` represented
-held-out testing.
-
-## Distributed execution
-
-`gnn4colliders.distributed` contains the small DDP boundary used by the
-application layer. `DistributedContext` reads the standard `torchrun`
-environment (`RANK`, `LOCAL_RANK`, and `WORLD_SIZE`), selects the rank-local
-device, and owns process-group cleanup. Graph samples are sharded before
-batching; training may pad rank shards for equal step counts, while validation
-and prediction use unpadded shards so events are not counted twice.
-
-The configured graph `batch_size` is per process. DDP wraps an otherwise
-ordinary model after device placement, and checkpoint state is normalized to
-the underlying model keys. Loss gradients are synchronized by DDP; epoch
-metrics and evaluation outputs are gathered across ranks. Rank 0 writes
-resolved configuration, checkpoints, and NPZ predictions. Moderate-size
-prediction gathering is in-memory; streaming/sharded output is a future
-extension.
-
-
-## Randomness, external services, and coupling
-
-The CLI exposes `--seed`, but `main` passes it to model construction rather
-than globally seeding Python, NumPy, or Torch
-([`training_script.py:638-753`](../legacy/root_gnn_dgl/scripts/training_script.py)).
-Pre-batching has explicit seeds, but `AugmentedDataset` mutates the process-wide
-NumPy seed ([`dataset.py:716-827`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py)),
-clustering uses unseeded `torch.randperm`/`randint` (`loss.py:259-295`), and
-model reset/fine-tuning hardcodes `torch.manual_seed(2)`
-([`GCN.py:58-65`](../legacy/root_gnn_dgl/models/GCN.py),
-[`GCN.py:900-915`](../legacy/root_gnn_dgl/models/GCN.py)). CUDA kernels, DDP,
-and DataLoader behavior are not made deterministic.
-
-Implicit coupling includes repository-relative `sys.path` insertion
-([`training_script.py:14-20`](../legacy/root_gnn_dgl/scripts/training_script.py)),
-dynamic imports, mutable default lists/dicts, global `FEATURE_DTYPE`
-([`dataset.py:13`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py)),
-in-place `tracking_info` mutation ([`dataset.py:176-181`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py)),
-and in-place DGL graph mutation during forward.
-
-The legacy environment assumes Python 3.8, PyTorch 2.0.1, CUDA 11.8, DGL
-1.1.1, ROOT, Awkward, Uproot, PyYAML, and scikit-learn
-([`setup/environment.yml:1-10`](../legacy/root_gnn_dgl/setup/environment.yml),
-[`setup/environment.yml:240-295`](../legacy/root_gnn_dgl/setup/environment.yml)).
-The active Linux development environment is intentionally separate: Python
-3.12, PyTorch 2.2.2/CUDA 12.1, and DGL 2.4.0 from the official DGL wheel
-repository. CUDA runtime wheels do not replace the compatible host NVIDIA
-driver and do not encode Perlmutter module settings.
-Standard configs assume `/global/cfs/` and `/pscratch/` paths, CUDA/NCCL,
-Slurm, and optionally Podman-HPC. `setup/download_data.sh` downloads the
-external Hugging Face dataset `HWresearch/Delphes`
-([`download_data.sh:13-67`](../legacy/root_gnn_dgl/setup/download_data.sh)).
-
-## Apparent unused or secondary code
-
-Not selected by the standard stats/Delphes configs, or only reachable from
-optional workflows, are `GCN_global`, `GCN_global_2way`, most transfer variants,
-attention models, `MultiModel`, and `Clustering`
-([`GCN.py:122-1933`](../legacy/root_gnn_dgl/models/GCN.py)); `UprootDataset`,
-`tHbbEdgeDataset`, `AugmentedDataset`, and photon-ID paths
-([`uproot_dataset.py:10-31`](../legacy/root_gnn_dgl/root_gnn_base/uproot_dataset.py),
-[`dataset.py:484-827`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py),
-[`photon_ID_dataset.py:1-33`](../legacy/root_gnn_dgl/root_gnn_base/photon_ID_dataset.py));
-optional loss and similarity utilities; and the no-op
-`root_gnn_base.utils.graph_augmentation` ([`utils.py:393-395`](../legacy/root_gnn_dgl/root_gnn_base/utils.py)).
-The main path evaluates `test_loaders`; validation loaders are only assembled
-when a config has a validation fold ([`training_script.py:682-747`](../legacy/root_gnn_dgl/scripts/training_script.py)).
-
-## 3. De facto interfaces to preserve
-
-### Metadata-aware dataset boundary
-
-The rewrite uses named `EventMetadata` (`fold`, `weight`, and stable
-`sample_id`) instead of exposing the legacy positional tracking tensor.
-`GraphSample`, `GraphBatch`, `SplitDefinition`, and `GraphDataLoader` form the
-ROOT-GNN orchestration boundary. Graph caches carry feature, graph, and cache
-schema versions and reject incompatible artifacts before loading.
-
-The current cache implementation stores processed `GraphSample` values (the
-Level-2 cache). The separation from `RootEventDataset` is intentional: a
-future Level-1 cache can store normalized `EventSample`/feature data for
-sequence or transformer representations without requiring DGL graph caches.
-
-The rewrite's shared data boundary is `gnn4colliders.data`: it reads selected
-ROOT/Awkward branches and returns architecture-neutral event samples. Feature
-construction and graph building remain separate downstream boundaries, so the
-same samples can be reused by non-graph model families.
-
-### Configuration and CLI boundary
-
-Hydra composes semantic YAML groups under `configs/` and passes the resolved
-configuration to explicit application factories in `gnn4colliders.config`.
-Those factories allow-list supported models, tasks, and trainer components;
-YAML is never treated as an arbitrary Python import specification. The thin
-`gnn4colliders` CLI selects `prepare`, `train`, `evaluate`, or `predict` and
-delegates to the stable data, training, checkpoint, and inference APIs. A new
-experiment should generally be a YAML change; new behavior belongs in Python.
-
-1. YAML `module`, `class`, `args`, plus runtime `sample_graph` and
-   `sample_global` injection.
-2. Dataset items `(DGLGraph, label, tracking, global_features)`.
-3. `ndata['features']`, `edata['features']`, seven node columns, and three edge
-   columns in `[deta, dphi, dR]` order.
-4. Tracking column 0 fold and column 1 weight semantics.
-5. `model(graph, global_feats)`, logits shape `[batch, out_size]`, and
-   `representation` where used.
-6. Weighted per-label loss, metric thresholds, checkpoint keys/prefix cleanup,
-   epoch filenames, and `.npz`/ROOT output fields.
-
-## 4. Ambiguous behavior
-
-- Historical edge order/self-loop expectations; empty and padding graph inputs.
-- Whether negative weights are meaningful or should always be absolute.
-- Whether “validation” is intended to differ from the active test-loader path.
-- Shape semantics of multi-label finishers and experimental transfer classes.
-- Whether chunk IDs must match historical `np.array_split` boundaries.
-- Required behavior for missing branches and dynamic selection expressions.
+It should include a deterministic fixture, unit tests for the representation,
+an integration path through the shared task/trainer interfaces, and explicit
+checkpoint/inference behavior. Shared infrastructure should be generalized
+only when the second model demonstrates a real common use case.
 
-## 5. Recommended rewrite boundaries
+## Frozen baseline
 
-Separate typed configuration; ROOT/Awkward I/O; selections/folds/features/
-edges; DGL cache/lazy loading/batching; active models and checkpoint adapters;
-objectives/metrics; training lifecycle; and inference/ONNX applications.
-Establish parity for the active `LazyDataset -> PreBatchedDataset ->
-Edge_Network` binary/multiclass path first. Add experimental classes only when
-a config or consumer proves they are required.
+The complete ROOT-GNN parity campaign is recorded by the
+`root-gnn-parity-baseline` tag. The historical implementation is no longer
+part of the active source tree. Reference data and compatibility adapters are
+kept so existing checkpoints and scientific observations remain usable while
+development moves to new architectures.

docs/compatibility.md

@@ -18,5 +18,6 @@ adapted into the new representation and are never rewritten implicitly.
 | Historical `tracking_info` NPZ output | no | — | No active consumer remains; positional output is intentionally unsupported |
 | Legacy YAML `module`/`class`/`args` | compatibility only | configuration boundary | Accepted only where the semantic factory can safely interpret it; new configs use semantic model names |
 
-The frozen `legacy/` tree remains available to parity tests and historical
-investigation. Production modules do not import executable code from it.
+The frozen parity fixtures and `root-gnn-parity-baseline` tag preserve the
+historical observations for parity tests and investigation. Production modules
+do not import executable historical model code.

docs/end_to_end_validation.md

@@ -1,4 +1,4 @@
-# End-to-end legacy/rewrite validation
+# End-to-end reference validation
 
 Task 21 compares staged event identity, labels, folds, weights, globals, node
 features, topology, edge features, batching, fixed-weight forward, loss and

docs/migration.md

@@ -1,312 +1,49 @@
-# Incremental migration plan: `root_gnn_dgl`
-
-The target is `legacy/root_gnn_dgl/`. `legacy/physicsnemo/` is a prior rewrite
-attempt and may inspire abstractions, but it is not a parity target. Neither
-legacy tree should be modified during migration.
-
-Each phase should add focused unit tests, a deterministic fixture in
-`data/fixtures/`, and parity tests under `tests/parity/` before moving upward.
-Record intentional differences and checkpoint consequences here.
+# Migration and model-family roadmap
 
 ## Frozen ROOT-GNN baseline
 
-The active ROOT-GNN rewrite and the legacy implementation are frozen at the
-`root-gnn-parity-baseline` tag. The legacy tree is now a read-only behavioral
-reference; new model development must not add production dependencies on it.
-The completed no-selection ttH CP-even versus CP-odd campaign covered full
-event preprocessing and graph parity, binary losses and metrics, one-step and
-multi-epoch fine-tuning, full-split training, checkpoint reload and resume,
-reproducibility, and chunked legacy `.bin` serialization.
-
-The next migration boundary is to replace live legacy imports in validation
-workflows with frozen reference fixtures. Until that boundary is complete,
-`legacy/` remains in the repository so the strict parity gate and historical
-checkpoint investigations remain reproducible. New architecture families may
-reuse shared data, features, tasks, training, and inference interfaces without
-depending on ROOT-GNN or the legacy tree.
-
-## Phase 0 — freeze observations and fixtures
-
-Capture a small representative ROOT-equivalent fixture containing the seven
-active node features, three edge features, labels, fold values, weights, and
-globals. Record outputs of `node_features_from_tree`, `full_connected_graph`,
-`EdgeDataset.make_graph`, and `fold_selection`
-([`dataset.py:15-59`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py),
-[`dataset.py:471-482`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py),
-[`utils.py:121-143`](../legacy/root_gnn_dgl/root_gnn_base/utils.py)). Preserve
-one `.bin`, one `model_epoch_N.pt`, one evaluation `.npz`, and one inference
-`.npz` fixture if available.
-
-Task 3 characterization records the active-path observations. Node rows are
-concatenated by object type in the configured order (jets, electrons, muons,
-photons, MET), and the seven columns are `[pt, eta, phi, energy, btag, charge,
-node_type]`. `CALC_E` is `pt*cosh(eta)` before the configured column scale is
-applied. The graph is directed and uses all ordered pairs except self-loops
-for graphs with more than one node; edge order is source-major. A one-node
-graph is a special case: the no-self-loop branch retains its sole self-loop.
-Edge columns are `[deta, dphi, dR]`, with `dphi` wrapped into `[-pi, pi]`.
-Dataset items expose `(graph, label, tracking, global_features)`; tracking
-column 0 is the fold identifier and column 1 is the event weight. These are
-compatibility observations, not proposed fixes.
-
-## Phase 1 — configuration boundary
-
-Implement a typed configuration layer that reads `Training`, `Model`,
-optional `Loss`, and `Datasets`. Initially retain a compatibility adapter for
-`module`/`class`/`args` and runtime injection of `sample_graph` and
-`sample_global`, matching `buildFromConfig`
-([`utils.py:10-43`](../legacy/root_gnn_dgl/root_gnn_base/utils.py)). Keep
-dynamic imports isolated at this boundary rather than spreading reflection
-through new code.
-
-## Phase 2 — pure preprocessing parity
-
-Port and test, in isolation:
-
-- branch-to-node conversion, `CALC_E`, `NODE_TYPE`, constants, scaling, empty
-  objects, and dtypes (`node_features_from_tree`,
-  [`dataset.py:15-50`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py));
-- string/tuple selections and cutflow (`check_selection`, `selection_mask`,
-  `compute_cutflow`, [`dataset.py:75-158`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py));
-- fold masks and cache suffixes (`fold_selection`, `fold_selection_name`,
-  [`utils.py:121-143`](../legacy/root_gnn_dgl/root_gnn_base/utils.py));
-- deterministic chunk partitioning (`hash_partition`,
-  [`batched_dataset.py:27-31`](../legacy/root_gnn_dgl/root_gnn_base/batched_dataset.py)).
-
-This is the highest-value parity layer: model parity is invalid if graph inputs
-differ.
-
-Task 4 implements the shared branch-to-node feature builder under
-`gnn4colliders.features`. It preserves the active seven-column schema,
-object-type ordering, explicit scales, derived `CALC_E`, node-type codes,
-float32 output, and supported empty vector collections. Selection, fold, and
-chunk helpers remain deferred to later data-infrastructure work.
-
-Task 6 implements the shared ROOT/Awkward ingestion boundary under
-`gnn4colliders.data`. `RootEventDataset` returns immutable, architecture-neutral
-`EventSample` values with selected branch data, labels, tracking, and globals;
-events are ordered by input file order with a global zero-based index. Fold
-filtering, caching, batching, and model-specific conversion remain deferred.
-
-## Phase 3 — graph construction and cache format
-
-Implement graph construction with tests for node/edge counts, directed edge
-ordering, self-loop policy, `[deta, dphi, dR]` order, metadata, and empty graphs.
-Preserve the dataset item contract `(graph, label, tracking, global_features)`
-from `RootDataset.__getitem__`
-([`dataset.py:465-469`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py)).
-
-Then implement DGL `.bin` serialization, lazy chunk loading, pre-batching, and
-padding. Compare against `RootDataset.save/load`, `LazyDataset`, and
-`PreBatchedDataset` ([`dataset.py:396-469`](../legacy/root_gnn_dgl/root_gnn_base/dataset.py),
-[`batched_dataset.py:129-174`](../legacy/root_gnn_dgl/root_gnn_base/batched_dataset.py)).
-Treat `NONE`, `STEPS`, `FIXED`, and `NODE` as explicit features; do not hide
-the hardcoded fixed padding sizes.
-
-Task 7 establishes the metadata-aware orchestration boundary around this
-phase: `EventMetadata`, `GraphSample`, `GraphBatch`, fold-based split
-selection, deterministic batching, and a version-checked graph-sample cache.
-The cache is deliberately Level 2; normalized event caching remains a future
-extension so non-graph model families can reuse ROOT preprocessing.
-
-## Phase 4 — active model parity
-
-Task 8 adds the active `EdgeNetwork` and `FineTunedEdgeNetwork` under
-`gnn4colliders.models.root_gnn`. The update order and MLP ordering follow the
-legacy active path. The rewrite uses an explicit backbone/classifier boundary,
-local DGL graph scope, and does not mutate global RNG state in constructors.
-Model parity now covers fixed-weight pretraining and transfer paths, including
-historical checkpoint prefixes. The legacy transfer implementation has an
-active bug when nonempty globals are supplied (`Pretrained_Output` ignores its
-argument); parity therefore characterizes its supported no-global path, while
-the rewritten model supports both global and fallback modes.
-
-Task 22 adds `TensorEdgeNetwork`, a native PyTorch realization of the same
-message-passing equations. It receives explicit node/edge index tensors from
-the batching boundary, retains parameter names for legacy checkpoint loading,
-and has fixed-weight CPU parity coverage against the DGL backend. It is the
-default training backend for CPU, CUDA, and Apple MPS; DGL remains available as
-a cache and compatibility adapter during the transition.
-
-Port `models.GCN.Edge_Network` first. Preserve constructor parameters,
-`forward(graph, global_feats)`, feature keys, processor order, MLP LayerNorm
-placement, and logits shape. Compare intermediate and final tensors on fixed
-graphs using the legacy architecture
-([`GCN.py:18-35`](../legacy/root_gnn_dgl/models/GCN.py),
-[`GCN.py:182-251`](../legacy/root_gnn_dgl/models/GCN.py)).
-
-Next port `Transferred_Learning_Finetuning`, including pretrained
-`model_state_dict` loading, removal of the final classifier, and new classifier
-initialization ([`GCN.py:884-997`](../legacy/root_gnn_dgl/models/GCN.py)). Test
-both frozen and unfrozen modes. Defer other model classes until an active
-config or consumer proves they are needed.
-
-## Phase 5 — objectives and metrics
-
-Implement the default objective exactly: elementwise configured loss,
-tracking-column weights, per-unique-label normalization, and averaging across
-labels ([`training_script.py:320-359`](../legacy/root_gnn_dgl/scripts/training_script.py)).
-Add parity cases for positive, zero, and negative weights and binary versus
-multiclass shapes.
-
-Port metric behavior from
-[`training_script.py:438-510`](../legacy/root_gnn_dgl/scripts/training_script.py):
-sigmoid threshold 0.5, argmax, weight masking, weighted ROC AUC, one-vs-rest
-multiclass AUC, and NaN behavior when AUC is undefined. Add `models/loss.py`
-classes only with dedicated tests; do not substitute their reductions.
-
-## Phase 6 — checkpoint and lifecycle
-
-Task 10 implemented the in-memory single-process training lifecycle before the
-checkpoint portion of this phase: `Trainer`, explicit optimizer/scheduler
-builders, `EarlyStopping`, reproducibility seeding, `GraphBatch.to`, and
-epoch/history result types. Checkpoint persistence/resume and the Python
-inference/evaluation and named NPZ/ROOT output layers are now implemented.
-Distributed execution and CLI wiring were completed in the later phases.
-
-Task 10 also establishes corrected split semantics: validation is evaluated
-every epoch and is the only split used for model selection or early stopping;
-the test split remains held out and is evaluated separately after fitting. The
-legacy loader naming inversion (`test` used for selection and `val` held out)
-is not carried into the rewrite.
-
-Create a checkpoint adapter preserving `model_epoch_<epoch>.pt` and keys
-`epoch`, `model_state_dict`, `optimizer_state_dict`, and `early_stop`
-([`training_script.py:565-604`](../legacy/root_gnn_dgl/scripts/training_script.py)).
-Support legacy DDP/compiled prefixes (`module.` and `_orig_mod.`) as exercised
-by checkpoint lookup and inference
-([`utils.py:145-248`](../legacy/root_gnn_dgl/root_gnn_base/utils.py),
-[`inference.py:274-290`](../legacy/root_gnn_dgl/scripts/inference.py)). Port
-`EarlyStop` state and log parsing separately
-([`utils.py:325-390`](../legacy/root_gnn_dgl/root_gnn_base/utils.py)). Verify
-resume, restart, early termination, and `.npz` fields before distributed work.
-
-## Phase 7 — CLI, inference, and export
-
-Task 12 implemented ordered prediction/evaluation, task-owned score
-semantics, checkpoint weight-only loading, named metadata retention, NPZ
-output, and explicit ROOT entry alignment. The semantic CLI and the validated
-ROOT-GNN ONNX export adapter are implemented.
-
-Task 13 adds Hydra composition and a single-process CLI around those existing
-APIs. The current application data boundary is a versioned
-`GraphSampleCache`; ROOT preparation converts events through the shared
-feature and graph builders before writing that cache.
-
-Build thin new applications around tested library interfaces in this order:
-
-1. preprocessing/cache generation (`scripts/prep_data.py`);
-2. training/evaluation (`scripts/training_script.py`);
-3. inference to `.npz` and ROOT (`scripts/inference.py`);
-4. ONNX export after PyTorch parity (`gnn4colliders export`).
-
-Use subprocess integration tests with tiny fixtures. Preserve CLI options only
-where they serve an active workflow; document removed diagnostic/cluster-only
-options.
-
-## Phase 8 — reproducibility and deployment
-
-Task 14 adds the initial deployment boundary: CPU/GPU DDP through standard
-`torchrun` variables, rank-local graph-sample sharding, global metric/output
-gathering, rank-0 checkpoint/config writing, and Perlmutter-oriented Slurm
-examples. Evaluation deliberately avoids sampler padding duplicates. The
-remaining follow-up is a streaming or sharded output path for very large
-distributed inference jobs.
-
-The seed policy remains explicit: the configured seed is offset by rank for
-process-local randomness, while distributed sample assignment is derived from
-the configured seed, world size, and epoch. GPU kernel nondeterminism and
-exact per-rank RNG checkpoint replay remain environment-dependent. Slurm/NCCL,
-Podman-HPC, ROOT, and Hugging Face integrations stay in launcher/adapters
-rather than package code.
-
-## Checkpoint compatibility checklist
-
-- [x] Load a checked-in or generated multiclass pretrained checkpoint.
-- [x] Load a legacy fine-tuning checkpoint after prefix normalization.
-- [x] Resume optimizer and early-stop state.
-- [x] Produce equivalent logits on a deterministic graph fixture.
-- [x] Produce equivalent `.npz` score, label, and metadata fields.
-- [x] Preserve ROOT scalar/vector score branch conventions in the Python adapter.
-
-Known risks are documented in [`architecture.md`](architecture.md): edge order,
-self-loops, weight semantics, validation/test naming, padding, dynamic
-selection evaluation, reproducibility, and the experimental model/loss surface.
-
-## Migration closure status
+The `root-gnn-parity-baseline` tag records the completed migration of the
+active ROOT-GNN behavior into `src/gnn4colliders.models.root_gnn`. The
+campaign covered full event preprocessing and graph parity, binary objectives
+and metrics, deterministic fine-tuning, full-split training, checkpoint reload
+and resume, reproducibility, and serialized graph-cache checks.
 
-### Task 18 compatibility closure
+The historical implementation is no longer in the active source tree. Its
+observable behavior is represented by committed fixtures, tests, and the
+one-way checkpoint/metadata compatibility adapters. New work must not add
+imports from historical implementation paths.
 
-The compatibility boundary is now explicit in `gnn4colliders.compat`.
-Production ingestion stores named `EventMetadata`; legacy two-column tracking
-is converted only at the compatibility boundary. Checkpoint prefix cleanup and
-the historical ROOT-GNN `classify` to `classifier` mapping have one canonical
-implementation. The new checkpoint schema and named NPZ output remain
-canonical. See [`compatibility.md`](compatibility.md) for the supported and
-intentionally unsupported historical artifacts.
+## Shared contracts
 
-The following matrix describes the supported new stack, rather than every
-class that exists in `legacy/`:
+New model families should consume these boundaries:
 
-| Legacy area | New-stack status | Notes |
-| --- | --- | --- |
-| ROOT/Awkward ingestion | migrated | `RootEventDataset` returns `EventSample` in file/event order |
-| node features | migrated + parity-tested | seven-column schema, `CALC_E`, ordering, scales, float32 |
-| edge construction | migrated + parity-tested | directed source-major topology and `[deta,dphi,dR]` |
-| graph cache | migrated | versioned `GraphSampleCache`; graph-level cache only |
-| folds and weights | migrated | named `EventMetadata.fold` and `.weight` |
-| batching | migrated | deterministic local loader and DDP sharding |
-| legacy padding modes | deferred | no active new-stack consumer |
-| `Edge_Network` | migrated + parity-tested | `EdgeNetwork`, raw logits |
-| transfer/fine-tuning | migrated + parity-tested | frozen or trainable backbone |
-| loss and metrics | migrated + parity-tested | task-owned weighted reductions and full-split AUC |
-| training lifecycle | migrated | `Trainer`, validation semantics, scheduler, early stopping |
-| checkpoints/resume | migrated | schema v1; historical weight/prefix adapter |
-| inference/NPZ | migrated | named output fields and ordered accumulation |
-| ROOT score output | compatibility adapter | Python API supported; CLI currently NPZ-only |
-| DDP | migrated | torchrun boundary, rank-0 artifacts, gathered metrics |
-| Slurm/Perlmutter | launcher examples | site policy remains outside package code |
-| ONNX export | migrated for ROOT-GNN | tensor-only adapter, ONNX Runtime validation, and `export` CLI; raw graph tensors are the input contract |
+- `EventSample` and named `EventMetadata` from `data`;
+- shared collider feature builders from `features`;
+- a representation-specific sample/batch type from the relevant adapter;
+- task-owned loss, score, prediction, and metric semantics;
+- the shared `Trainer`, checkpoint, reproducibility, and inference APIs.
 
-### Intentional redesigns
+The graph path is the current ROOT-GNN representation. A sequence or token
+model should add a separate representation boundary rather than placing
+sequence behavior in graph modules or generic data code.
 
-These are deliberate new-stack contracts, not accidental parity failures:
+## Next model-family milestone
 
-* `tracking[:, 0]` and `tracking[:, 1]` become named `metadata.fold` and
-  `metadata.weight`; public consumers do not depend on positional columns.
-* Dynamic legacy YAML `module`/`class` construction becomes allow-listed
-  semantic Hydra configuration.
-* The monolithic training script becomes `Task` + `Trainer` + checkpoint and
-  inference adapters.
-* Graph state is scoped to the forward pass rather than relying on persistent
-  mutation of shared graph state.
-* Model constructors do not mutate global RNG state; seeding is explicit in
-  the training/application boundary.
-* Validation is the selection/early-stopping split and test is held out. This
-  corrects the legacy loader-name inversion.
+The next vertical slice is a minimal `root_transformer` implementation:
 
-Compatibility preserves externally observable scientific behavior where it is
-validated; it does not promise to preserve every legacy implementation bug.
-The characterized legacy transfer path had a nonempty-global handling bug;
-the rewrite supports named globals. Negative weights, rare empty graphs,
-historical checkpoint variants, and legacy padding edge cases remain areas to
-audit when a supported consumer requires them.
+1. Define a small `SequenceSample` contract and deterministic fixture.
+2. Implement token construction using shared event/features infrastructure.
+3. Add the transformer model under `models/root_transformer/`.
+4. Connect it to the existing binary task and trainer on a tiny fixture.
+5. Add checkpoint, prediction, and reproducibility tests.
 
-## ROOT-GNN v1 completion checklist
+Do not generalize shared interfaces until this second representation exercises
+the proposed common behavior.
 
-- [x] active ROOT data path and graph cache
-- [x] validated feature, graph, model, task, and metric behavior
-- [x] train from scratch and fine-tune a pretrained backbone
-- [x] resume new-stack checkpoints and load supported historical weights
-- [x] evaluate and predict named outputs
-- [x] single-process and DDP application boundaries
-- [x] Perlmutter/Slurm launcher examples and profiling guidance
-- [x] ROOT-GNN ONNX export and CPU Runtime parity
-- [ ] streaming/sharded large-scale prediction output
-- [ ] removal of frozen legacy reference
-- [ ] ROOT-Transformer representation/model
+## Validation requirements
 
-ROOT-GNN v1 is complete when the checked-in new stack can prepare active data,
-reproduce validated legacy behavior, train, transfer, resume, evaluate,
-predict, and run single-process or DDP workflows. The remaining unchecked
-items are intentionally deferred rather than undocumented promises.
+Every new model family must provide unit tests for its representation and
+model, a small end-to-end integration test, checkpoint reload coverage, and a
+deterministic repeatability check. Scientific behavior that is intentionally
+shared with ROOT-GNN should be compared against the frozen reference fixture;
+architecture-specific behavior should have its own reference outputs.

legacy/LICENSE

@@ -1,21 +0,0 @@
-MIT License
-
-Copyright (c) 2025 LBL ATLAS
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.

legacy/README.md

@@ -1,358 +0,0 @@
----
-license: mit
-tags: 
-- arXiv:2412.10665
----
-
-This is a demo is of the approach described in the paper, ["Pretrained Event Classification Model for High Energy Physics Analysis"](https://arxiv.org/abs/2412.10665)
-```
-@misc{ho2024pretrained,
-  title={Pretrained Event Classification Model for High Energy Physics Analysis},
-  author={Joshua Ho, Benjamin Ryan Roberts, Shuo Han, Haichen Wang},
-  year={2024},
-  eprint={2412.10665},
-  archivePrefix={arXiv}
-}
-```
-
-## Abstract
-
-We introduce a foundation model for event classification in high-energy physics, built on a **Graph Neural Network** architecture and trained on **120 million simulated proton-proton collision events** spanning 12 distinct physics processes. The model is *pretrained* to learn a general and robust representation of collision data using challenging multiclass and multilabel classification tasks.
-
-Its performance is evaluated across five event classification tasks, which include both physics processes used during pretraining and new processes not encountered during pretraining. Fine-tuning the pretrained model significantly improves classification performance, particularly in scenarios with limited training data, demonstrating gains in both accuracy and computational efficiency.
-
-To investigate the underlying mechanisms behind these performance improvements, we employ a representational similarity evaluation framework based on *Centered Kernel Alignment*. This analysis reveals notable differences in the learned representations of fine-tuned pretrained models compared to baseline models trained from scratch.
-
-## Introduction
-
-Machine learning has become a ubiquitous tool in particle physics, employed in a variety of tasks including triggering, simulation, reconstruction, and offline analysis. While its utility spans classification, regression, and generative tasks, the current paradigm of developing machine learning models from scratch for each specific application presents several challenges. This approach not only demands specialized expertise and substantial computing resources but can also result in suboptimal performance due to limited training data. The from-scratch development of models necessitates individual validation studies to ensure that neural networks utilize well-modeled information from training samples, whether derived from Monte Carlo simulations or control samples from experimental data.
-
-Foundation models offer a promising direction to address these limitations. These models, pre-trained on large, diverse datasets across various tasks, provide robust and general representations of underlying data structures. Notable examples in other fields include GPT-4 [OpenAI et al., 2024](#ref-openai-2024-gpt4) and BERT [Devlin et al., 2018](#ref-devlin-2018-bert) in natural language processing, Stable Diffusion [Rombach et al., 2021](#ref-rombach-2021-latentdiffusion) in image processing, and AlphaFold [Jumper et al., 2021](#ref-jumper-2021-alphafold) in structural biology. The foundation model approach offers several advantages for particle physics applications: reduced computing resources for fine-tuning [Yosinski et al., 2014](#ref-yosinski-2014-transfer) compared to training from scratch, superior performance on specific tasks (particularly with limited training data), and potentially simplified validation procedures as downstream tasks inherit verified representations from the pre-trained model.
-
-Current literature on pretrained models for particle physics can be categorized based on the data representation they handle. Models operating on particle- or event-level numerical data use features like particle four momenta or jets, leveraging self-supervised or generative methods to learn versatile representations. Detector-focused models operate on high-dimensional responses such as calorimeter deposits or pixel hits, employing geometry-aware techniques for accurate simulation and analysis. Finally, models using textual or code representations apply large language model architectures to integrate domain knowledge, enabling tasks like question answering and code generation.
-
-Recent studies have begun exploring foundation models tailored to particle physics data, which has a variety of distinct structures and properties across many experiments and data processing stages, including:
-
-- particle-level & event-level numeric data [Wildridge et al., 2024](#ref-wildridge-2024-bumblebee), [Katel et al., 2024](#ref-katel-2024-jet), [Golling et al., 2024](#ref-golling-2024-maskedset), [Mikuni & Nachman, 2024](#ref-mikuni-2024-omnilearn), [Harris et al., 2024](#ref-harris-2024-resimulation), [Birk et al., 2024](#ref-birk-2024-omnijet), [Vigl et al., 2024](#ref-vigl-2024-finetune),
-- detector-level & geometry-aware data [Araz et al., 2024](#ref-araz-2024-pointcloud), [Liu et al., 2023](#ref-liu-2023-gaam), [Hashemi et al., 2024](#ref-hashemi-2024-gen), [Huang et al., 2024](#ref-huang-2024-lmtracking),
-- textual or code data [Zhang et al., 2024](#ref-zhang-2024-xiwu).
-
-This paper presents a foundation model designed specifically for collider event-level data. In modern collider experiments, final-stage analysis processes information from reconstructed objects that either directly correspond to particles in collision final states (such as leptons and photons) or serve as proxies (such as jets and missing transverse energy). While traditional approaches often relied on "high-level" variables calculated from object features, recent trends favor direct input of event objects and their features into neural networks for analysis tasks. A notable example is [ATLAS Collaboration, 2023](#ref-atlas-2023-4top), which established the observation of simultaneous production of four top quarks with the ATLAS experiment by employing a graph neural network (GNN) architecture to process event-level object information.
-
-We present foundation models that adopt an architecture similar to that used for [ATLAS Collaboration, 2023](#ref-atlas-2023-4top). Our models are pre-trained using either multiclass classification or multi-label learning tasks across 12 distinct physics processes. We evaluate these models through fine-tuning and testing on five classification tasks, including both familiar and novel processes not seen during pre-training. Our analysis benchmarks the models' performance improvements, their scaling behavior with training sample size, and computational efficiency, representing the first prototype of a foundation model operating on collider final-state object data.
-
-## Data Samples
-
-To provide a diverse set of physics processes for the pretraining, we use Madgraph@NLO 2.7.3 [Alwall et al., 2014](#ref-alwall-2014hca) to generate proton-proton collision events at next-to-leading order (NLO) in Quantum Chromodynamics (QCD). We generate 12 distinct Standard Model (SM) physics processes, including six major Higgs boson production mechanisms: gluon fusion production \\(ggF\\), vector boson fusion \\(VBF\\), associated production of the Higgs boson with a W boson \\(WH\\) or a Z boson \\(ZH\\), associated production of the Higgs boson with a top-quark pair \\(t\bar{t}H\\), and associated production of the Higgs boson with a single top quark and a forward quark \\(tHq\\). Additionally, we simulate six top quark production processes: single top production, top-quark pair production \\(t\bar{t}\\), top quark pair production in association with a pair of photons \\(t\bar{t}\gamma\gamma\\), associated production of a top-quark pair with a W boson \\(t\bar{t}W\\), simultaneous production of three top quarks \\(t\bar{t}t\\), and simultaneous production of four top quarks \\(t\bar{t}t\bar{t}\\). In these samples, the Higgs boson and top quarks decay inclusively. These 12 Higgs and top quark production processes constitute the pretraining dataset.
-
-To test the pretrained model, we further generated four processes including three beyond Standard Model (SM) processes: a SM \\(t\bar{t}H\\) production where the Higgs boson decays exclusively to a pair of photons, a \\(t\bar{t}H\\) production with the Higgs boson decaying to a pair of photons, where the top-Yukawa coupling is CP-odd, implemented using the Higgs Characterization model [Artoisenet et al., 2013](#ref-artoisinet-2013puc), the production of a pair of superpartners of the top quark (s-top) using the Minimal Supersymmetric Standard Model (MSSM) [Rosiek, 1990](#ref-rosiek-1990), [Allanach et al., 2009](#ref-allanach-2009), and flavor changing neutral current (FCNC) processes [Degrande et al., 2015](#ref-degrande-2015), [Durieux et al., 2015](#ref-durieux-2015). For the s-top process, we simulate the production of heavier s-top pairs \\(t_2\bar{t_2}\\), where each heavier s-top (mass 582 GeV) decays into a lighter s-top \\(t_1\\) or \\(\bar{t_1}\\), mass 400 GeV) and a Higgs boson. The FCNC process involves \\(t\bar{t}\\) production where one top quark decays to a Higgs boson and a light quark. We generate 10 million events for each process, except for \\(tHq\\) and \\(t\bar{t}t\bar{t}\\), where 5 million events were produced.
-
-In all simulation samples, the center of mass energy of the proton-proton collision is set to 13 TeV. The Higgs boson, top quarks, and vector bosons are set to decay inclusively (except the \\(t\bar{t}H \rightarrow \gamma\gamma\\) samples), with MadSpin [Artoisenet et al., 2012](#ref-artoisinet-2012st) handling the decays of top quarks and W bosons. The generated events are processed through Pythia 8.235 [Sjostrand et al., 2015](#ref-sjostrand-2015) for parton showering and heavy particle decays, followed by Delphes 3.4.2 [de Favereau et al., 2014](#ref-defavereau-2014) configured to emulate the ATLAS detector [ATLAS Collaboration, 2008](#ref-atlas-2008) for fast detector simulation.
-
-The detector-level object selection criteria are defined to align with typical experimental conditions. Photons are required to have transverse momentum \\(p_T \geq 20~\mathrm{GeV}\\) and pseudorapidity \\(|\eta| \leq 2.37\\), excluding the electromagnetic calorimeter crack region \\(1.37 < |\eta| < 1.52\\). Electrons must have \\(p_T \geq 10~\mathrm{GeV}\\) and \\(|\eta| \leq 2.47\\) (excluding the same crack region), while muons are selected with \\(p_T \geq 10~\mathrm{GeV}\\) and \\(|\eta| \leq 2.7\\). Jets are reconstructed using the anti-\\(k_t\\) algorithm [Cacciari et al., 2008](#ref-cacciari-2008gp) with radius parameter \\(\Delta R=0.4\\), where \\(\Delta R\\) is defined as \\(\sqrt{\Delta\eta ^2 + \Delta\phi^2}\\), with \\(\Delta\eta\\) being the difference in pseudorapidity and \\(\Delta\phi\\) the difference in azimuthal angle. Jets must satisfy \\(p_T \geq 25~\mathrm{GeV}\\) and \\(|\eta| \leq 2.5\\). To avoid double-counting, jets are removed if they are within \\(\Delta R < 0.4\\) of a photon or lepton. The identification of jets originating from b-quark decays (b-tagging) is performed by matching jets within \\(\Delta R = 0.4\\) of a b-quark, with efficiency corrections applied to match the performance of the ATLAS experiment's b-tagging algorithm [ATLAS Collaboration, 2019](#ref-atlas-2019bwq).
-
-## Methods
-
-### Overview
-
-We present a methodology for developing and evaluating a foundation model for particle collision event analysis. The approach centers on pretraining a Graph Neural Network (GNN) architecture using a comprehensive dataset that spans multiple physics tasks, enabling the model to learn robust and transferable features. For task-specific applications, we employ a fine-tuning strategy that combines output layer adaptation with carefully calibrated learning rates for updating the pretrained parameters.
-
-Given the prevalence of classification problems in particle physics data analysis, we evaluate the model's efficacy through a systematic assessment across five binary classification tasks:
-
-- \\(t\bar{t}H(\rightarrow \gamma\gamma)\\) with CP-even versus CP-odd t-H interaction
-- \\(t\bar{t}\\) with FCNC top quark decays versus $tHq$ processes
-- \\(t\bar{t}W\\) versus $ttt$ processes
-- Stop pair production with Higgs bosons in the decay chain versus \\(t\bar{t}H\\) processes
-- \\(WH\\) versus \\(ZH\\) production modes
-
-Our evaluation metrics encompass classification performance, computational efficiency, and model interpretability. The investigation extends to analyzing the model's scaling behavior with respect to training dataset size, benchmarked against models trained without pretraining. Although we explored transfer learning through parameter freezing of pretrained layers, this approach did not yield performance improvements, leading us to focus our detailed analysis on fine-tuning strategies.
-
-This methodological framework demonstrates the potential of foundation models to enhance the efficiency of particle physics analyses while improving task-specific performance, offering a promising direction for future high-energy physics research.
-
----
-
-### GNN Architecture
-
-We implement a Graph Neural Network (GNN) architecture that naturally accommodates the point-cloud structure of particle physics data, employing the DGL framework with a PyTorch backend [Wang et al., 2019][ref-dgl-2019], [Paszke et al., 2019][ref-pytorch-2019]. A fully connected graph is constructed for each event, with nodes corresponding to reconstructed jets, electrons, muons, photons, and \\(\vec{E}_T^{\text{miss}}\\). The features of each node include the four-momentum \\((p_T, \eta, \phi, E)\\) of the object with a massless assumption (\\(E = p_T \cosh \eta\\)), the b-tagging label (for jets), the charge (for leptons), and an integer labeling the type of object represented by the node. We use a placeholder value of 0 for features which are not defined for every node type such as the b-jet tag, lepton charge, or the pseudorapidity of \\(\vec{E}_T^{\text{miss}}\\). We assign the angular distances (\\(\Delta \eta, \Delta \phi, \Delta R\\)) as edge features and the number of nodes $N$ in the graph as a global feature. We denote the node features \\(\{\vec x_i\}\\), edge features \\(\{\vec y_{ij}\}\\), and global features \\(\{\vec z\}\\).
-
-The GNN model is based on the graph network architecture described in [Battaglia et al., 2018][ref-graphnets-2018] using simple multilayer perceptron (MLP) feature functions and summation aggregation. The model is comprised of three primary components: an encoder, the graph network, and a decoder. In the encoder, three MLPs embed the nodes, edges, and global features into a latent space of dimension 64. The graph network block, which is designed to facilitate message passing between different domains of the graph, performs an edge update $f_e$, followed by a node update $f_n$, and finally a global update $f_g$, all defined below. The inputs to each update MLP are concatenated.
-
-$$
-\vec {y'}_{ij} = f_e\left(\{\vec x_k\},\vec y_{ij},\vec z\right) = \mathrm{MLP}\left(\vec x_i,\vec x_j,\vec y_{ij},\vec z\right)
-$$
-
-$$
-\vec{x'}_{i} = f_n\left(\vec x_i,\{\vec{y'}_{jk}\},\vec z\right) = \mathrm{MLP}\left(\vec x_i,\sum_j\vec{y'}_{ij},\vec z\right)
-$$
-
-$$
-\vec{z'} = f_g\left(\{\vec{x'}_i\},\{\vec{y'}_{ij}\},\vec z\right) = \mathrm{MLP}\left(\sum_i\vec{x'}_i,\sum_{i,j}\vec{y'}_{ij},\vec z\right)
-$$
-
-This graph block is iterated four times with the same update MLPs. Finally, the global features are passed through a decoder MLP and a final layer linear to produce the desired model outputs. Each MLP consists of 4 linear layers, each with an output width of 64, with the `ReLU` activation function. The output of the MLP is then passed through a `LayerNorm` layer [Ba et al., 2016][ref-layernorm-2016]. The total number of trainable parameters in this model is about 400,000.
-
-As a performance benchmark, a baseline GNN model is trained from scratch for each classification task. The initial learning rate is set to \\(10^{-4}\\) with an exponential decay following \\(LR(x) = LR_{\text{initial}}\cdot(0.99)^x\\), where \\(x\\) represents the epoch number.
-
----
-
-### Pretraining Strategy
-
-We explore two complementary pretraining approaches to develop robust representations of collision events: (1) multi-class classification, which trains the model to distinguish between different physics processes, and (2) multi-label classification, which predicts the existence and kinematics of heavy particles with prompt decays. The pretraining dataset consists of approximately 120 million events, evenly distributed across 12 distinct physics processes, including all major Higgs boson production mechanisms and top quark processes as described in [Data Samples](#sec-data). This large-scale pretraining effort was conducted on the Perlmutter supercomputer at NERSC.
-
-#### Multi-class Classification
-
-For Monte Carlo simulated events, the underlying physics process that generated each event is known precisely, providing natural labels for supervised learning. However, the challenge lies in the complexity of collision events: different physics processes can produce similar kinematics and event topologies, particularly in certain regions of phase space. No single observable can unambiguously identify the underlying process. By training the model to distinguish between 12 different processes simultaneously, we challenge it to learn subtle differences in kinematics and topology that collectively characterize each process. The model is trained using categorical cross entropy as the loss function. The output layer of the multiclass classification model has 832 trainable parameters.
-
-#### Multi-label Classification
-
-This approach combines both classification and regression tasks to characterize collision events. For discrete properties like particle presence in specific kinematic regions, we employ classification labels with binary cross-entropy loss. For continuous quantities like particle multiplicities, we use regression labels with mean-squared error loss. This hybrid approach enables the model to learn both categorical and continuous aspects of the physics processes simultaneously.
-
-We develop a comprehensive set of 41 labels that capture both particle multiplicities and kinematic properties. This approach increases prediction granularity and enhances model interpretability. By training the model to predict event kinematics rather than event identification, we create a task-independent framework that can potentially generalize better to novel scenarios not seen during pretraining.
-
-The particle multiplicity labels count the number of Higgs bosons (\\(n_{\text{higgs}}\\)), top quarks (\\(n_{\text{tops}}\\)), vector bosons (\\(n_V\\)), \\(W\\) bosons (\\(n_W\\)), and \\(Z\\) bosons (\\(n_Z\\)). The kinematic labels characterize the transverse momentum (\\(p_T\\)), pseudorapidity (\\(\eta\\)), and azimuthal angle (\\(\phi\\)) of Higgs bosons and top quarks through binned classifications.
-
-For Higgs bosons, $p_T$ is categorized into three ranges: (0, 30) GeV, (30, 200) GeV, and (200, \\(\infty\\)) GeV, with the upper range particularly sensitive to potential BSM effects. Similarly, both leading and subleading top quarks have $p_T$ classifications spanning (0, 30) GeV, (30, 300) GeV, and (300, \\(\infty\\)) GeV. When no particle exists within a specific \\(p_T\\) range, the corresponding label is set to \\([0, 0, 0]\\). For all particles, \\(\eta\\) measurements are divided into 4 bins with boundaries at \\([-1.5, 0, 1.5]\\), while \\(\phi\\) measurements use 4 bins with boundaries at \\([-\frac{\pi}{2}, 0, \frac{\pi}{2}]\\). As with \\(p_T\\), both \\(\eta\\) and \\(\phi\\) labels default to \\([0, 0, 0, 0]\\) in the absence of a particle. This comprehensive labeling schema enables fine-grained learning of kinematic distributions and particle multiplicities, essential for characterizing complex collision events.
-
-The loss function combines individual losses from all 41 labels through weighted averaging. Binary cross-entropy is applied to classification labels, while mean-squared error is used for regression labels. The model generates predictions for all labels simultaneously, with individual losses calculated according to their respective types. The final loss is computed as an equally-weighted average across all labels, with weights set to 1 to ensure uniform contribution to the optimization process. The output layer of the multilabel model has 2,688 trainable parameters.
-
-#### Pretraining
-
-During pre-training, the initial learning rate is \\(10^{-4}\\), and the learning rate decays by 1% each epoch following the power law function \\(LR(x) = 10^{-4}\cdot(0.99)^x\\), where \\(x\\) is the number of epochs. Both pre-trained models reach a plateau in loss by epoch 50, at which point the training is stopped.
-
----
-### Fine-tuning Methodology
-
-For downstream tasks, we adjust the model architecture for fine-tuning by replacing the original output layer (final linear layer) with a newly initialized linear layer while retaining the pre-trained weights for all other layers. This modification allows the model to specialize in the specific downstream task while leveraging the general features learned during pretraining.
-
-The fine-tuning process begins with distinct learning rate setups for different parts of the model. The newly initialized linear layer is trained with an initial learning rate of \\(10^{-4}\\), matching the rate used for models trained from scratch. Meanwhile, the pre-trained layers are fine-tuned more cautiously with a lower initial learning rate of \\(10^{-5}\\). This approach ensures that the pre-trained layers adapt gradually without losing their general features, while the new layer learns effectively from scratch. Both learning rates decay over time following the same power law function, \\(LR(x) = LR_{initial} \cdot (0.99)^x\\), to promote stable convergence as training progresses.
-
-We also evaluated a transfer learning setup in which either the decoder MLP or the final linear layer was replaced with a newly initialized component. During this process, all other model parameters remained frozen, leveraging the pre-trained features without further updating them. However, we did not observe performance improvements using the transfer learning setup. Consequently, we focus on reporting results obtained with the fine-tuning approach.
-
----
-
-### Performance Evaluation
-
-We assess model performance using two figures of merit: the classification accuracy and the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve. The accuracy is defined as the fraction of correctly classified events when applying a threshold of 0.5 to the neural network output score. Both metrics demonstrate consistent trends in our analysis.
-
-To obtain reliable performance estimates and uncertainties, we employ an ensemble training approach where 5 independent models are trained for each configuration with random weight initialization and random subsets of the training dataset. This enables us to evaluate both the models' sensitivity to initial parameters and to quantify uncertainties in their performance.
-
-To investigate how model performance scales with training data, we conducted training runs using sample sizes ranging from \\(10^3\\) to \\(10^7\\) events per class (\\(10^3\\), \\(10^4\\), \\(10^5\\), \\(10^6\\), and \\(10^7\\)) for each model setup: the from-scratch baseline and models fine-tuned from multi-class or multi-label pretrained models. For the \\(10^7\\) case, only the initialization was randomized due to dataset size limitations. All models were evaluated on the same testing dataset, consisting of 2 million events per class, which remained separate from the training process.
-
-| **Name of Task**     | **Pretraining Task** | \\(10^3\\)         | \\(10^4\\)         | \\(10^5\\)         | \\(10^6\\)         | \\(10^7\\)         |
-|----------------------|----------------------|--------------------|--------------------|--------------------|--------------------|--------------------|
-| **ttH CP Even vs Odd** | Baseline Accuracy     | 56.5 ± 1.1 | 62.2 ± 0.1 | 64.3 ± 0.0 | 65.7 ± 0.0 | 66.2 ± 0.0 |
-|                        | Multiclass (%)        | +4.8 ± 1.1 | +3.4 ± 0.1 | +1.3 ± 0.0 | +0.2 ± 0.0 | −0.0 ± 0.0 |
-|                        | Multilabel (%)        | +2.1 ± 1.2 | +1.9 ± 0.1 | +0.8 ± 0.1 | +0.0 ± 0.0 | −0.1 ± 0.0 |
-| **FCNC vs tHq**        | Baseline Accuracy     | 63.6 ± 0.7 | 67.8 ± 0.4 | 68.4 ± 0.3 | 69.3 ± 0.3 | 67.9 ± 0.0 |
-|                        | Multiclass (%)        | +5.8 ± 0.8 | +1.2 ± 0.4 | +1.4 ± 0.3 | +0.5 ± 0.3 | −0.0 ± 0.0 |
-|                        | Multilabel (%)        | −5.3 ± 0.8 | −1.3 ± 0.4 | +0.9 ± 0.4 | +0.3 ± 0.3 | +0.4 ± 0.1 |
-| **ttW vs ttt**         | Baseline Accuracy     | 75.8 ± 0.1 | 77.6 ± 0.1 | 78.9 ± 0.0 | 79.8 ± 0.0 | 80.3 ± 0.0 |
-|                        | Multiclass (%)        | +3.7 ± 0.1 | +2.7 ± 0.1 | +1.3 ± 0.0 | +0.4 ± 0.0 | +0.0 ± 0.0 |
-|                        | Multilabel (%)        | +2.2 ± 0.1 | +1.1 ± 0.1 | +0.5 ± 0.0 | +0.0 ± 0.0 | −0.1 ± 0.0 |
-| **stop vs ttH**        | Baseline Accuracy     | 83.0 ± 0.2 | 86.3 ± 0.1 | 87.6 ± 0.0 | 88.5 ± 0.0 | 88.8 ± 0.0 |
-|                        | Multiclass (%)        | +0.4 ± 0.2 | +1.9 ± 0.1 | +1.0 ± 0.0 | +0.3 ± 0.0 | +0.0 ± 0.0 |
-|                        | Multilabel (%)        | +2.8 ± 0.2 | +1.0 ± 0.1 | +0.5 ± 0.0 | +0.0 ± 0.0 | −0.0 ± 0.0 |
-| **WH vs ZH**           | Baseline Accuracy     | 51.4 ± 0.1 | 53.9 ± 0.1 | 55.8 ± 0.0 | 57.5 ± 0.0 | 58.0 ± 0.0 |
-|                        | Multiclass (%)        | +5.2 ± 0.1 | +5.3 ± 0.1 | +3.1 ± 0.0 | +0.6 ± 0.0 | +0.1 ± 0.0 |
-|                        | Multilabel (%)        | −1.1 ± 0.1 | −0.9 ± 0.2 | +0.5 ± 0.1 | +0.1 ± 0.0 | −0.1 ± 0.0 |
-
-> **Table 1**: Accuracy of the traditional model versus the accuracy increase due to fine-tuning from various pretraining tasks.  
-> The accuracies are averaged over 5 independently trained models with randomly initialized weights and trained on a random subset of the data. One exception is the \\(10^7\\) training where all models use the same dataset due to limitations on our dataset size. The random subsets are allowed to overlap, but this overlap should be very minimal because all models take an independent random subset of \\(10^7\\) events. The testing accuracy is calculated from the same testing set of 2 million events per class across all models for a specific training task. The errors are the propagated errors (root sum of squares) of the standard deviation of accuracies for each model.
-
-## Results
-
-### Classification Performance
-
-Since the observations of AUC and accuracy show similar trends, we focus the presentation of the results using accuracy here for conciseness in Table 1.
-
-In general, the fine-tuned pretrained model achieves at least the same level of classification performance as the baseline model. Notably, there are significant improvements, particularly when the sample size is small, ranging from \\(10^3\\) to \\(10^4\\) events. In some cases, the accuracy improvements exceed five percentage points, demonstrating that pretrained models provide a strong initial representation that compensates for limited data. The numerical values of the improvements in accuracy may not fully capture the impact on the sensitivity of the measurements for which the neural network classifier is used, and the final sensitivity improvement is likely to be greater.
-
-As the training sample size grows to \\(10^5\\), \\(10^6\\), and eventually \\(10^7\\) events, the added benefit of pretraining diminishes. With abundant data, models trained from scratch approach or even match the accuracy of fine-tuned pretrained models. This suggests that large datasets enable effective learning from scratch, rendering the advantage of pretraining negligible in such scenarios.
-
-Although both pretraining approaches offer benefits, multiclass pretraining tends to provide more consistent improvements across tasks, especially in the low-data regime. In contrast, multilabel pretraining can sometimes lead to neutral or even slightly negative effects for certain tasks and data sizes. This highlights the importance of the pretraining task design, as the similarity between pretraining and fine-tuning tasks in the multiclass approach appears to yield better-aligned representations.
-
-Finally, the spread of accuracy across the five tasks for the baseline model is quite large, offering a robust test of fine-tuning across tasks of varying difficulty. The consistent observation of these trends across tasks confirms the reliability and robustness of the findings.
-
----
-
-### Model Interpretability
-
-We aim to understand whether pretrained and baseline models learn the same underlying representations. If the two models exhibit high similarity, a plausible interpretation is that pretraining provides the pretrained model with an advantageous initialization, allowing it to converge to a similar state as the baseline model more efficiently. Conversely, significant differences between the models would indicate that pretraining facilitates the development of a more general and robust latent space, which serves as a foundation for fine-tuning to effectively adapt to the downstream task. To investigate this, we analyzed the representational similarity between a pretrained model fine-tuned for the downstream task and a baseline model trained directly on the downstream task without pretraining.
-
-We use Centered Kernel Alignment (CKA) [Kornblith et al., 2019][ref-kornblith-2019-cka] to analyze model similarity and interpretability. CKA is a robust metric that quantifies the similarity between the internal representations of neural networks by comparing their feature matrices in a manner that is invariant to scaling, rotation, and alignment. This invariance makes CKA particularly effective for studying relationships between network layers, even across networks of different sizes or those trained from varying initializations.
-
-The similarity is evaluated using a 64-dimensional latent representation after the decoder stage of the GNN model. This choice allows us to compare the internal states of the models at a fine-grained level and understand how training strategies impact the representations directly used for the output task.
-
-To provide an intuitive understanding of CKA values, we construct a table of the CKA scores for various transformations performed on a set of dummy data.
-
-- **A:** randomly initialized matrix with shape (1000, 64), following a normal distribution (\\(\sigma = 1, \mu = 0\\))
-- **B:** matrix with shape (1000, 64) constructed via various transformations performed on \\(A\\)
-- **Noise:** randomly initialized noise matrix with shape (1000, 64), following a normal distribution (\\(\sigma = 1, \mu = 0\\))
-
-| Dataset | CKA Score |
-|---------|-----------|
-| \\(A, B = A\\) | 1.00 |
-| \\(A, B =\\) permutation on columns of \\(A\\) | 1.00 |
-| \\(A, B = A + \mathrm{Noise}(0.1)\\) | 0.99 |
-| \\(A, B = A + \mathrm{Noise}(0.5)\\) | 0.80 |
-| \\(A, B = A + \mathrm{Noise}(0.75)\\) | 0.77 |
-| \\(A, B = A \cdot \mathrm{Noise}(1)\\) (Linear Transformation) | 0.76 |
-| \\(A, B = A + \mathrm{Noise}(1)\\) | 0.69 |
-| \\(A, B = A + \mathrm{Noise}(2)\\) | 0.51 |
-| \\(A, B = A + \mathrm{Noise}(5)\\) | 0.39 |
-
-**Table 2:** CKA scores for a dummy dataset \\(A\\) and \\(B\\), where \\(B\\) is created via various transformations performed on \\(A\\).
-
-As seen in Table 2 and in the definition of the CKA, the CKA score is permutation-invariant. We will use the CKA score to evaluate the similarity between various models and gain insight into the learned representation of detector events in each model (i.e., the information that each model learns).
-
-We train ensembles of models for each training task to observe how the CKA score changes due to the random initialization of our models. The CKA score between two models is then defined to be:
-
-\\[
-CKA(A, B) = \frac{1}{n^2} \sum_i^n \sum_j^n CKA(A_i, B_j)
-\\]
-
-where \\(A_i\\) is the representation learned by the \\(i^{\text{th}}\\) model in an ensemble with \\(n\\) total models. The error in CKA is the standard deviation of \\(CKA(A_i, B_j)\\).
-
-Here we present results for the CKA similarity between the final model in each setup with the final model in the baseline, shown in Table 3.
-
-| Training Task         | Baseline         | Multiclass      | Multilabel      |
-|-----------------------|------------------|-----------------|-----------------|
-| ttH CP Even vs Odd    | 0.94 ± 0.05      | 0.82 ± 0.01     | 0.77 ± 0.06     |
-| FCNC vs tHq           | 0.96 ± 0.03      | 0.76 ± 0.01     | 0.81 ± 0.01     |
-| ttW vs ttt            | 0.91 ± 0.08      | 0.75 ± 0.10     | 0.72 ± 0.05     |
-| stop vs ttH           | 0.87 ± 0.11      | 0.79 ± 0.12     | 0.71 ± 0.08     |
-| WH vs ZH              | 0.90 ± 0.07      | 0.53 ± 0.03     | 0.44 ± 0.06     |
-
-**Table 3:** CKA Similarity of the latent representation before the decoder with the baseline model, averaged over 3 models per training setup, and all models trained with the full dataset (\\(10^7\\)). The baseline column is not guaranteed to be 1.0 because of the random initialization of the model. Each baseline model converges to a slightly different representation as seen in the CKA values in that column.
-
-The baseline models with different initializations exhibit high similarity values, ranging from approximately 0.87 to 0.96, which indicates that independently trained baseline models tend to converge on similar internal representations despite random initialization. Across the considered tasks, models trained as multi-class or multi-label classifiers exhibit noticeably lower CKA similarity scores when compared to the baseline model. For example, in the WH vs ZH task, the baseline model and another baseline trained model have a high similarity of 0.90, whereas the multi-class and multi-label models show significantly reduced similarities (0.53 and 0.44, respectively). This pattern suggests that the representational spaces developed by multi-class or multi-label models differ substantially from those learned by the baseline model that was trained directly on the downstream classification task.
-
-### Computational Efficiency
-
-To estimate the computational resources required for each approach, we measured the wall time needed for a model to reach its final performance. For baseline models, this is defined as the wall time from the start of training until the loss of the model plateaus. For the foundation model approach, the estimate includes both the pretraining time and the fine-tuning time, each measured from the start of training until the loss plateaus. This approach ensures a consistent and comprehensive evaluation of the computational demands.
-
-![The ratio of the fine-tuning time required to achieve 99% of the baseline model's final classification accuracy to the total time spent training the baseline model.](training_time.png)
-*Fig. 1: The ratio of the fine-tuning time required to achieve 99% of the baseline model's final classification accuracy to the total time spent training the baseline model.*
-
-Figure 1 shows the fine-tuning time for the model pretrained with multiclass classification, relative to the time required for the baseline model, as a function of training sample size. In general, the fine-tuning time is significantly shorter than the training time required by the baseline model approach. For smaller training sets, on the order of \\(10^5\\) events, tasks such as FCNC vs. tHq and ttW vs. ttt benefit substantially from the pretrained model’s “head start,” achieving their final performance in only about 1% of the baseline time. For large training datasets, the fine-tuning time relative to the baseline training time becomes larger; however, given that the large training sample typically requires longer training time, fine-tuning still yields much faster training convergence. The ttH CP-even vs. ttH CP-odd task, with a training sample size of \\(10^7\\) events, is an exception where the fine-tuning time exceeds the training time required for the baseline model. This is likely because the processes involved in this task include photon objects in the final states, which are absent from the events used during pretraining.
-
-To accurately evaluate the total time consumption, it is necessary to include the pretraining time required for the foundation model approach. The pretraining times are as follows:
-
-- **Multi-class pretraining:** 45.5 GPU hours
-- **Multi-label pretraining:** 60.0 GPU hours
-
-The GPU hours recorded for the multi-label model represent the total time required when training the model in parallel on 16 GPUs. This includes a model synchronization step, which results in higher GPU hours compared to the multi-class pretraining model.
-
-The foundation model approach becomes increasingly efficient when a large number of tasks are fine-tuned using the same pretrained model, compared to training each task independently from scratch. To illustrate this, we evaluate the computational time required for a scenario where the training sample contains \\(10^7\\) events. For the five tasks tested in this study, the baseline training time (training from scratch) ranges from 1.68 GPU hours (WH vs. ZH) to 5.30 GPU hours (ttW vs. ttt), with an average baseline training time of 2.94 GPU hours. In contrast, the average fine-tuning time for the foundation model approach, relative to the baseline, is 38% of the baseline training time for \\(10^7\\) events. Based on these averages, we estimate that the foundation model approach becomes more computationally efficient than the baseline approach when fine-tuning is performed for more than 41 tasks.
-
-As a practical example, the ATLAS measurement of Higgs boson couplings using the \\(H \rightarrow \gamma\gamma\\) decay channel [ATLAS Collaboration, 2023][ref-atlas-2023-higg] involved training 42 classifiers for event categorization. This coincides with our estimate, suggesting that the foundation model approach can reduce computational costs even for a single high-energy physics measurement.
-
-## Conclusions
-
-We presented an in-depth study of a particle physics foundation model designed to operate on the four-momentum and identification properties of event final-state objects. This model is built on a Graph Neural Network (GNN) architecture and trained on a dataset comprising 120 million simulated proton-proton collision events across 12 distinct physics processes. The pretraining phase explored both multiclass and multilabel classification tasks, providing a robust foundation for downstream applications. Notably, the pretrained models demonstrated significant improvements in event classification performance when fine-tuned, particularly for tasks with limited training samples.
-
-The foundation model approach also offers substantial computational advantages. By leveraging fine-tuning, this methodology reduces the computational resources required for large-scale applications across multiple tasks. Our estimates indicate that significant resource savings can be achieved even for single particle physics measurements, making this approach both scalable and efficient.
-
-To better understand the learned representations of the pretrained model and guide future optimization efforts, we employed a representational similarity evaluation framework using Centered Kernel Alignment (CKA). This metric allowed us to investigate the source of the performance gains observed in the foundation model. Our analysis revealed notable differences in the learned representations between the fine-tuned pretrained model and a baseline model trained from scratch. In deep learning, it is well-established that multiple equally valid solutions can exist. Future studies are necessary to determine whether the low similarity in latent representations reflects complementary information uniquely captured by the foundation and baseline models, or if it can simply be attributed to connected local minima in the loss landscape.
-
-## Acknowledgments
-
-This work is supported by the U.S. National Science Foundation under the Award No. 2046280, and by U.S. Department of Energy, Office of Science under contract DE-AC02-05CH11231.
-
-## References
-
-- <span id="ref-openai-2024-gpt4"></span> **OpenAI et al.** GPT-4 Technical Report. arXiv:2303.08774 (2024). [https://arxiv.org/abs/2303.08774](https://arxiv.org/abs/2303.08774)
-
-- <span id="ref-yosinski-2014-transfer"></span> **Jason Yosinski, Jeff Clune, Yoshua Bengio, Hod Lipson.** How transferable are features in deep neural networks? CoRR abs/1411.1792 (2014). [http://arxiv.org/abs/1411.1792](http://arxiv.org/abs/1411.1792)
-
-- <span id="ref-rombach-2021-latentdiffusion"></span> **Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, Björn Ommer.** High-Resolution Image Synthesis with Latent Diffusion Models. CoRR abs/2112.10752 (2021). [https://arxiv.org/abs/2112.10752](https://arxiv.org/abs/2112.10752)
-
-- <span id="ref-podell-2023-sdxl"></span> **Dustin Podell, Zion English, Kyle Lacey et al.** SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. arXiv:2307.01952 (2023). [https://arxiv.org/abs/2307.01952](https://arxiv.org/abs/2307.01952)
-
-- <span id="ref-jumper-2021-alphafold"></span> **John Jumper, Richard Evans, Alexander Pritzel et al.** Highly accurate protein structure prediction with AlphaFold. Nature 596, 583-589 (2021). [https://doi.org/10.1038/s41586-021-03819-2](https://doi.org/10.1038/s41586-021-03819-2)
-
-- <span id="ref-devlin-2018-bert"></span> **Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova.** BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. CoRR abs/1810.04805 (2018). [http://arxiv.org/abs/1810.04805](http://arxiv.org/abs/1810.04805)
-
-- <span id="ref-atlas-2023-higg"></span> **ATLAS Collaboration.** Measurement of the properties of Higgs boson production at \\(\sqrt{s} = 13\,\text{TeV}\\) in the \\(H \to \gamma\gamma\\) channel using \\(139\,\text{fb}^{-1}\\) of \\(pp\\) collision data with the ATLAS experiment. JHEP 07 (2023) 088. [arXiv:2207.00348](https://arxiv.org/abs/2207.00348), [https://doi.org/10.1007/JHEP07(2023)088](https://doi.org/10.1007/JHEP07(2023)088)
-
-- <span id="ref-atlas-2023-4top"></span> **ATLAS Collaboration.** Observation of four-top-quark production in the multilepton final state with the ATLAS detector. Eur. Phys. J. C 83 (2023) 496. [arXiv:2303.15061](https://arxiv.org/abs/2303.15061), [https://doi.org/10.1140/epjc/s10052-023-11573-0](https://doi.org/10.1140/epjc/s10052-023-11573-0)
-
-- <span id="ref-kornblith-2019-cka"></span> **Simon Kornblith, Mohammad Norouzi, Honglak Lee, Geoffrey Hinton.** Similarity of Neural Network Representations Revisited. CoRR abs/1905.00414 (2019). [http://arxiv.org/abs/1905.00414](http://arxiv.org/abs/1905.00414)
-
----
-
-<!-- Historical/General Physics foundational texts -->
-
-- <span id="ref-birell-1982-qfields"></span> **N. D. Birell, P. C. W. Davies.** Quantum Fields in Curved Space. Cambridge Univ. Press (1982).
-
-- <span id="ref-feynman-1954"></span> **R. P. Feynman.** Phys. Rev. 94, 262 (1954).
-
-- <span id="ref-einstein-1935-epr"></span> **A. Einstein, Yu. Podolsky, N. Rosen.** Phys. Rev. 47, 777 (1935).
-
-- <span id="ref-berman-1983-stability"></span> **G. P. Berman, Jr., F. M. Izrailev, Jr.** Stability of nonlinear modes. Physica D 88, 445 (1983).
-
-- <span id="ref-davies-1988-trapped"></span> **E. B. Davies, L. Parns.** Trapped modes in acoustic waveguides. Q. J. Mech. Appl. Math. 51, 477–492 (1988).
-
-- <span id="ref-witten-2001"></span> **Edward Witten.** hep-th/0106109 (2001). [https://arxiv.org/abs/hep-th/0106109](https://arxiv.org/abs/hep-th/0106109)
-
----
-
-<!-- Particle physics/data science foundational models -->
-
-- <span id="ref-beutler-1994-hem"></span> **E. Beutler.** Williams Hematology, 5th Edition, Chapter 7, pp. 654–662. McGraw-Hill, New York (1994).
-
-- <span id="ref-knuth-1973-fa"></span> **Donald E. Knuth.** The Art of Computer Programming vol. 1: Fundamental Algorithms, 2nd Ed., Addison-Wesley (1973).
-
-- <span id="ref-smith-2005-philos"></span> **J. S. Smith, G. W. Johnson.** Philos. Trans. R. Soc. London, Ser. B 777, 1395 (2005).
-
-- <span id="ref-smith-2010-jap-unpub"></span> **W. J. Smith, T. J. Johnson, B. G. Miller.** Surface chemistry and preferential crystal orientation on a silicon surface. J. Appl. Phys. (unpublished, 2010).
-
-- <span id="ref-smith-2010-jap-sub"></span> **V. K. Smith, K. Johnson, M. O. Klein.** Surface chemistry and preferential crystal orientation on a silicon surface. J. Appl. Phys. (submitted, 2010).
-
-- <span id="ref-underwood-1988-lowerbounds"></span> **Ulrich Underwood, Ned Net, Paul Pot.** Lower Bounds for Wishful Research Results. Talk at Fanstord University (1988).
-
-- <span id="ref-johnson-2007-comm"></span> **M. P. Johnson, K. L. Miller, K. Smith.** Personal communication (Jan-May 2007).
-
----
-
-<!-- Prototypical collider software and tools -->
-
-- <span id="ref-pytorch-2019"></span> **Adam Paszke et al.** PyTorch: An Imperative Style, High-Performance Deep Learning Library. arXiv:1912.01703 (2019). [http://arxiv.org/abs/1912.01703](http://arxiv.org/abs/1912.01703)
-
-- <span id="ref-dgl-2019"></span> **Minjie Wang et al.** Deep Graph Library: Towards Efficient and Scalable Deep Learning on Graphs. arXiv:1909.01315 (2019). [http://arxiv.org/abs/1909.01315](http://arxiv.org/abs/1909.01315)
-
-- <span id="ref-graphnets-2018"></span> **Peter W. Battaglia et al.** Relational inductive biases, deep learning, and graph networks. arXiv:1806.01261 (2018). [http://arxiv.org/abs/1806.01261](http://arxiv.org/abs/1806.01261)
-
-- <span id="ref-layernorm-2016"></span> **Jimmy Lei Ba, Jamie Ryan Kiros, Geoffrey E. Hinton.** Layer Normalization. arXiv:1607.06450 (2016). [https://arxiv.org/abs/1607.06450](https://arxiv.org/abs/1607.06450)
-
----
-
-<!-- Recent & foundation models in HEP ML -->
-
-- <span id="ref-wildridge-2024-bumblebee"></span> **Andrew J. Wildridge et al.** Bumblebee: Foundation Model for Particle Physics Discovery. arXiv:2412.07867 (2024). [https://arxiv.org/abs/2412.07867](https://arxiv.org/abs/2412.07867)
-
-- <span id="ref-katel-2024-jet"></span> **Subash Katel et al.** Learning Symmetry-Independent Jet Representations via Jet-Based Joint Embedding Predictive Architecture. arXiv:2412.05333 (2024). [https://arxiv.org/abs/2412.05333](https://arxiv.org/abs/2412.05333)
-
-- <span id="ref-araz-2024-pointcloud"></span> **Jack Y. Araz et al.** Point cloud-based diffusion models for the Electron-Ion Collider. arXiv:2410.22421 (2024). [https://arxiv.org/abs/2410.22421](https://arxiv.org/abs/2410.22421)
-
-- <span id="ref-leigh-2024-maskedparticle"></span> **Matthew Leigh et al.** Is Tokenization Needed for Masked Particle Modelling? arXiv:2409.12589 (2024). [https://arxiv.org/abs/2409.12589](https://arxiv.org/abs/2409.12589)
-
-- <span id="ref-mikuni-2024-omnilearn"></span> **Vinicius Mikuni, Benjamin Nachman.** OmniLearn: A Method to Simultaneously Facilitate All Jet Physics Tasks. arXiv:2404.16091 (2024). [https://arxiv.org/abs/2404.16091](https://arxiv.org/abs/2404.16091)
-
-- <span id="ref-zhang-2024-xiwu"></span> **Zhengde Zhang et al.** Xiwu: A Basis Flexible and Learnable LLM for High Energy Physics. arXiv:2404.08001 (2024). [https://arxiv.org/abs/2404.08001](https://arxiv.org/abs/2404.08001)
-
-- <span id="ref-harris-2024-resimulation"></span> **Philip Harris et al.** Re-Simulation-based Self-Supervised Learning for Pre-Training Foundation Models. arXiv:2403.07066 (2024). [https://arxiv.org/abs/2403.07066](https://arxiv.org/abs/2403.07066)
-
-- <span id="ref-birk-2024-omnijet"></span> **Joschka Birk, Anna Hallin, Gregor Kasieczka.** OmniJet-$\alpha$: the first cross-task foundation model for particle physics. Machine Learning: Science and Technology. 5(3), 035031 (Aug 2024). [https://doi.org/10.1088/2632-2153/ad66ad](https://doi.org/10.1088/2632-2153/ad66ad)
-
-- <span id="ref-huang-2024-lmtracking"></span> **Andris Huang et al.** A Language Model for Particle Tracking. arXiv:2402.10239 (2024). [https://arxiv.org/abs/2402.10239](https://arxiv.org/abs/2402.10239)
-
-- <span id="ref-golling-2024-maskedset"></span> **Tobias Golling et al.** Masked Particle Modeling on Sets: Towards Self-Supervised High Energy Physics Foundation Models. arXiv:2401.13537 (2024). [https://arxiv.org/abs/2401.13537](https://arxiv.org/abs/2401.13537)
-
-- <span id="ref-liu-2023-gaam"></span> **Junze Liu et al.** Generalizing to new geometries with Geometry-Aware Autoregressive Models (GAAMs) for fast calorimeter simulation. Journal of Instrumentation 18(11), P11003 (Nov 2023). [https://doi.org/10.1088/1748-0221/18/11/p11003](https://doi.org/10.1088/1748-0221/18/11/p11003)
-
-- <span id="ref-hashemi-2024-gen"></span> **Baran Hashemi et al.** Ultra-high-granularity detector simulation with intra-event aware generative adversarial network and self-supervised relational reasoning. Nature Communications 15(1) (June 2024). [https://doi.org/10.1038/s41467-024-49104-4](https://doi.org/10.1038/s41467-024-49104-4)
-
-- <span id="ref-vigl-2024-finetune"></span> **Matthias Vigl et al.** Finetuning Foundation Models for Joint Analysis Optimization. arXiv:2401.13536 (2024). [https://arxiv.org/abs/2401.13536](https://arxiv.org/abs/2401.13536)
-
-- <span id="ref-li-2024-refine"></span> **Chen Li, Hao Cai, Xianyang Jiang.** Refine neutrino events reconstruction with BEiT-3. Journal of Instrumentation 19(6), T06003 (Jun 2024). [https://doi.org/10.1088/1748-0221/19/06/t06003](https://doi.org/10.1088/1748-0221/19/06/t06003)

legacy/physicsnemo/configs/config.yaml

@@ -1,64 +0,0 @@
-# ignore_header_test
-# Copyright 2023 Stanford University
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-random_seed: 2
-
-scheduler:
-  lr: 1.E-3
-  lr_decay: 1.E-3
-
-training:
-  epochs: 100
-  
-checkpoints:
-  ckpt_path: "checkpoints"
-  ckpt_name: "config"
-
-performance:
-  amp: False
-  jit: False
-
-architecture:
-  processor_size: 8
-  hidden_dim_node_encoder: 128
-  hidden_dim_edge_encoder: 128
-  hidden_dim_processor: 128
-  hidden_dim_node_decoder: 128
-  out_dim: 1
-
-paths:
-  data_dir: /global/cfs/projectdirs/atlas/joshua/hackathon_data/stats_100K
-  save_dir: /pscratch/sd/j/joshuaho/physicsnemo/graphs/stats_100K
-  training_dir: ./training_stats_100K/
-
-datasets:
-  - name: ttH_cp_even
-    load_path: ${paths.data_dir}/ttH_NLO.root
-    label: 0
-  - name: ttH_cp_odd
-    load_path: ${paths.data_dir}/ttH_CPodd.root
-    label: 1
-
-root_dataset:
-  ttree: output
-  type: torch.bfloat16
-  particles: ["jet", "ele", "mu", "ph", "MET"]
-  features: ["pt", "eta", "phi", "energy", "btag", "charge", "node_type"]
-  globals: []
-  weights: ""
-  tracking: []
-  step_size: 8192
-  batch_size: 8192
-  train_val_test_split: [0.75, 0.24, 0.01]

legacy/physicsnemo/configs/config_stats_all.yaml

@@ -1,65 +0,0 @@
-# ignore_header_test
-# Copyright 2023 Stanford University
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-random_seed: 2
-
-scheduler:
-  lr: 1.E-4
-  lr_decay: 1.E-3
-
-training:
-  epochs: 100
-  
-checkpoints:
-  ckpt_path: "checkpoints"
-  ckpt_name: "config_stats_all"
-
-performance:
-  amp: False
-  jit: False
-
-architecture:
-  processor_size: 5
-  hidden_dim_node_encoder: 64
-  hidden_dim_edge_encoder: 64
-  hidden_dim_processor: 64
-  hidden_dim_node_decoder: 64
-  out_dim: 1
-
-paths:
-  data_dir: /global/cfs/projectdirs/atlas/joshua/hackathon_data/stats_all
-  save_dir: /pscratch/sd/j/joshuaho/physicsnemo/graphs/stats_all
-  training_dir: ./training_stats_all/
-
-datasets:
-  - name: ttH_cp_even
-    load_path: ${paths.data_dir}/ttH_NLO.root
-    label: 0
-  - name: ttH_cp_odd
-    load_path: ${paths.data_dir}/ttH_CPodd.root
-    label: 1
-
-root_dataset:
-  ttree: output
-  type: torch.bfloat16
-  particles: ["jet", "ele", "mu", "ph", "MET"]
-  features: ["pt", "eta", "phi", "energy", "btag", "charge", "node_type"]
-  globals: []
-  weights: ""
-  tracking: []
-  step_size: 81920
-  batch_size: 8192
-  train_val_test_split: [0.75, 0.24, 0.01]
-  prebatch: True

legacy/physicsnemo/configs/tHjb_CP_0_vs_45.yaml

@@ -1,79 +0,0 @@
-# ignore_header_test
-# Copyright 2023 Stanford University
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-random_seed: 2
-
-scheduler:
-  lr: 1.E-3
-  lr_decay: 1.E-3
-
-training:
-  epochs: 100
-  
-checkpoints:
-  ckpt_path: "checkpoints"
-  ckpt_name: "config"
-
-performance:
-  amp: False
-  jit: False
-
-architecture:
-  processor_size: 8
-  hidden_dim_node_encoder: 128
-  hidden_dim_edge_encoder: 128
-  hidden_dim_processor: 128
-  hidden_dim_node_decoder: 128
-  global_emb_dim: 128
-  out_dim: 1
-
-paths:
-  data_dir: /global/cfs/projectdirs/atlas/joshua/ttHCP/ntuples/v02/preselection/merged_fixed/train/
-  save_dir: /pscratch/sd/j/joshuaho/physicsnemo/ttHCP/graphs/tHjb_CP_0_vs_45/
-  training_dir: ./training_tHjb_CP_0_vs_45/
-
-datasets:
-  - name: tHjb_cp_0_had
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_0_AF3_had_scaled.root
-    label: 0
-  - name: tHjb_cp_0_lep
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_0_AF3_lep_scaled.root
-    label: 0
-  - name: tHjb_cp_45_had
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_45_AF3_had_scaled.root
-    label: 1
-  - name: tHjb_cp_45_lep
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_45_AF3_lep_scaled.root
-    label: 1
-
-root_dataset:
-  ttree: output
-  dtype: torch.bfloat16
-  features:
-    # pt, eta, phi, energy, btag, charge, node_type
-    jet: [m_jet_pt, m_jet_eta, m_jet_phi, CALC_E, m_jet_PCbtag, 0, 0]
-    electron: [m_el_pt, m_el_eta, m_el_phi, CALC_E, 0, m_el_charge, 1]
-    muon: [m_mu_pt, m_mu_eta, m_mu_phi, CALC_E, 0, m_mu_charge, 2]
-    photon: [ph_pt_myy, ph_eta, ph_phi, CALC_E, 0, 0, 3]
-    met: [m_met, 0, m_met_phi, CALC_E, 0, 0, 4]
-  globals: [NUM_NODES]
-  weights: m_weightXlumi
-  tracking: []
-  step_size: 16384
-  batch_size: 16384
-  train_val_test_split: [0.5, 0.25, 0.25]
-  prebatch: 
-    enabled: True
-    chunk_size: 512

legacy/physicsnemo/configs/tHjb_CP_0_vs_90.yaml

@@ -1,87 +0,0 @@
-# ignore_header_test
-# Copyright 2023 Stanford University
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-random_seed: 2
-
-scheduler:
-  lr: 1.E-3
-  lr_decay: 1.E-3
-
-training:
-  epochs: 100
-  
-checkpoints:
-  ckpt_path: "checkpoints"
-  ckpt_name: "tHjb_CP_0_vs_90"
-
-performance:
-  amp: False
-  jit: False
-
-architecture:
-  module: models.MeshGraphNet
-  class: MeshGraphNet
-  args:
-    base_gnn:
-      input_dim_nodes: 7
-      input_dim_edges: 3
-      output_dim: 128
-      processor_size: 8
-      hidden_dim_node_encoder: 128
-      hidden_dim_edge_encoder: 128
-      hidden_dim_processor: 128
-      hidden_dim_node_decoder: 128
-    global_emb_dim: 128
-    global_feat_dim: 1
-    out_dim: 1
-
-paths:
-  data_dir: /global/cfs/projectdirs/atlas/joshua/ttHCP/ntuples/v02/preselection/merged_fixed/train/
-  save_dir: /pscratch/sd/j/joshuaho/physicsnemo/ttHCP/graphs/tHjb_CP_0_vs_90/
-  training_dir: ./tHjb_CP_0_vs_90/
-
-datasets:
-  - name: tHjb_cp_0_had
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_0_AF3_had_scaled.root
-    label: 0
-  - name: tHjb_cp_0_lep
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_0_AF3_lep_scaled.root
-    label: 0
-  - name: tHjb_cp_90_had
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_90_AF3_had_scaled.root
-    label: 1
-  - name: tHjb_cp_90_lep
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_90_AF3_lep_scaled.root
-    label: 1
-
-root_dataset:
-  ttree: output
-  dtype: torch.bfloat16
-  features:
-    # pt, eta, phi, energy, btag, charge, node_type
-    jet: [m_jet_pt, m_jet_eta, m_jet_phi, CALC_E, m_jet_PCbtag, 0, 0]
-    electron: [m_el_pt, m_el_eta, m_el_phi, CALC_E, 0, m_el_charge, 1]
-    muon: [m_mu_pt, m_mu_eta, m_mu_phi, CALC_E, 0, m_mu_charge, 2]
-    photon: [ph_pt_myy, ph_eta, ph_phi, CALC_E, 0, 0, 3]
-    met: [m_met, 0, m_met_phi, CALC_E, 0, 0, 4]
-  globals: [NUM_NODES]
-  weights: 1
-  tracking: []
-  step_size: 16384
-  batch_size: 16384
-  train_val_test_split: [0.5, 0.25, 0.25]
-  prebatch: 
-    enabled: True
-    chunk_size: 512

legacy/physicsnemo/configs/tHjb_CP_0_vs_90_edge_network.yaml

@@ -1,82 +0,0 @@
-# ignore_header_test
-# Copyright 2023 Stanford University
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-random_seed: 2
-
-scheduler:
-  lr: 1.E-3
-  lr_decay: 1.E-3
-
-training:
-  epochs: 100
-  
-checkpoints:
-  ckpt_path: "checkpoints"
-  ckpt_name: "tHjb_CP_0_vs_90_edge_network"
-
-performance:
-  amp: False
-  jit: False
-
-architecture:
-  module: models.Edge_Network
-  class: Edge_Network
-  args:
-    input_dim_nodes: 7
-    input_dim_edges: 3
-    input_dim_globals: 1
-    hid_size: 64
-    n_layers: 4
-    n_proc_steps: 4
-    out_dim: 1
-
-paths:
-  data_dir: /global/cfs/projectdirs/atlas/joshua/ttHCP/ntuples/v02/preselection/merged_fixed/train/
-  save_dir: /pscratch/sd/j/joshuaho/physicsnemo/ttHCP/graphs/tHjb_CP_0_vs_90/
-  training_dir: ./tHjb_CP_0_vs_90_edge_network/
-
-datasets:
-  - name: tHjb_cp_0_had
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_0_AF3_had_scaled.root
-    label: 0
-  - name: tHjb_cp_0_lep
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_0_AF3_lep_scaled.root
-    label: 0
-  - name: tHjb_cp_90_had
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_90_AF3_had_scaled.root
-    label: 1
-  - name: tHjb_cp_90_lep
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_90_AF3_lep_scaled.root
-    label: 1
-
-root_dataset:
-  ttree: output
-  dtype: torch.bfloat16
-  features:
-    # pt, eta, phi, energy, btag, charge, node_type
-    jet: [m_jet_pt, m_jet_eta, m_jet_phi, CALC_E, m_jet_PCbtag, 0, 0]
-    electron: [m_el_pt, m_el_eta, m_el_phi, CALC_E, 0, m_el_charge, 1]
-    muon: [m_mu_pt, m_mu_eta, m_mu_phi, CALC_E, 0, m_mu_charge, 2]
-    photon: [ph_pt_myy, ph_eta, ph_phi, CALC_E, 0, 0, 3]
-    met: [m_met, 0, m_met_phi, CALC_E, 0, 0, 4]
-  globals: [NUM_NODES]
-  weights: 1
-  tracking: []
-  step_size: 16384
-  batch_size: 16384
-  train_val_test_split: [0.5, 0.25, 0.25]
-  prebatch: 
-    enabled: True
-    chunk_size: 512

legacy/physicsnemo/configs/tHjb_CP_0_vs_90_globals.yaml

@@ -1,84 +0,0 @@
-# ignore_header_test
-# Copyright 2023 Stanford University
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-random_seed: 2
-
-scheduler:
-  lr: 1.E-3
-  lr_decay: 1.E-3
-
-training:
-  epochs: 100
-  
-checkpoints:
-  ckpt_path: "checkpoints"
-  ckpt_name: "tHjb_CP_0_vs_90_globals"
-
-performance:
-  amp: False
-  jit: False
-
-architecture:
-  base_gnn:
-    input_dim_nodes: 7
-    input_dim_edges: 3
-    output_dim: 128
-    processor_size: 8
-    hidden_dim_node_encoder: 128
-    hidden_dim_edge_encoder: 128
-    hidden_dim_processor: 128
-    hidden_dim_node_decoder: 128
-  global_emb_dim: 128
-  global_feat_dim: 5
-  out_dim: 1
-
-paths:
-  data_dir: /global/cfs/projectdirs/atlas/joshua/ttHCP/ntuples/v02/preselection/merged_fixed/train/
-  save_dir: /pscratch/sd/j/joshuaho/physicsnemo/ttHCP/graphs/tHjb_CP_0_vs_90_globals/
-  training_dir: ./tHjb_CP_0_vs_90_globals/
-
-datasets:
-  - name: tHjb_cp_0_had
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_0_AF3_had_scaled.root
-    label: 0
-  - name: tHjb_cp_0_lep
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_0_AF3_lep_scaled.root
-    label: 0
-  - name: tHjb_cp_90_had
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_90_AF3_had_scaled.root
-    label: 1
-  - name: tHjb_cp_90_lep
-    load_path: ${paths.data_dir}/merged_aMCPy8_tHjb125_CP_90_AF3_lep_scaled.root
-    label: 1
-
-root_dataset:
-  ttree: output
-  dtype: torch.bfloat16
-  features:
-    # pt, eta, phi, energy, btag, charge, node_type
-    jet: [m_jet_pt, m_jet_eta, m_jet_phi, CALC_E, m_jet_PCbtag, 0, 0]
-    electron: [m_el_pt, m_el_eta, m_el_phi, CALC_E, 0, m_el_charge, 1]
-    muon: [m_mu_pt, m_mu_eta, m_mu_phi, CALC_E, 0, m_mu_charge, 2]
-    photon: [ph_pt_myy, ph_eta, ph_phi, CALC_E, 0, 0, 3]
-    met: [m_met, 0, m_met_phi, CALC_E, 0, 0, 4]
-  globals: [NUM_NODES, eta_H, pt_H, eta_recotop1, pT_recotop1]
-  weights: 1
-  tracking: []
-  step_size: 16384
-  batch_size: 16384
-  train_val_test_split: [0.5, 0.25, 0.25]
-  prebatch: 
-    enabled: True
-    chunk_size: 512

legacy/physicsnemo/dataset/Dataset.py

@@ -1,243 +0,0 @@
-import os
-import uproot
-import dgl
-import torch
-import numpy as np
-from omegaconf import DictConfig
-from typing import List
-from concurrent.futures import ProcessPoolExecutor, as_completed
-from tqdm import tqdm
-
-from dataset import GraphBuilder
-from dataset import Graphs
-from dataset import Normalization
-
-from dgl.dataloading import GraphDataLoader
-
-class Dataset:
-    def __init__(
-        self,
-        name: str,
-        label: int,
-        load_path: str,
-        save_path: str,
-        dtype: torch.dtype,
-        device: str,
-        cfg: DictConfig
-    ):
-        self.name = name
-        self.label = label
-        self.load_path = load_path
-        self.save_path = save_path
-        self.dtype = dtype
-        self.data = None
-        self.device = device
-
-        self.ttree = cfg.ttree
-        self.features = cfg.features
-        self.weights = cfg.weights
-        self.globals = cfg.globals
-        self.tracking = cfg.tracking
-        self.step_size = cfg.step_size
-        self.batch_size = cfg.batch_size
-
-        self.prebatch = cfg.get('prebatch', {'enabled': False})
-
-        self.train_val_test_split = cfg.train_val_test_split
-        assert np.sum(self.train_val_test_split) == 1, "train_val_test_split must sum to 1"
-
-        print(f"initializing dataset {name} with dtype {self.dtype}")
-
-    def get_branches(self) -> List[str]:
-        node_branches = [
-            branches
-            for particle in self.features.values()
-            for branches in particle
-            if isinstance(branches, str) and (branches != "CALC_E" or branches != "NUM_NODES")
-        ]
-        global_branches = [x for x in self.globals if isinstance(x, str)]
-        weight_branch = [self.weights] if isinstance(self.weights, str) else []
-        tracking_branches = [x for x in self.tracking if isinstance(x, str)]
-        label_branch = [self.label] if isinstance(self.label, str) else []
-
-        return node_branches + global_branches + weight_branch + tracking_branches + label_branch
-
-    def process(self):
-        branches = self.get_branches()
-        with uproot.open(f"{self.load_path}:{self.ttree}") as tree:
-            available_branches = set(tree.keys())
-            num_entries = tree.num_entries
-
-        print(f"getting branches: {branches}")
-        
-        num_cpus = os.cpu_count()
-        total_chunks = np.ceil(num_entries / self.step_size)
-
-        with ProcessPoolExecutor(max_workers=num_cpus) as executor:
-            futures = []
-
-            with tqdm(
-                uproot.iterate(
-                        f"{self.load_path}:{self.ttree}",
-                        expressions=[b for b in branches if b in available_branches],
-                        step_size=self.step_size,
-                        library="ak"
-                    ),
-                desc="loading root file",
-                total=total_chunks,
-                position=0,
-                leave=True
-            ) as pbar:
-                
-                for chunk_id, arrays in enumerate(pbar):
-
-                    cfg = GraphBuilder.ChunkConfig(
-                        name=self.name,
-                        label=self.label,
-                        chunk_id=chunk_id,
-                        batch_size=self.batch_size,
-                        arrays=arrays,
-                        features=self.features,
-                        globals=self.globals,
-                        tracking=self.tracking,
-                        weights=self.weights,
-                        branches=branches,
-                        dtype=self.dtype,
-                        save_path=self.save_path,
-                        prebatch = self.prebatch,
-                    )
-                    
-                    futures.append(executor.submit(GraphBuilder.process_chunk, cfg))
-
-        for idx, future in enumerate(as_completed(futures)):
-            try:
-                future.result()
-            except Exception as e:
-                import traceback
-                print(f"exception in chunk: {idx}")
-                traceback.print_exception(type(e), e, e.__traceback__)
-        return
-    
-    def load(self):
-        with uproot.open(f"{self.load_path}:{self.ttree}") as tree:
-            num_entries = tree.num_entries
-        total_chunks = int(np.ceil(num_entries / self.step_size))
-        
-        chunk_files = [f"{self.save_path}/{self.name}_{chunk_id:04d}.bin" for chunk_id in range(total_chunks)]
-        if not all(os.path.exists(f) for f in chunk_files):
-            print("graphs not found. processing root file...")
-            self.process()
-
-        graph_tuple_list = []
-
-        for chunk_id, f in enumerate(chunk_files):
-            if chunk_id < total_chunks - 1:
-                if (self.prebatch.enabled):
-                    n_graphs = self.step_size // self.prebatch.chunk_size
-                else:
-                    n_graphs = self.step_size
-            else:
-                if (self.prebatch.enabled):
-                    n_graphs = (num_entries - self.step_size * (total_chunks - 1)) // self.prebatch.chunk_size + 1
-                else:
-                    n_graphs = num_entries - self.step_size * (total_chunks - 1)
-            graph_tuple_list.extend((f, idx) for idx in range(n_graphs))
-        
-        split = self.train_val_test_split
-        n_total = len(graph_tuple_list)
-        n_train = int(split[0] * n_total)
-        n_val = int(split[1] * n_total)
-
-        train_tuples = graph_tuple_list[:n_train]
-        val_tuples   = graph_tuple_list[n_train:n_train + n_val]
-        test_tuples  = graph_tuple_list[n_train + n_val:]
-        return train_tuples, val_tuples, test_tuples
-    
-class GraphTupleDataset:
-    def __init__(self, tuple_list, stats):
-        self.tuple_list = tuple_list
-        self.stats = stats
-        self.cache = {}
-
-    def __len__(self):
-        return len(self.tuple_list)
-
-    def __getitem__(self, idx):
-        f, graph_idx = self.tuple_list[idx]
-        if f in self.cache:
-            g = self.cache[f]
-        else: 
-            g = Graphs.load_graphs(f)
-            g.normalize(self.stats)
-            self.cache[f] = g
-        return g[graph_idx]
-
-    @staticmethod
-    def collate_fn(samples):
-        all_graphs = []
-        all_metadata = {}
-
-        # Initialize keys in all_metadata from the first sample
-        for k in samples[0][1]:
-            all_metadata[k] = []
-
-        for graph, metadata in samples:
-            all_graphs.append(graph)
-            for k, v in metadata.items():
-                all_metadata[k].append(v)
-
-        # Stack or concatenate metadata for each key
-        for k in all_metadata:
-            # If v is a tensor, stack or cat as appropriate
-            # Use torch.cat if v is already [N, ...] (e.g. labels, features)
-            # Use torch.stack if v is scalar or needs new dimension
-            try:
-                all_metadata[k] = torch.cat(all_metadata[k], dim=0)
-            except Exception:
-                all_metadata[k] = torch.stack(all_metadata[k], dim=0)
-
-        batched_graph = dgl.batch(all_graphs)
-        return batched_graph, all_metadata
-        
-def get_dataset(cfg: DictConfig, device):
-
-    all_train = []
-    all_val = []
-    all_test = []
-
-    dtype_str = getattr(cfg.root_dataset, "dtype", "torch.float32")
-    if isinstance(dtype_str, str) and dtype_str.startswith("torch."):
-        dtype = getattr(torch, dtype_str.split(".")[-1], torch.float32)
-    else:
-        dtype = torch.float32
-
-    for ds in cfg.datasets:
-        name = ds['name']
-        load_path = ds.get('load_path', f"{cfg.paths.data_dir}/{name}.root")
-        save_path = ds.get('save_path', f"{cfg.paths.save_dir}/")
-        datastet = Dataset(name, ds.get('label'), load_path, save_path, dtype, device, cfg.root_dataset)
-        train, val, test = datastet.load()
-        all_train.extend(train)
-        all_val.extend(val)
-        all_test.extend(test)
-
-    stats = Normalization.global_stats(f"{cfg.paths.save_dir}/stats/", dtype=dtype)
-
-    train_dataset = GraphTupleDataset(all_train, stats)
-    val_dataset = GraphTupleDataset(all_val, stats)
-    test_dataset = GraphTupleDataset(all_test, stats)
-
-    if (cfg.root_dataset.get('prebatch', False)):
-        batch_size = cfg.root_dataset.batch_size // cfg.root_dataset.prebatch.chunk_size
-        collate_fn = GraphTupleDataset.collate_fn
-    else:
-        batch_size = cfg.root_dataset.batch_size
-        collate_fn = None
-    
-    train_loader = GraphDataLoader(train_dataset, batch_size=batch_size, shuffle=True, pin_memory=True, num_workers=5, drop_last=False, collate_fn=collate_fn)
-    val_loader   = GraphDataLoader(val_dataset, batch_size=batch_size, shuffle=False, pin_memory=True, num_workers=5, drop_last=False, collate_fn=collate_fn)
-    test_loader  = GraphDataLoader(test_dataset, batch_size=batch_size, shuffle=False, pin_memory=True, num_workers=0, drop_last=False, collate_fn=collate_fn)
-
-    print("all data loaded successfully")
-    print(f"train: {len(train_dataset)}, val: {len(val_dataset)}, test: {len(test_dataset)}")
-    return train_loader, val_loader, test_loader

legacy/physicsnemo/dataset/GraphBuilder.py

@@ -1,162 +0,0 @@
-import dgl
-import torch
-import numpy as np
-import awkward as ak
-from dataclasses import dataclass
-from typing import List, Any, Union
-
-from dataset.Graphs import Graphs, save_graphs
-from dataset import Normalization
-
-@dataclass
-class ChunkConfig:
-    name:           str
-    label:          Union[str, int]
-    chunk_id:       int
-    batch_size:     int
-    arrays:         List[Any]
-    features:       List[Any]
-    globals:        List[Any]
-    weights:        Union[str, float]
-    tracking:       List[Any]
-    branches:       List[Any]
-    dtype:          torch.dtype
-    save_path:      str
-    prebatch:       dict
-
-def process_chunk(cfg: ChunkConfig):
-    # Collect everything as lists first
-    graph_list = []
-    meta_dict = {
-        'globals': [],
-        'label': [],
-        'weight': [],
-        'tracking': [],
-        'batch_num_nodes': [],
-        'batch_num_edges': [],
-    }
-
-    for i in range(len(cfg.arrays)):
-        g, meta = process_single_entry(cfg, i)
-        graph_list.append(g)
-        for k in meta_dict:
-            meta_dict[k].append(meta[k])
-
-    # Stack all metadata fields into tensors
-    for k in meta_dict:
-        meta_dict[k] = torch.stack(meta_dict[k])
-
-    graphs = Graphs(graphs=graph_list, metadata=meta_dict)
-    Normalization.save_stats(graphs, f"{cfg.save_path}/stats/{cfg.name}_{cfg.chunk_id:04d}.json")
-
-    if getattr(cfg.prebatch, "enabled", False):
-        graphs.shuffle()
-        graphs.batch(cfg.prebatch["chunk_size"])
-
-    save_graphs(graphs, f"{cfg.save_path}/{cfg.name}_{cfg.chunk_id:04d}.bin")
-
-def process_single_entry(cfg, i):
-    # 1) node features
-    node_features: List[torch.Tensor] = []
-
-    for particle, branch_list in cfg.features.items():
-        feature_tensors: List[torch.Tensor] = []
-        for branch in branch_list:
-            if branch == "CALC_E":
-                pT  = feature_tensors[0]
-                eta = feature_tensors[1]
-                val = pT * torch.cosh(eta)
-            elif isinstance(branch, str):
-                arr = cfg.arrays[branch][i]
-                val = torch.from_numpy(ak.to_numpy(arr)).to(cfg.dtype)
-            else:
-                length = feature_tensors[0].shape[0]
-                val = torch.full((length,), float(branch), dtype=cfg.dtype)
-            feature_tensors.append(val)
-
-        if feature_tensors and feature_tensors[0].numel() > 0:
-            block = torch.stack(feature_tensors, dim=1)
-            node_features.append(block)
-
-    node_features = torch.cat(node_features, dim=0) if node_features else torch.empty((0, len(cfg.features)), dtype=cfg.dtype)
-
-    # 2) global features
-    global_feat_list: List[torch.Tensor] = []
-    for b in cfg.globals:
-        if b == "NUM_NODES":
-            global_feat_list.append(torch.tensor([len(node_features)], dtype=cfg.dtype))
-        else:
-            arr = cfg.arrays[b][i]
-            global_feat_list.append(torch.from_numpy(ak.to_numpy(arr)).to(cfg.dtype))
-    global_feat = torch.cat(global_feat_list, dim=0) if global_feat_list else torch.zeros((1,), dtype=cfg.dtype)
-
-    # 3) tracking
-    tracking_list: List[torch.Tensor] = []
-    for b in cfg.tracking:
-        arr = cfg.arrays[b][i]
-        tracking_list.append(torch.from_numpy(ak.to_numpy(arr)).to(cfg.dtype))
-    tracking = torch.cat(tracking_list, dim=0) if tracking_list else torch.zeros((1,), dtype=cfg.dtype)
-
-    # 4) weight
-    weight = float(cfg.arrays[cfg.weights][i]) if isinstance(cfg.weights, str) else cfg.weights
-    weight = torch.tensor(weight, dtype=cfg.dtype)
-
-    # 5) label
-    label = float(cfg.arrays[cfg.label][i]) if isinstance(cfg.label, str) else cfg.label
-    label = torch.tensor(label, dtype=cfg.dtype)
-
-    # 6) make the DGLGraph
-    g = make_graph(node_features, dtype=cfg.dtype)
-
-    # 7) batch_num_nodes and batch_num_edges
-    batch_num_nodes = g.batch_num_nodes()
-    batch_num_edges = g.batch_num_edges()
-
-    meta = {
-        'globals': global_feat,
-        'label': label,
-        'weight': weight,
-        'tracking': tracking,
-        'batch_num_nodes': batch_num_nodes,
-        'batch_num_edges': batch_num_edges,
-    }
-    return g, meta
-
-src_dst_cache = {}
-def get_src_dst(num_nodes):
-    if num_nodes not in src_dst_cache:
-        src, dst = torch.meshgrid(torch.arange(num_nodes), torch.arange(num_nodes), indexing='ij')
-        src_dst_cache[num_nodes] = (src.flatten(), dst.flatten())
-    return src_dst_cache[num_nodes]
-
-@torch.jit.script
-def compute_edge_features(eta, phi, src, dst):
-    deta = eta[src] - eta[dst]
-    dphi = phi[src] - phi[dst]
-    dphi = torch.remainder(dphi + np.pi, 2 * np.pi) - np.pi
-    dR = torch.sqrt(deta ** 2 + dphi ** 2)
-    edge_features = torch.stack([dR, deta, dphi], dim=1)
-    return edge_features
-
-def make_graph(node_features: torch.tensor, dtype=torch.float32):
-
-    num_nodes = node_features.shape[0]
-    if num_nodes == 0:
-        g = dgl.graph(([], []))
-        g.ndata['features'] = node_features
-        g.edata['features'] = torch.empty((0, 3), dtype=dtype)
-        g.globals = torch.tensor([0], dtype=dtype)
-        return g
-    
-    src, dst = get_src_dst(num_nodes)
-    src = src.flatten()
-    dst = dst.flatten()
-    g = dgl.graph((src, dst))
-    g.ndata['features'] = node_features
-
-    eta = node_features[:, 1]
-    phi = node_features[:, 2]
-    edge_features = compute_edge_features(eta, phi, src, dst)
-    g.edata['features'] = edge_features
-
-    return g

legacy/physicsnemo/dataset/Graphs.py

@@ -1,88 +0,0 @@
-import dgl
-import torch
-from dataclasses import dataclass, field
-from typing import List, Dict
-
-@dataclass
-class Graphs:
-    graphs: List[dgl.DGLGraph]
-    metadata: Dict[str, torch.Tensor]
-
-    def __len__(self):
-        return len(self.graphs)
-
-    def __getitem__(self, idx):
-        meta = {k: v[idx] for k, v in self.metadata.items()}
-        return self.graphs[idx], meta
-
-    def shuffle(self):
-        idx = torch.randperm(len(self.graphs))
-        self.graphs = [self.graphs[i] for i in idx]
-        for k in self.metadata:
-            self.metadata[k] = self.metadata[k][idx]
-
-    def batch(self, batch_size, node_feature_dim=None, dtype=None):
-        """
-        In-place batching: after this, self.graphs is a list of batched DGLGraphs,
-        and self.metadata[k] is a tensor of shape [num_batches, batch_size, ...].
-        """
-        batched_graphs = []
-        batched_meta = {k: [] for k in self.metadata}
-        N = len(self.graphs)
-
-        # Infer node_feature_dim and dtype if not specified
-        if node_feature_dim is None and N > 0:
-            feats = self.graphs[0].ndata['features']
-            node_feature_dim = feats.shape[1] if feats.ndim > 1 else 1
-        if dtype is None and N > 0:
-            dtype = self.graphs[0].ndata['features'].dtype
-
-        for start in range(0, N, batch_size):
-            end = start + batch_size
-            batch_graphs = self.graphs[start:end]
-            batch_meta = {k: v[start:end] for k, v in self.metadata.items()}
-
-            # Padding if needed
-            pad_count = batch_size - len(batch_graphs)
-            if pad_count > 0:
-                dummy_graph = dgl.graph(([], []))
-                dummy_graph.ndata['features'] = torch.empty((0, node_feature_dim), dtype=dtype)
-                dummy_graph.edata['features'] = torch.empty((0, 3), dtype=dtype)  # assuming 3 edge features
-                batch_graphs += [dummy_graph] * pad_count
-
-                # Pad metadata with zeros
-                for k, v in batch_meta.items():
-                    shape = list(v[0].shape) if len(v) > 0 else []
-                    pad_tensor = torch.zeros([pad_count] + shape, dtype=v.dtype, device=v.device)
-                    batch_meta[k] = torch.cat([v, pad_tensor], dim=0)
-            else:
-                for k, v in batch_meta.items():
-                    batch_meta[k] = torch.stack(v, dim=0) if isinstance(v, list) else v
-
-            batched_graphs.append(dgl.batch(batch_graphs))
-            for k in batched_meta:
-                batched_meta[k].append(batch_meta[k])
-
-        # Now stack along a new axis: [num_batches, batch_size, ...]
-        for k in batched_meta:
-            self.metadata[k] = torch.stack(batched_meta[k], dim=0)
-
-        self.graphs = batched_graphs
-
-    def normalize(self, stats):
-        node_mean, node_std, _ = stats['node']
-        edge_mean, edge_std, _ = stats['edge']
-        for g in self.graphs:
-            g.ndata['features'] = (g.ndata['features'] - node_mean) / node_std
-            g.edata['features'] = (g.edata['features'] - edge_mean) / edge_std
-
-def save_graphs(graphs: Graphs, f: str):
-    meta_to_save = {k: v for k, v in graphs.metadata.items()}
-    dgl.save_graphs(f, graphs.graphs, meta_to_save)
-
-def load_graphs(f: str) -> Graphs:
-    g, meta = dgl.load_graphs(f)
-    for k in meta:
-        if not isinstance(meta[k], torch.Tensor):
-            meta[k] = torch.stack(meta[k])
-    return Graphs(graphs=g, metadata=meta)

legacy/physicsnemo/dataset/Normalization.py

@@ -1,144 +0,0 @@
-import torch
-import json
-import os
-from dataset.Graphs import Graphs
-from typing import List, Dict, Tuple
-
-def combine_feature_stats(chunks: List[Dict]) -> Tuple[torch.Tensor, torch.Tensor, int]:
-    """
-    Combine mean/std/count from multiple chunks using Welford's algorithm.
-    Returns combined mean, std, and total count.
-    """
-    n_total = 0
-    mean_total = None
-    M2_total = None
-
-    for chunk in chunks:
-        n_k = chunk['count']
-        if n_k == 0:
-            continue
-
-        mean_k = torch.tensor(chunk['mean'])
-        std_k = torch.tensor(chunk['std'])
-        M2_k = (std_k ** 2) * n_k
-
-        if n_total == 0:
-            mean_total = mean_k
-            M2_total = M2_k
-            n_total = n_k
-        else:
-            delta = mean_k - mean_total
-            N = n_total + n_k
-            mean_total += delta * (n_k / N)
-            M2_total += M2_k + (delta ** 2) * (n_total * n_k / N)
-            n_total = N
-
-    if n_total == 0:
-        return torch.tensor([]), torch.tensor([]), 0
-
-    std_total = torch.sqrt(M2_total / n_total)
-    return mean_total, std_total, n_total
-
-def global_stats(dirpath: str, dtype: torch.dtype) -> Dict[str, Tuple[torch.Tensor, torch.Tensor, int]]:
-    """
-    Load all JSON stats files in a directory, combine node, edge, and global stats,
-    and optionally save the combined stats as JSON to `save_path`.
-    """
-
-    combined_stats_path = os.path.join(dirpath, "global_stats.json")
-
-    if not os.path.exists(combined_stats_path):
-        stats_list = []
-        for fname in os.listdir(dirpath):
-            if fname.endswith('.json'):
-                with open(os.path.join(dirpath, fname), 'r') as f:
-                    stats_list.append(json.load(f))
-
-        node_stats = [s['node'] for s in stats_list]
-        edge_stats = [s['edge'] for s in stats_list]
-
-        combined = {
-            'node': combine_feature_stats(node_stats),
-            'edge': combine_feature_stats(edge_stats),
-        }
-
-        combined_json = {}
-        for key, (mean, std, count) in combined.items():
-            combined_json[key] = {
-                'mean': mean.tolist() if mean.numel() > 0 else [],
-                'std': std.tolist() if std.numel() > 0 else [],
-                'count': count,
-            }
-
-        with open(combined_stats_path, 'w') as f:
-            json.dump(combined_json, f, indent=4)
-
-    with open(combined_stats_path, 'r') as f:
-        combined_json = json.load(f)
-
-    def to_tensor(d):
-        mean = torch.tensor(d['mean'], dtype=dtype) if d['mean'] else torch.tensor([], dtype=dtype)
-        std = torch.tensor(d['std'], dtype=dtype) if d['std'] else torch.tensor([], dtype=dtype)
-        count = d['count']
-        return mean, std, count
-
-    return {
-        'node': to_tensor(combined_json['node']),
-        'edge': to_tensor(combined_json['edge']),
-    }
-
-def compute_stats(feats, eps=1e-6):
-    mean = feats.mean(dim=0)
-    if feats.size(0) > 1:
-        var = ((feats - mean) ** 2).mean(dim=0)
-    else:
-        var = torch.zeros_like(mean)
-    std = torch.sqrt(var)
-    std = torch.where(std < eps, torch.full_like(std, eps), std)
-
-    return mean, std
-
-def save_stats(graphs: 'Graphs', filepath: str, categorical_unique_threshold=50):
-    """
-    Compute and save normalization stats (mean, std, counts) for node and edge features.
-    Categorical features (few unique values) have normalization disabled (mean=0, std=1).
-    """
-    if len(graphs) == 0:
-        raise ValueError("No graphs to compute stats from.")
-
-    # Node and edge features
-    all_node_feats = torch.cat([g.ndata['features'] for g, _ in graphs], dim=0)
-    all_edge_feats = torch.cat([g.edata['features'] for g, _ in graphs], dim=0)
-
-    counts = {
-        'node': all_node_feats.size(0),
-        'edge': all_edge_feats.size(0),
-    }
-
-    node_mean, node_std = compute_stats(all_node_feats)
-    edge_mean, edge_std = compute_stats(all_edge_feats)
-
-    categorical_mask = torch.tensor([
-        torch.unique(all_node_feats[:, i]).numel() < categorical_unique_threshold
-        for i in range(node_mean.size(0))
-    ], dtype=torch.bool)
-    node_mean[categorical_mask] = 0.0
-    node_std[categorical_mask] = 1.0
-
-    stats = {
-        'node': {
-            'mean': node_mean.tolist(),
-            'std': node_std.tolist(),
-            'count': counts['node'],
-        },
-        'edge': {
-            'mean': edge_mean.tolist(),
-            'std': edge_std.tolist(),
-            'count': counts['edge'],
-        },
-    }
-
-    os.makedirs(os.path.dirname(filepath), exist_ok=True)
-
-    with open(filepath, 'w') as f:
-        json.dump(stats, f, indent=4)

legacy/physicsnemo/metrics.py

@@ -1,110 +0,0 @@
-import torch
-import numpy as np
-import torch.nn.functional as F
-
-def bce(input, target, weights=None):
-
-    if input.shape != target.shape:
-        if input.shape[-1] == 1 and input.shape[:-1] == target.shape:
-            input = input.squeeze(-1)
-        elif target.shape[-1] == 1 and target.shape[:-1] == input.shape:
-            target = target.squeeze(-1)
-
-    loss = F.binary_cross_entropy_with_logits(input, target, reduction='none')
-    return torch.mean(loss)
-
-def weighted_bce(input, target, weights=None):
-    """
-    Compute a weighted and label-normalized binary cross entropy (BCE) loss.
-
-    For each unique label in the target tensor, the BCE loss is computed and weighted,
-    then normalized by the sum of weights for that label. The final loss is the mean
-    of these per-label normalized losses.
-
-    Args:
-        input (Tensor): Predicted logits of shape (N, ...).
-        target (Tensor): Ground truth labels of shape (N, ...), with discrete label values.
-        weights (Tensor or None): Optional tensor of per-sample weights, same shape as input/target.
-
-    Returns:
-        Tensor: Scalar tensor representing the normalized weighted BCE loss.
-    """
-
-    if input.shape != target.shape:
-        if input.shape[-1] == 1 and input.shape[:-1] == target.shape:
-            input = input.squeeze(-1)
-        elif target.shape[-1] == 1 and target.shape[:-1] == input.shape:
-            target = target.squeeze(-1)
-
-    # Compute per-element BCE loss (no reduction)
-    loss = F.binary_cross_entropy_with_logits(input, target, reduction='none')
-
-    # If weights not provided, use ones
-    if weights is None:
-        weights = torch.ones_like(loss)
-
-    unique_labels = torch.unique(target)
-    normalized_losses = []
-    for label in unique_labels:
-        label_mask = (target == label)  # This will be a bool tensor
-        # Defensive: make sure mask is bool
-        if label_mask.dtype != torch.bool:
-            label_mask = label_mask.bool()
-        label_weights = weights[label_mask]
-        label_losses = loss[label_mask]
-        weight_sum = label_weights.sum()
-        if weight_sum > 0:
-            label_loss = (label_weights * label_losses).sum() / weight_sum
-            normalized_losses.append(label_loss)
-
-    if normalized_losses:
-        return torch.stack(normalized_losses).mean()
-    else:
-        return torch.tensor(0.0, device=input.device)
-    
-
-def roc_auc_score(classes : np.ndarray,
-               predictions : np.ndarray,
-               weights : np.ndarray = None) -> float:
-    """
-    Calculating ROC AUC score as the probability of correct ordering
-    """
-
-    if weights is None:
-        weights = np.ones_like(predictions)
-
-    assert len(classes) == len(predictions) == len(weights)
-    assert classes.ndim == predictions.ndim == weights.ndim == 1
-    class0, class1 = sorted(np.unique(classes))
-
-    data = np.empty(
-            shape=len(classes),
-            dtype=[('c', classes.dtype),
-                   ('p', predictions.dtype),
-                   ('w', weights.dtype)]
-        )
-    data['c'], data['p'], data['w'] = classes, predictions, weights
-
-    data = data[np.argsort(data['c'])]
-    data = data[np.argsort(data['p'], kind='mergesort')] # here we're relying on stability as we need class orders preserved
-
-    correction = 0.
-    # mask1 - bool mask to highlight collision areas
-    # mask2 - bool mask with collision areas' start points
-    mask1 = np.empty(len(data), dtype=bool)
-    mask2 = np.empty(len(data), dtype=bool)
-    mask1[0] = mask2[-1] = False
-    mask1[1:] = data['p'][1:] == data['p'][:-1]
-    if mask1.any():
-        mask2[:-1] = ~mask1[:-1] & mask1[1:]
-        mask1[:-1] |= mask1[1:]
-        ids, = mask2.nonzero()
-        correction = sum([((dsplit['c'] == class0) * dsplit['w'] * msplit).sum() * 
-                          ((dsplit['c'] == class1) * dsplit['w'] * msplit).sum()
-                          for dsplit, msplit in zip(np.split(data, ids), np.split(mask1, ids))]) * 0.5
- 
-    weights_0 = data['w'] * (data['c'] == class0)
-    weights_1 = data['w'] * (data['c'] == class1)
-    cumsum_0 = weights_0.cumsum()
-
-    return ((cumsum_0 * weights_1).sum() - correction) / (weights_1.sum() * cumsum_0[-1])

legacy/physicsnemo/models/Edge_Network.py

@@ -1,72 +0,0 @@
-import torch
-import torch.nn as nn
-import dgl
-
-from models import utils
-
-class Edge_Network(nn.Module):
-    def __init__(self, cfg):
-        super().__init__()
-        hid_size = cfg.hid_size
-        n_layers = cfg.n_layers
-        self.n_proc_steps = cfg.n_proc_steps
-
-        #encoder
-        self.node_encoder = utils.Make_MLP(cfg.input_dim_nodes, hid_size, hid_size, n_layers)
-        self.edge_encoder = utils.Make_MLP(cfg.input_dim_edges, hid_size, hid_size, n_layers)
-        self.global_encoder = utils.Make_MLP(cfg.input_dim_globals, hid_size, hid_size, n_layers)
-
-        #GNN
-        self.node_update = utils.Make_MLP(3*hid_size, hid_size, hid_size, n_layers)
-        self.edge_update = utils.Make_MLP(4*hid_size, hid_size, hid_size, n_layers)
-        self.global_update = utils.Make_MLP(3*hid_size, hid_size, hid_size, n_layers)
-
-        #decoder
-        self.global_decoder = utils.Make_MLP(hid_size, hid_size, hid_size, n_layers)
-        self.classify = nn.Linear(hid_size, cfg.out_dim)
-
-    def forward(self, node_feats, edge_feats, global_feats, batched_graph, metadata={}):
-        # encoders
-        batched_graph.ndata['h'] = self.node_encoder(node_feats)
-        batched_graph.edata['e'] = self.edge_encoder(edge_feats)
-
-        if global_feats.ndim == 3:
-            global_feats = global_feats.view(-1, global_feats.shape[-1])
-        h_global = self.global_encoder(global_feats)
-
-        # message passing
-        for _ in range(self.n_proc_steps):
-            batched_graph.apply_edges(dgl.function.copy_u('h', 'm_u'))
-            batched_graph.apply_edges(utils.copy_v)
-
-            # edge update
-            edge_inputs = torch.cat([
-                batched_graph.edata['e'],
-                batched_graph.edata['m_u'],
-                batched_graph.edata['m_v'],
-                utils.broadcast_global_to_edges(h_global, edge_split=metadata.get("batch_num_edges", None))
-            ], dim=1)
-            batched_graph.edata['e'] = self.edge_update(edge_inputs)
-
-            # node update
-            batched_graph.update_all(dgl.function.copy_e('e', 'm'), dgl.function.sum('m', 'h_e'))
-            node_inputs = torch.cat([
-                batched_graph.ndata['h'],
-                batched_graph.ndata['h_e'],
-                utils.broadcast_global_to_nodes(h_global, node_split=metadata.get("batch_num_nodes", None))
-            ], dim=1)
-            batched_graph.ndata['h'] = self.node_update(node_inputs)
-
-            # global update
-            graph_node_feat = utils.mean_nodes(
-                batched_graph, 'h', node_split=metadata.get("batch_num_nodes", None)
-            )
-            graph_edge_feat = utils.mean_edges(
-                batched_graph, 'e', edge_split=metadata.get("batch_num_edges", None)
-            )
-            h_global = self.global_update(torch.cat([h_global, graph_node_feat, graph_edge_feat], dim=1))
-
-        h_global = self.global_decoder(h_global)
-        out = self.classify(h_global)
-        return out
-

legacy/physicsnemo/models/MeshGraphNet.py

@@ -1,51 +0,0 @@
-import torch
-import torch.nn as nn
-import dgl
-
-from models import utils
-
-# Import the PhysicsNemo MeshGraphNet model
-from physicsnemo.models.meshgraphnet import MeshGraphNet as PhysicsNemoMeshGraphNet
-
-class MeshGraphNet(nn.Module):
-    def __init__(self, cfg):
-        super().__init__()
-        base_gnn_cfg = cfg.base_gnn
-        self.base_gnn = PhysicsNemoMeshGraphNet(**base_gnn_cfg)
-
-        self.global_mlp = nn.Sequential(
-            nn.Linear(cfg.global_feat_dim, cfg.global_emb_dim),
-            nn.ReLU(),
-        )
-
-        self.mlp = nn.Linear(
-            base_gnn_cfg['output_dim'] + base_gnn_cfg['input_dim_edges'] + cfg.global_emb_dim,
-            cfg.out_dim
-        )
-    
-    def forward(self, node_feats, edge_feats, global_feats, batched_graph, metadata={}):
-        """
-        node_feats: [total_num_nodes, node_feat_dim]
-        edge_feats: [total_num_edges, edge_feat_dim]
-        global_feats: [num_graphs, global_feat_dim]
-        batched_graph: DGLGraph, representing the collection of graphs in a batch
-        metadata: dict, may contain 'batch_num_nodes', 'batch_num_edges', etc.
-        Returns:
-            graph_pred: [num_graphs, out_dim]
-        """
-        node_pred = self.base_gnn(node_feats, edge_feats, batched_graph)
-        batched_graph.ndata['h'] = node_pred
-        batched_graph.edata['e'] = edge_feats
-
-        graph_node_feat = utils.mean_nodes(batched_graph, 'h', node_split=metadata.get("batch_num_nodes", None))
-        graph_edge_feat = utils.mean_edges(batched_graph, 'e', edge_split=metadata.get("batch_num_edges", None))
-
-        # Flatten global_feats if needed
-        if global_feats.ndim == 3:
-            global_feats = global_feats.view(-1, global_feats.shape[-1])
-        global_emb = self.global_mlp(global_feats)  # [num_graphs, global_emb_dim]
-
-        combined_feat = torch.cat([graph_node_feat, graph_edge_feat, global_emb], dim=-1)
-        graph_pred = self.mlp(combined_feat)
-        return graph_pred
-    

legacy/physicsnemo/models/utils.py

@@ -1,135 +0,0 @@
-import torch
-import torch.nn as nn
-import dgl
-
-def mean_nodes(batched_graph, feat_key='h', op='mean', node_split=None):
-    """
-    Aggregates node features per disjoint graph in a batched DGLGraph.
-
-    Args:
-        batched_graph: DGLGraph
-        feat_key: str, node feature key
-        op: 'mean', 'sum', or 'max'
-        node_split: 1D tensor or list of ints (num nodes per graph)
-
-    Returns:
-        Tensor of shape [num_graphs, node_feat_dim]
-    """
-    h = batched_graph.ndata[feat_key]
-    if node_split is None or len(node_split) == 0:
-        if op == 'mean':
-            return dgl.mean_nodes(batched_graph, feat_key)
-        elif op == 'sum':
-            return dgl.sum_nodes(batched_graph, feat_key)
-        elif op == 'max':
-            return dgl.max_nodes(batched_graph, feat_key)
-        else:
-            raise ValueError(f"Unknown op: {op}")
-    else:
-        # Ensure node_split is a flat list of ints
-        if isinstance(node_split, torch.Tensor):
-            splits = node_split.view(-1).tolist()
-        else:
-            splits = [int(x) for x in node_split]
-        chunks = torch.split(h, splits, dim=0)
-        if op == 'mean':
-            out = torch.stack([chunk.mean(0) if chunk.shape[0] > 0 else torch.zeros_like(h[0]) for chunk in chunks])
-        elif op == 'sum':
-            out = torch.stack([chunk.sum(0) if chunk.shape[0] > 0 else torch.zeros_like(h[0]) for chunk in chunks])
-        elif op == 'max':
-            out = torch.stack([chunk.max(0).values if chunk.shape[0] > 0 else torch.zeros_like(h[0]) for chunk in chunks])
-        else:
-            raise ValueError(f"Unknown op: {op}")
-        return out
-    
-def mean_edges(batched_graph, feat_key='e', op='mean', edge_split=None):
-    """
-    Aggregates edge features per disjoint graph in a batched DGLGraph.
-
-    Args:
-        batched_graph: DGLGraph
-        feat_key: str, edge feature key
-        op: 'mean', 'sum', or 'max'
-        edge_split: 1D tensor or list of ints (num edges per graph)
-
-    Returns:
-        Tensor of shape [num_graphs, edge_feat_dim]
-    """
-    e = batched_graph.edata[feat_key]
-    if edge_split is None or len(edge_split) == 0:
-        if op == 'mean':
-            return dgl.mean_edges(batched_graph, feat_key)
-        elif op == 'sum':
-            return dgl.sum_edges(batched_graph, feat_key)
-        elif op == 'max':
-            return dgl.max_edges(batched_graph, feat_key)
-        else:
-            raise ValueError(f"Unknown op: {op}")
-    else:
-        # Ensure edge_split is a flat list of ints
-        if isinstance(edge_split, torch.Tensor):
-            splits = edge_split.view(-1).tolist()
-        else:
-            splits = [int(x) for x in edge_split]
-        chunks = torch.split(e, splits, dim=0)
-        if op == 'mean':
-            out = torch.stack([chunk.mean(0) if chunk.shape[0] > 0 else torch.zeros_like(e[0]) for chunk in chunks])
-        elif op == 'sum':
-            out = torch.stack([chunk.sum(0) if chunk.shape[0] > 0 else torch.zeros_like(e[0]) for chunk in chunks])
-        elif op == 'max':
-            out = torch.stack([chunk.max(0).values if chunk.shape[0] > 0 else torch.zeros_like(e[0]) for chunk in chunks])
-        else:
-            raise ValueError(f"Unknown op: {op}")
-        return out
-    
-def Make_SLP(in_size, out_size, activation = nn.ReLU, dropout = 0):
-    layers = []
-    layers.append(nn.Linear(in_size, out_size))
-    layers.append(activation())
-    layers.append(nn.Dropout(dropout))
-    return layers
-
-def Make_MLP(in_size, hid_size, out_size, n_layers, activation = nn.ReLU, dropout = 0):
-    layers = []
-    if n_layers > 1:
-        layers += Make_SLP(in_size, hid_size, activation, dropout)
-        for i in range(n_layers-2):
-            layers += Make_SLP(hid_size, hid_size, activation, dropout)
-        layers += Make_SLP(hid_size, out_size, activation, dropout)
-    else:
-        layers += Make_SLP(in_size, out_size, activation, dropout)
-    layers.append(torch.nn.LayerNorm(out_size))
-    return nn.Sequential(*layers)
-
-def broadcast_global_to_nodes(globals, node_split):
-    """
-    globals: [num_graphs, global_dim]
-    node_split: list/1D tensor of length num_graphs, number of nodes per graph
-    Returns: [total_num_nodes, global_dim]
-    """
-    if node_split is None:
-        raise ValueError("node_split must be provided")
-    if not torch.is_tensor(node_split):
-        node_split = torch.tensor(node_split, dtype=torch.long, device=globals.device)
-    else:
-        node_split = node_split.to(device=globals.device, dtype=torch.long)
-    node_split = node_split.flatten()
-    return torch.repeat_interleave(globals, node_split, dim=0)
-
-def broadcast_global_to_edges(globals, edge_split):
-    """
-    globals: [num_graphs, global_dim] (on CUDA or CPU)
-    edge_split: list/1D tensor of length num_graphs, number of edges per graph (CPU or CUDA)
-    Returns: [total_num_edges, global_dim]
-    """
-    if edge_split is None:
-        raise ValueError("edge_split must be provided")
-    if not torch.is_tensor(edge_split):
-        edge_split = torch.tensor(edge_split, dtype=torch.long, device=globals.device)
-    else:
-        edge_split = edge_split.to(device=globals.device, dtype=torch.long)
-    edge_split = edge_split.flatten()
-    return torch.repeat_interleave(globals, edge_split, dim=0)
-
-def copy_v(edges):
-    return {'m_v': edges.dst['h']}

legacy/physicsnemo/setup/Dockerfile

@@ -1,23 +0,0 @@
-FROM nvcr.io/nvidia/physicsnemo/physicsnemo:25.06
-
-WORKDIR /global/cfs/projectdirs/atlas/joshua/GNN4Colliders
-
-LABEL maintainer.name="Joshua Ho"
-LABEL maintainer.email="ho22joshua@berkeley.edu"
-
-ENV LANG=C.UTF-8
-
-# Install system dependencies: vim, OpenMPI, and build tools
-RUN apt-get update -qq \
- && apt-get install -y --no-install-recommends \
-    wget lsb-release gnupg software-properties-common \
-    vim \
-    g++-11 gcc-11 libstdc++-11-dev \
-    openmpi-bin openmpi-common libopenmpi-dev \
- && rm -rf /var/lib/apt/lists/*
-
-# Install Python packages: mpi4py and jupyter
-RUN pip install --no-cache-dir mpi4py jupyter uproot
-
-# (Optional) Expose Jupyter port
-EXPOSE 8888

legacy/physicsnemo/setup/build_image.sh

@@ -1,4 +0,0 @@
-tag=$1
-echo $tag
-podman-hpc build -t joshuaho/nemo:$tag --platform linux/amd64 .
-podman-hpc migrate joshuaho/nemo:$tag

legacy/physicsnemo/train.py

@@ -1,246 +0,0 @@
-import time, os
-
-start = time.time()
-import torch
-from torch.nn.parallel import DistributedDataParallel
-from dgl.dataloading import GraphDataLoader
-from torch.amp import GradScaler
-import numpy as np
-import hydra
-from omegaconf import DictConfig
-from physicsnemo.launch.logging import (
-    PythonLogger,
-    RankZeroLoggingWrapper,
-)
-from physicsnemo.launch.utils import load_checkpoint, save_checkpoint
-from physicsnemo.distributed.manager import DistributedManager
-
-import json
-from tqdm import tqdm
-import random
-
-import models.MeshGraphNet as MeshGraphNet
-from dataset.Dataset import get_dataset
-import metrics
-
-import utils
-
-class MGNTrainer:
-    def __init__(self, logger, cfg, dist):
-        # set device
-        self.device = dist.device
-        logger.info(f"Using {self.device} device")
-
-        start = time.time()
-        self.trainloader, self.valloader, self.testloader = get_dataset(cfg, self.device)
-        print(f"total time loading dataset: {time.time() - start:.2f} seconds")
-
-        dtype_str = getattr(cfg.root_dataset, "dtype", "torch.float32")
-        if isinstance(dtype_str, str) and dtype_str.startswith("torch."):
-            self.dtype = getattr(torch, dtype_str.split(".")[-1], torch.float32)
-        else:
-            self.dtype = torch.float32
-
-        self.model = utils.build_from_module(cfg.architecture)
-        self.model = self.model.to(dtype=self.dtype, device=self.device)
-        # num_params = sum(p.numel() for p in self.model.parameters() if p.requires_grad)
-        # print(f"Number of trainable parameters: {num_params}")
-
-        if cfg.performance.jit:
-            self.model = torch.jit.script(self.model).to(self.device)
-        else:
-            self.model = self.model.to(self.device)
-
-        # instantiate loss, optimizer, and scheduler
-        self.optimizer = torch.optim.Adam(self.model.parameters(), lr=cfg.scheduler.lr)
-        self.scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(
-            self.optimizer,
-            T_max=cfg.training.epochs,
-            eta_min=cfg.scheduler.lr * cfg.scheduler.lr_decay,
-        )
-        self.scaler = GradScaler('cuda')
-
-        # load checkpoint
-        self.epoch_init = load_checkpoint(
-            os.path.join(cfg.checkpoints.ckpt_path, cfg.checkpoints.ckpt_name),
-            models=self.model,
-            optimizer=self.optimizer,
-            scheduler=self.scheduler,
-            scaler=self.scaler,
-            device=self.device,
-        )
-
-        self.cfg = cfg
-
-    def backward(self, loss):
-        """
-        Perform backward pass.
-
-        Arguments:
-            loss: loss value.
-
-        """
-        # backward pass
-        if self.cfg.performance.amp:
-            self.scaler.scale(loss).backward()
-            self.scaler.step(self.optimizer)
-            self.scaler.update()
-        else:
-            loss.backward()
-            self.optimizer.step()
-
-    def train(self, graph, metadata):
-        """
-        Perform one training iteration over one graph. The training is performed
-        over multiple timesteps, where the number of timesteps is specified in
-        the 'stride' parameter.
-
-        Arguments:
-            graph: the desired graph.
-
-        Returns:
-            loss: loss value.
-
-        """
-        graph = graph.to(self.device, non_blocking=True)
-        globals = metadata['globals'].to(self.device, non_blocking=True)
-        label = metadata['label'].to(self.device, non_blocking=True)
-        weight =  metadata['weight'].to(self.device, non_blocking=True)
-
-        self.optimizer.zero_grad()
-        pred = self.model(graph.ndata["features"], graph.edata["features"], globals, graph, metadata)
-        loss = metrics.weighted_bce(pred, label, weights=weight)
-        self.backward(loss)
-        return loss.detach()
-
-    @torch.no_grad()
-    def eval(self):
-        """
-        Evaluate the model on one batch.
-
-        Args:
-            graph (DGLGraph): The input graph.
-            label (Tensor): The target labels.
-
-        Returns:
-            loss (Tensor): The computed loss value (scalar).
-        """
-        predictions = []
-        labels = []
-        weights = []
-
-        for graph, metadata in self.valloader:
-            
-            graph = graph.to(self.device, non_blocking=True)
-            globals = metadata['globals'].to(self.device, non_blocking=True)
-            label = metadata['label'].to(self.device, non_blocking=True)
-            weight =  metadata['weight'].to(self.device, non_blocking=True)
-            
-            pred = self.model(graph.ndata["features"], graph.edata["features"], globals, graph, metadata)
-            predictions.append(pred)
-            labels.append(label)
-            weights.append(weight)     
-
-        predictions = torch.cat(predictions, dim=0)
-        labels = torch.cat(labels, dim=0)
-        weights = torch.cat(weights, dim=0)
-
-        loss = metrics.weighted_bce(predictions, labels, weights=weights)
-        
-        # Convert logits to probabilities
-        prob = torch.sigmoid(predictions)
-
-        # Flatten to 1D arrays
-        prob_flat = prob.detach().to(torch.float32).cpu().numpy().flatten()
-        labels_flat = labels.detach().to(torch.float32).cpu().numpy().flatten()
-
-        # Calculate AUC
-        try:
-            auc = metrics.roc_auc_score(labels_flat, prob_flat)
-        except ValueError:
-            auc = float('nan')  # Not enough classes present for AUC
-
-        return loss, auc
-
-@hydra.main(version_base=None, config_path="./configs/", config_name="tHjb_CP_0_vs_45")
-def do_training(cfg: DictConfig):
-    """
-    Perform training over all graphs in the dataset.
-
-    Arguments:
-        cfg: Dictionary of parameters.
-
-    """
-    random.seed(cfg.random_seed)
-    np.random.seed(cfg.random_seed)
-    torch.manual_seed(cfg.random_seed)
-
-    # initialize distributed manager
-    DistributedManager.initialize()
-    dist = DistributedManager()
-
-    # initialize loggers
-    os.makedirs(cfg.checkpoints.ckpt_path, exist_ok=True)
-    logger = PythonLogger("main")
-    logger.file_logging(os.path.join(cfg.checkpoints.ckpt_path, "train.log"))
-
-    # initialize trainer
-    trainer = MGNTrainer(logger, cfg, dist)
-
-    if dist.distributed:
-        ddps = torch.cuda.Stream()
-        with torch.cuda.stream(ddps):
-            trainer.model = DistributedDataParallel(
-                trainer.model,
-                device_ids=[dist.local_rank],  # Set the device_id to be
-                                               # the local rank of this process on
-                                               # this node
-                output_device=dist.device,
-                broadcast_buffers=dist.broadcast_buffers,
-                find_unused_parameters=dist.find_unused_parameters,
-            )
-        torch.cuda.current_stream().wait_stream(ddps)
-
-    # training loop
-    start = time.time()
-    logger.info("Training started...")
-    for epoch in range(trainer.epoch_init, cfg.training.epochs):
-
-        # Training
-        train_loss = []
-        for graph, metadata in tqdm(trainer.trainloader, desc=f"epoch {epoch} trianing"):
-            trainer.model.train()
-            loss = trainer.train(graph, metadata)
-            train_loss.append(loss.item())
-        
-        val_loss, val_auc = trainer.eval()
-
-        train_loss = torch.tensor(train_loss).mean()
-
-        logger.info(
-            f"epoch: {epoch}, loss: {train_loss:10.3e}, val_loss: {val_loss:10.3e}, val_auc = {val_auc:10.3e}, time per epoch: {(time.time()-start):10.3e}"
-        )
-
-        # save checkpoint
-        save_checkpoint(
-            os.path.join(cfg.checkpoints.ckpt_path, cfg.checkpoints.ckpt_name),
-            models=trainer.model,
-            optimizer=trainer.optimizer,
-            scheduler=trainer.scheduler,
-            scaler=trainer.scaler,
-            epoch=epoch,
-        )
-        start = time.time()
-        trainer.scheduler.step()
-    logger.info("Training completed!")
-
-
-"""
-    Perform training over all graphs in the dataset.
-
-    Arguments:
-        cfg: Dictionary of parameters.
-
-    """
-if __name__ == "__main__":
-    do_training()

legacy/physicsnemo/utils.py

@@ -1,11 +0,0 @@
-import importlib
-from types import SimpleNamespace
-
-def build_from_module(cfg):
-    modname = cfg['module']
-    classname = cfg['class']
-    args = cfg['args']
-    module = importlib.import_module(modname)
-    model_cls = getattr(module, classname)
-    cfg_obj = SimpleNamespace(**args)
-    return model_cls(cfg_obj)

legacy/root_gnn_dgl/.codex/skills/root-gnn-dgl-data-preparation/SKILL.md

@@ -1,202 +0,0 @@
----
-name: root-gnn-dgl-data-preparation
-description: Use when the user asks to build graphs or prebatched .bin files, rerun failed data prep, verify missing graph chunks, or use scripts/prep_data.py, scripts/check_dataset_files.py, or jobs/prep_data/prep_data.sh before training in root_gnn_dgl.
----
-
-# root-gnn-dgl-data-preparation
-
-Use this skill for graph creation and graph-readiness checks before training.
-
-## Fast parallel graph creation
-
-For the Delphes baseline configs, use the bounded launcher rather than
-`jobs/prep_data/run_processing.py` or a hand-written unbounded process loop:
-
-```bash
-python jobs/prep_data/parallel_prep.py \
-  configs/delphes/*_baseline.yaml \
-  --phase raw \
-  --workers 8
-```
-
-The launcher runs one subprocess per graph chunk, with a bounded worker pool.
-The graph-building path streams ROOT entries for the requested chunk, so
-parallel workers do not each materialize the complete input ROOT file.
-Start with 8 workers per CPU session; 16 is reasonable when memory and I/O
-are healthy. Do not equate the number of workers with all available CPUs:
-filesystem contention and per-process graph memory generally become the
-bottleneck first. Set numerical-library thread counts to one before launch:
-
-```bash
-export OMP_NUM_THREADS=1
-export MKL_NUM_THREADS=1
-export OPENBLAS_NUM_THREADS=1
-export NUMEXPR_NUM_THREADS=1
-```
-
-For two independent CPU sessions, split the task list without overlap:
-
-```bash
-# Session 1
-python jobs/prep_data/parallel_prep.py configs/delphes/*_baseline.yaml \
-  --phase raw --split 1/2 --workers 16
-
-# Session 2
-python jobs/prep_data/parallel_prep.py configs/delphes/*_baseline.yaml \
-  --phase raw --split 2/2 --workers 16
-```
-
-Run prebatching only after both raw phases finish:
-
-```bash
-python jobs/prep_data/parallel_prep.py configs/delphes/*_baseline.yaml \
-  --phase shuffle --split 1/2 --workers 8 --buffer-size 1
-python jobs/prep_data/parallel_prep.py configs/delphes/*_baseline.yaml \
-  --phase shuffle --split 2/2 --workers 8 --buffer-size 1
-```
-
-`--split PART/TOTAL` partitions dataset tasks, not events. Use the same
-config glob and split value in both sessions. Run only baseline configs for
-graph creation; finetuning configs share the same processed data directories.
-
-## Primary entry point
-
-Run from the repo root:
-
-```bash
-python scripts/prep_data.py --config <config.yaml> --dataset <dataset_name> --chunk <chunk_index>
-```
-
-Use `--shuffle_mode` when you want preshuffled, prebatched graph files for training:
-
-```bash
-python scripts/prep_data.py --config <config.yaml> --dataset <dataset_name> --shuffle_mode --chunk <chunk_index>
-```
-
-## Chunk and memory semantics
-
-- `args.chunks` is the number of ordinary raw graph `.bin` files. It controls
-  the size of each graph-creation task and must match the configs used later
-  by training.
-- `shuffle_chunks` is the number of shuffled/prebatched output partitions.
-- `buffer_size` is the number of raw graph `.bin` chunks cached in memory by a
-  lazy dataset during shuffling. It does not need to be less than `chunks`,
-  but use `1` (or `2`) for memory-constrained runs.
-- Reducing `chunks` makes each raw task larger; it usually increases, rather
-  than reduces, per-worker memory. Change it consistently in baseline and
-  finetuning configs, and do not mix old cache files from a different chunk
-  layout.
-- For Delphes configs, the repository currently uses `chunks: 10` and
-  `shuffle_chunks: 10`.
-
-The old `jobs/prep_data/prep_data.sh` wrapper is sequential and previously
-ran chunk 0 twice. Prefer `parallel_prep.py`; if the wrapper is needed, it
-now runs each requested chunk once.
-
-## Recommended single-dataset run pattern
-
-- Read dataset names from `config["Datasets"]`.
-- Read the chunk count from each dataset's `args.chunks`.
-- For a single raw chunk, run the command without `--shuffle_mode`.
-- Add `--shuffle_mode` only after the ordinary graph chunks exist.
-
-Example pattern:
-
-```bash
-python scripts/prep_data.py --config <config.yaml> --dataset <dataset> --chunk 0
-python scripts/prep_data.py --config <config.yaml> --dataset <dataset> \
-  --shuffle_mode --chunk 0 --buffer_size 1
-```
-
-Use the repo wrapper when you want the standard loop:
-
-```bash
-bash jobs/prep_data/prep_data.sh <config> <dataset> <chunks> [extra_args]
-```
-
-## Important flags and caveats
-
-- `--shuffle_mode` creates the prebatched artifacts consumed by `scripts/training_script.py --preshuffle`.
-- `scripts/prep_data.py` accepts `--buffer_size` and `--shuffle_chunks` as
-  runtime overrides; changing `--shuffle_chunks` changes output filenames,
-  so update training configs before using that override for production.
-- `--drop_last` is inverted by the CLI definition: passing the flag sets `drop_last=False`.
-- Dataset configs can override training batch size during prebatching with a dataset-level `batch_size`.
-- The README says a `list index out of range` after graph saving is currently expected in some prep runs. Treat it as non-fatal if the output `.bin` files were written successfully.
-
-## Dataset selections
-
-Datasets may define event selections at the dataset level:
-
-```yaml
-Datasets:
-  signal:
-    args:
-      ...
-    selections:
-      - [n_jets, 4, ">="]
-      - "met_met_NOSYS > 30000"
-```
-
-Selection behavior:
-
-- Selections are applied during data prep before graph chunking.
-- `scripts/prep_data.py` prints a cutflow for each dataset before processing.
-- The streaming optimization applies to the no-selection Delphes path. Configs
-  with selections still use the legacy full-array selection path and should
-  be tested with one worker before parallelizing.
-- Tuple/list selections use `[branch, cut, op]`, where `op` can be `>`, `>=`, `<`, `<=`, `==`, or `!=`.
-- String selections are evaluated against loaded ROOT branches, so referenced branch names must exist.
-- Selection branches are added automatically to the branch list through `selection_branches()`.
-- Empty or omitted `selections` means all events pass.
-- If a selection references vector branches, verify the result is one boolean per event; jagged per-object masks will not index event arrays correctly.
-
-## Audit the outputs
-
-Run from the repo root:
-
-```bash
-python scripts/check_dataset_files.py --configs stats_100K/pretraining_multiclass.yaml
-```
-
-The `--configs` argument must be a comma-separated list of paths relative to `configs/`.
-
-This checker validates:
-
-- chunk files named `${dataset}_${chunk}.bin`
-- prebatched fold files named `${dataset}_prebatched_padded_${i}_n_${n_folds}_f_${foldlist}.bin`
-
-Use rerun mode to repair missing artifacts:
-
-```bash
-python scripts/check_dataset_files.py --configs stats_100K/pretraining_multiclass.yaml --rerun
-```
-
-For bulk prep over every dataset in one or more configs, use the bounded
-launcher above. Avoid the legacy bulk helper for large Delphes files because
-it can start too many full-file readers.
-
-The legacy helper is:
-
-```bash
-python jobs/prep_data/run_processing.py configs/run_3_ttH/scratch.yaml configs/run_3_ttH/finetuning.yaml
-```
-
-This calls `jobs/prep_data/prep_data.sh` for each dataset using the dataset-level `shuffle_chunks` value.
-
-Treat data prep as ready only if:
-
-- every required chunk file exists
-- every required prebatched fold file exists when training will use `--preshuffle`
-- save paths match the config
-- any post-save `IndexError` did not prevent the files from being written
-
-If stopping a Slurm run, prefer cancelling the whole allocation:
-
-```bash
-scancel "$SLURM_JOB_ID"
-```
-
-Completed `.bin` files remain on disk, but inspect files being written at the
-time of cancellation before restarting; a partially written file may exist
-and be mistaken for a valid cache.

legacy/root_gnn_dgl/.codex/skills/root-gnn-dgl-env-setup/SKILL.md

@@ -1,63 +0,0 @@
----
-name: root-gnn-dgl-env-setup
-description: Use when the user asks to set up or validate the root_gnn_dgl runtime, such as running conda setup from setup/environment.yml, setup/test_setup.py, import ROOT checks, podman-hpc image setup, or the interactive allocation scripts in jobs/ before data prep, training, or inference.
----
-
-# root-gnn-dgl-env-setup
-
-Use this skill from the repo root before any stage run.
-
-## Choose the runtime
-
-- Use the conda environment in `setup/environment.yml` for `scripts/inference.py`. The repo README says inference needs PyROOT, and the podman image does not include ROOT.
-- Use the `podman-hpc` image `joshuaho/pytorch:1.0` for training on Perlmutter when you want the containerized path.
-- For parallel inference, make sure `mpi4py` is available. The README notes it is not listed in the conda environment requirements; `setup/Dockerfile` installs it in the container image.
-
-## Conda path
-
-```bash
-cd setup
-conda env create -f environment.yml
-conda activate pytorch
-cd ..
-python setup/test_setup.py
-python -c "import ROOT"
-```
-
-Run `setup/test_setup.py` from the repo root. It appends the current working directory to `sys.path` and checks imports in `scripts`, `root_gnn_base`, and `models`.
-
-## Podman path
-
-```bash
-podman-hpc pull docker.io/joshuaho/pytorch:1.0
-```
-
-Or build locally:
-
-```bash
-cd setup
-source build_image.sh
-```
-
-The helper `setup/launch_image.sh` mounts `/pscratch/sd/j/joshuaho/` and `/global/cfs/projectdirs/atlas/joshua/` into the container and then runs the given entrypoint.
-
-## Interactive allocations
-
-- `source jobs/interactive.sh` for one shared interactive GPU node.
-- `source jobs/cpu.sh` for a CPU allocation that suits large prep loops.
-- `source jobs/salloc.sh` for a multi-node GPU allocation.
-
-## Runtime audit
-
-- Use `nvidia-smi` before training on login or interactive nodes to confirm memory availability.
-- Validate the basic repo imports with `python setup/test_setup.py`.
-- Validate PyROOT explicitly with `python -c "import ROOT"`.
-- For parallel inference, also validate `python -c "from mpi4py import MPI"`.
-- Some repo scripts hard-code NERSC-style paths under `/global/cfs/projectdirs/atlas/joshua/...`. If running elsewhere, fix those paths before assuming the environment is valid.
-
-Treat environment setup as passing only if:
-
-- imports succeed
-- the chosen runtime matches the stage you plan to run
-- required site-specific paths exist
-- GPU or CPU resources are actually available for the intended stage

legacy/root_gnn_dgl/.codex/skills/root-gnn-dgl-inference/SKILL.md

@@ -1,133 +0,0 @@
----
-name: root-gnn-dgl-inference
-description: Use when the user asks to score ROOT files, add GNN score branches, launch MPI inference, or verify inference outputs in root_gnn_dgl, including scripts/inference.py, jobs/inference/run_inference.py, ROOT branch checks, and NPZ or ROOT output audits.
----
-
-# root-gnn-dgl-inference
-
-Use this skill to score ntuples with trained models and verify the outputs.
-
-## Environment
-
-- Run inference in an environment with PyROOT available.
-- The repo README says the conda environment is required for inference because the podman training image does not include ROOT.
-
-## Entry point
-
-Run from the repo root:
-
-```bash
-python scripts/inference.py \
-  --target <input.root> \
-  --destination <output.root> \
-  --config <config.yaml> \
-  --branch_name <branch_name> \
-  --chunks 1 \
-  --chunkno 0 \
-  --write
-```
-
-## Multi-model inference
-
-The script accepts multiple configs and multiple branch names in one run:
-
-```bash
-python scripts/inference.py \
-  --target <input.root> \
-  --destination <output.root> \
-  --config config_a.yaml config_b.yaml \
-  --branch_name score_a score_b \
-  --chunks 1 \
-  --chunkno 0 \
-  --write
-```
-
-The number of configs and branch names must match.
-
-## Checkpoint selection
-
-- With the default `--ckpt -1`, the script selects the best epoch from `training.log` using `--var` and `--mode`.
-- Use `--ckpt <n>` to force a specific checkpoint.
-- If `--destination` is omitted, the script writes under `<Training_Directory>/inference/`.
-
-## Output modes
-
-- `--write` creates a new ROOT file and adds score branches.
-- Without `--write`, the script saves an `.npz` bundle containing scores, labels, and tracking info.
-- Use `--clobber` when reusing an existing destination path.
-
-## Parallel inference
-
-Use the generic MPI/local wrapper for many files or many models:
-
-```bash
-mpirun -np <num_ranks> python jobs/inference/run_inference.py \
-  --sample-config <config-with-target-datasets.yaml> \
-  --config-dir <directory-of-model-configs> \
-  --output-dir <output-directory> \
-  --write
-```
-
-Useful variants:
-
-```bash
-python jobs/inference/run_inference.py --target '<glob-or-file.root>' --config model.yaml --output-dir scores --write
-python jobs/inference/run_inference.py --sample-config samples.yaml --config model_a.yaml model_b.yaml --branch-name score_a score_b --output-dir scores --write
-python jobs/inference/run_inference.py --sample-config samples.yaml --config-dir configs/run_3_ttH --output-dir scores --write --test
-```
-
-Wrapper behavior:
-
-- `--sample-config` discovers target ROOT files from every dataset's `args.raw_dir` and `args.file_names`.
-- `--target` accepts explicit files or glob patterns.
-- `--config-dir` discovers model configs from `*.yaml`; `--config` accepts explicit config files.
-- Branch names default to each model config's `Training_Name` plus `_score`.
-- `--branch-name` may override branch names, but the count must match the config count.
-- `--test` prints the planned `scripts/inference.py` commands without running them.
-- With MPI, tasks are split by target file across ranks; without MPI it runs serially.
-- GPU assignment is local rank modulo 4 through `CUDA_VISIBLE_DEVICES`.
-
-## Repo-specific behavior
-
-- The first config's first dataset is used as the template dataset. The script rewrites `raw_dir`, `file_names`, `save_dir`, `chunks`, `process_chunks`, and optionally `tree_name` at runtime.
-- Pass `--tree <name>` if the ROOT tree name differs from the config default.
-- Chunked inference writes per-chunk outputs; merging those outputs is a separate step.
-- Job wrappers should derive the repo root from their own path; avoid adding hard-coded checkout paths.
-
-## Audit the outputs
-
-Start with basic runtime evidence if you have a log:
-
-- `Writing to file`
-- `Input entries:`
-- `Output entries:`
-- `Wrote scores to`
-- absence of `Traceback`
-
-For ROOT outputs, prefer `uproot`:
-
-```bash
-python - <<'PY'
-import numpy as np
-import uproot
-path = "<output.root>"
-branches = ["<score_branch>"]
-tree = uproot.open(path)["output"]
-print("entries", tree.num_entries)
-for branch in branches:
-    arr = tree[branch].array(library="np")
-    print(branch, len(arr), np.isnan(arr).sum(), float(np.nanmin(arr)), float(np.nanmax(arr)), float(np.nanmean(arr)))
-PY
-```
-
-Treat inference as valid only if:
-
-- the destination file exists
-- every requested score branch exists
-- output entry count matches the input tree
-- score arrays contain no NaNs
-- score arrays are not constant
-
-For multi-model inference, every branch must exist and branch statistics should usually differ unless the models are intentionally identical.
-
-Without `--write`, inspect the `.npz` keys `scores`, `labels`, and `tracking_info` and verify array lengths and NaN counts.

legacy/root_gnn_dgl/.codex/skills/root-gnn-dgl-plotting/SKILL.md

@@ -1,80 +0,0 @@
----
-name: root-gnn-dgl-plotting
-description: Use when the user asks to plot training curves, regenerate training.png, build the sweep PDF from plotting/training_performance.py, compare training runs, or extract Loss, Accuracy, Test_Loss, Test_AUC, and timing information from training.log files in root_gnn_dgl.
----
-
-# root-gnn-dgl-plotting
-
-Use this skill when the task is about plots or metrics derived from `training.log`.
-
-## Single-run plot regeneration
-
-The training script can regenerate the per-run PNG directly:
-
-```bash
-python scripts/training_script.py --config <config.yaml> --plot
-```
-
-That uses `root_gnn_base.utils.read_log()` and `root_gnn_base.utils.plot_log()` to rebuild `training.png` from `Training_Directory/training.log`.
-
-`plot_log()` produces a 2x2 figure with:
-
-- cumulative time in seconds
-- train and test loss
-- accuracy
-- test AUC
-
-Be aware that `plot_log()` fixes the accuracy axis to `(0.44, 0.56)`, which may be too narrow for some runs.
-
-## Sweep-level plotting
-
-Use the dedicated plotting script when the user wants a PDF comparing shipped sweeps:
-
-```bash
-python plotting/training_performance.py
-python plotting/training_performance.py --output <output.pdf>
-```
-
-The script currently plots two config groups:
-
-- `pretraining`
-- `higgs_production`
-
-It writes one PDF page per group and resolves each run's `Training_Directory` from its config.
-
-## What the plotting script reads from training.log
-
-`plotting/training_performance.py` parses rows that start with `Epoch` and extracts:
-
-- `Epoch`
-- `Loss`
-- `Accuracy`
-- `Test_Loss`
-- `Test_AUC`
-- `Time`
-
-It also computes cumulative time in hours.
-
-Baseline runs are drawn as lines. Non-baseline sweep variants are drawn as point clouds and labeled by the parameter change relative to the baseline.
-
-## Audit the log before plotting
-
-Treat plotting input as valid only if:
-
-- `training.log` exists
-- it contains at least one valid `Epoch ...` row
-- parsed metric arrays are finite
-- the referenced `Training_Directory` actually exists
-
-If the plotting script fails, inspect the log directly:
-
-```bash
-sed -n '1,40p' <training_dir>/training.log
-tail -n 25 <training_dir>/training.log
-```
-
-## When to use which plot path
-
-- Use `--plot` on `scripts/training_script.py` when the user wants the repo's standard per-run `training.png`.
-- Use `plotting/training_performance.py` when the user wants a cross-run PDF for the built-in sweep groups.
-- If the user wants custom comparisons outside the built-in groups, start from the parsing logic in `plotting/training_performance.py` and the metric schema in `training.log`.

legacy/root_gnn_dgl/.codex/skills/root-gnn-dgl-training/SKILL.md

@@ -1,156 +0,0 @@
----
-name: root-gnn-dgl-training
-description: Use when the user asks to train, finetune, resume, submit, queue-check, log-check, or validate training runs in root_gnn_dgl, including scripts/training_script.py, sqs, jobs/slurm/<JOBID>.out, training.log, checkpoints, and scripts/generate_multiclass_finetuning_configs.py.
----
-
-# root-gnn-dgl-training
-
-Use this skill for any training stage in the repo, from launch through monitoring and artifact review.
-
-## Run locally
-
-Run from the repo root:
-
-```bash
-python scripts/training_script.py --config <config.yaml> --preshuffle --nocompile --lazy
-```
-
-This matches the README, `run_demo.sh`, and the podman job wrappers.
-
-## Why these defaults
-
-- `--preshuffle` uses the saved prebatched graph files created during data prep.
-- `--nocompile` is recommended by the README because compiled mode requires padded graphs at prep time.
-- `--lazy` matches the common dataset classes used by the shipped configs.
-
-## Common runtime modes
-
-- `--restart` starts from scratch instead of resuming from the last checkpoint.
-- Without `--restart`, the script resumes from the last `model_epoch_<n>.pt` it finds in `Training_Directory`.
-- `--evaluate <epoch>` skips training and evaluates a specific checkpoint.
-- `--plot` regenerates `training.png` from `training.log`.
-- `--directory <suffix>` appends a suffix to `Training_Directory`.
-- `--cpu`, `--multigpu`, `--multinode`, `--statistics`, `--seed`, and `--abs` are available when needed.
-
-## Run multiple local trainings
-
-Use the single-node launcher when queueing many config files on the current node:
-
-```bash
-python jobs/training/run_parallel_trainings.py <config-or-directory> --split <K/N>
-```
-
-Useful examples:
-
-```bash
-python jobs/training/run_parallel_trainings.py configs/run_3_ttH --split 2/2
-python jobs/training/run_parallel_trainings.py configs/run_3_ttH --split 2/2 --test
-```
-
-Launcher behavior:
-
-- Discovers `.yaml` files from each target directory, or accepts explicit config files.
-- Uses fixed local GPU slots `0,1,2,3` via `CUDA_VISIBLE_DEVICES`.
-- Runs `scripts/training_script.py --config <config> --preshuffle --nocompile --lazy`.
-- Forwards unknown args to `training_script.py`, such as `--restart` or `--seed 7`.
-- Rejects configs with duplicate `Training_Directory` values.
-- `--split K/N` selects configs by deterministic index modulo `N`; use `--split 1/2` and `--split 2/2` for two complementary halves.
-- `--test` prints the launch plan without starting training.
-- Logs go under `jobs/slurm/parallel_training_logs/<timestamp>/`.
-
-## Submit on Perlmutter
-
-Prefer the podman path:
-
-```bash
-sbatch jobs/training/podman/run_job.sh <config>
-```
-
-The job wrappers derive the repo root from their own location. `jobs/training/podman/submit.sh` and `jobs/training/conda/submit.sh` accept config paths as arguments; without arguments they run their built-in default sweeps. Set `SLURM_ACCOUNT` when the cluster requires an account:
-
-```bash
-SLURM_ACCOUNT=atlas sbatch jobs/training/podman/run_job.sh configs/run_3_ttH/scratch.yaml
-bash jobs/training/podman/submit.sh configs/run_3_ttH/scratch.yaml configs/run_3_ttH/finetuning.yaml
-```
-
-For the conda wrapper, set `ROOT_GNN_CONDA_ENV` if the environment is not named `dgl`.
-
-For distributed training, pass `--multinode` and launch under an environment that sets `RANK`, `LOCAL_RANK`, and `WORLD_SIZE`.
-
-## Preconditions
-
-- If you use `--preshuffle`, run data preparation first and confirm the graph artifacts exist.
-- For finetuning configs, verify that `Model.args.pretraining_path` points to an existing checkpoint before launching training.
-- For multinode runs, pass `--multinode` and launch under the relevant distributed job environment.
-
-## Monitor queue and logs
-
-Check queue state:
-
-```bash
-sqs -u "$USER"
-sqs -u "$USER" | rg "<pattern>"
-```
-
-Useful interpretations:
-
-- `PD` means pending
-- `R` means running
-- `START_TIME N/A` with reason `Priority` means queued normally, not broken
-
-Once a job has a `JOBID`, inspect:
-
-```bash
-sed -n '1,80p' jobs/slurm/<JOBID>.out
-tail -n 80 jobs/slurm/<JOBID>.out
-rg -n "Traceback|Error|Exception|Epoch|Epoch Done|Early Termination|Done" jobs/slurm/<JOBID>.out
-```
-
-Healthy training logs usually show:
-
-- the `Executing: python -u ... scripts/training_script.py ...` line
-- dataset cache loads
-- repeated `Epoch ... | LR ... | Loss ... | Accuracy ... | Test_Loss ... | Test_AUC ... | Time ... s`
-- `Epoch Done.`
-- `Num batches trained = ...`
-- valid completion via `Done`, sometimes after `Early Termination at Epoch ...`
-
-Early stopping is a normal completion mode in this repo.
-
-## Audit training artifacts
-
-Training writes into `Training_Directory`:
-
-- `config.yaml`
-- `model_epoch_<n>.pt`
-- `model_epoch_<n>.npz`
-- `training.log`
-- `training.png`
-
-Primary checks:
-
-```bash
-sed -n '1,40p' <training_dir>/training.log
-tail -n 25 <training_dir>/training.log
-```
-
-Treat the run as healthy only if:
-
-- epoch numbers increase monotonically
-- `Loss`, `Test_Loss`, and `Test_AUC` stay finite
-- the latest logged epoch has a matching `model_epoch_<n>.pt`
-- the run produced real epoch rows rather than stopping before training started
-
-If `training.log` grows but checkpoints stop appearing, suspect a save-path or filesystem issue.
-
-Use `python plotting/training_performance.py` or the `root-gnn-dgl-plotting` skill when you want consolidated sweep-level PDFs instead of a single-run `training.png`.
-
-## Generate finetuning configs
-
-Use this when you want to derive `configs/higgs_production/multiclass_finetuning/*.yaml` from completed multiclass pretraining runs:
-
-```bash
-python scripts/generate_multiclass_finetuning_configs.py
-```
-
-Before using the generated configs, verify that the chosen best-epoch checkpoint paths still exist.

legacy/root_gnn_dgl/.codex/skills/root-gnn-dgl-workflow/SKILL.md

@@ -1,68 +0,0 @@
----
-name: root-gnn-dgl-workflow
-description: Use when the user asks to run or review the full root_gnn_dgl workflow, such as run_demo.sh, a full pretraining-to-finetuning-to-inference campaign, or a stage-by-stage pass, warning, fail audit across environment setup, data preparation, training, inference, and outputs.
----
-
-# root-gnn-dgl-workflow
-
-Use this skill when the user wants an end-to-end workflow rather than a single isolated stage.
-
-## Shipped demo
-
-Run from the repo root:
-
-```bash
-source run_demo.sh
-```
-
-The demo does:
-
-1. graph prep for multiclass pretraining
-2. multiclass pretraining
-3. graph prep for binary classification
-4. from-scratch binary training
-5. finetuned binary training
-6. inference with two output score branches
-
-## Before running the workflow
-
-- Check GPU availability with `nvidia-smi` or request an interactive node with `jobs/interactive.sh`.
-- Confirm the target data and output directories in `run_demo.sh` exist and are writable.
-- Confirm `configs/stats_100K/finetuning_ttH_CP_even_vs_odd.yaml` points at the checkpoint you actually want to finetune from.
-
-## Workflow audit order
-
-When reviewing a campaign, check stages in this order:
-
-1. environment readiness
-2. data-prep outputs
-3. training submission and queue state
-4. training logs and checkpoints
-5. inference outputs
-
-Use the retained stage skills for each check:
-
-- `root-gnn-dgl-env-setup`
-- `root-gnn-dgl-data-preparation`
-- `root-gnn-dgl-training`
-- `root-gnn-dgl-inference`
-
-## Output style
-
-Return a short status for each stage:
-
-- `pass`: evidence is consistent with a healthy stage
-- `warning`: stage likely worked but still needs a follow-up check
-- `fail`: concrete blocker or corrupted or missing artifact found
-
-Repo-specific workflow blockers:
-
-- pending jobs in `sqs` with `Priority` are waiting, not failed
-- missing prebatched `.bin` files block `--preshuffle` training
-- missing `pretraining_path` blocks finetuning
-- ROOT outputs without the requested score branches are inference failures even if the file exists
-
-## When to adapt instead of sourcing the demo
-
-- If you only want one stage, call the underlying prep, training, or inference script directly.
-- If dataset paths, branch names, or chunk counts differ, copy the command pattern from `run_demo.sh` and adjust the values instead of editing the demo in place.

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/config.yaml

@@ -1,319 +0,0 @@
-Datasets:
-  SingleT_schan:
-    args:
-      buffer_size: 11
-      chunks: 100
-      file_names: SingleT_schan.root
-      fold_var: Number
-      label: 8
-      name: SingleT_schan
-      node_branch_names: &id001
-      - - jet_pt
-        - ele_pt
-        - mu_pt
-        - ph_pt
-        - MET_met
-      - - jet_eta
-        - ele_eta
-        - mu_eta
-        - ph_eta
-        - 0
-      - - jet_phi
-        - ele_phi
-        - mu_phi
-        - ph_phi
-        - MET_phi
-      - CALC_E
-      - - jet_btag
-        - 0
-        - 0
-        - 0
-        - 0
-      - - 0
-        - ele_charge
-        - mu_charge
-        - 0
-        - 0
-      - NODE_TYPE
-      node_branch_types: &id002
-      - vector
-      - vector
-      - vector
-      - vector
-      - single
-      node_feature_scales: &id003
-      - 1e-1
-      - 1
-      - 1
-      - 1e-1
-      - 1
-      - 1
-      - 1
-      raw_dir: /global/cfs/projectdirs/atlas/joshua/root_gnn/root_gnn_dgl/data/ntuples/Hyy_pretraining/
-      save_dir: /pscratch/sd/j/joshuaho/root_gnn/root_gnn_dgl/data/processed_pretraining_multiclass_12_process/
-      tree_name: output
-      weight_var: weight
-    class: LazyDataset
-    folding: &id004
-      n_folds: 10
-      test:
-      - 0
-      - 1
-      train:
-      - 2
-      - 3
-      - 4
-      - 5
-      - 6
-      - 7
-      - 8
-      - 9
-    module: root_gnn_base.dataset
-    padding_mode: NONE
-    shuffle_chunks: 10
-  VBF:
-    args:
-      buffer_size: 11
-      chunks: 100
-      file_names: VBF_NLO_inc.root
-      fold_var: Number
-      label: 3
-      name: VBF
-      node_branch_names: *id001
-      node_branch_types: *id002
-      node_feature_scales: *id003
-      raw_dir: /global/cfs/projectdirs/atlas/joshua/root_gnn/root_gnn_dgl/data/ntuples/Hyy_pretraining/
-      save_dir: /pscratch/sd/j/joshuaho/root_gnn/root_gnn_dgl/data/processed_pretraining_multiclass_12_process/
-      tree_name: output
-      weight_var: weight
-    class: LazyDataset
-    folding: *id004
-    module: root_gnn_base.dataset
-    padding_mode: NONE
-    shuffle_chunks: 10
-  WH:
-    args:
-      buffer_size: 11
-      chunks: 100
-      file_names: WH_NLO_inc.root
-      fold_var: Number
-      label: 4
-      name: WH
-      node_branch_names: *id001
-      node_branch_types: *id002
-      node_feature_scales: *id003
-      raw_dir: /global/cfs/projectdirs/atlas/joshua/root_gnn/root_gnn_dgl/data/ntuples/Hyy_pretraining/
-      save_dir: /pscratch/sd/j/joshuaho/root_gnn/root_gnn_dgl/data/processed_pretraining_multiclass_12_process/
-      tree_name: output
-      weight_var: weight
-    class: LazyDataset
-    folding: *id004
-    module: root_gnn_base.dataset
-    padding_mode: NONE
-    shuffle_chunks: 10
-  ZH:
-    args:
-      buffer_size: 11
-      chunks: 100
-      file_names: ZH_NLO_inc.root
-      fold_var: Number
-      label: 5
-      name: ZH
-      node_branch_names: *id001
-      node_branch_types: *id002
-      node_feature_scales: *id003
-      raw_dir: /global/cfs/projectdirs/atlas/joshua/root_gnn/root_gnn_dgl/data/ntuples/Hyy_pretraining/
-      save_dir: /pscratch/sd/j/joshuaho/root_gnn/root_gnn_dgl/data/processed_pretraining_multiclass_12_process/
-      tree_name: output
-      weight_var: weight
-    class: LazyDataset
-    folding: *id004
-    module: root_gnn_base.dataset
-    padding_mode: NONE
-    shuffle_chunks: 10
-  ggF:
-    args:
-      buffer_size: 11
-      chunks: 100
-      file_names: ggF_NLO_inc.root
-      fold_var: Number
-      label: 2
-      name: ggF
-      node_branch_names: *id001
-      node_branch_types: *id002
-      node_feature_scales: *id003
-      raw_dir: /global/cfs/projectdirs/atlas/joshua/root_gnn/root_gnn_dgl/data/ntuples/Hyy_pretraining/
-      save_dir: /pscratch/sd/j/joshuaho/root_gnn/root_gnn_dgl/data/processed_pretraining_multiclass_12_process/
-      tree_name: output
-      weight_var: weight
-    class: LazyDataset
-    folding: *id004
-    module: root_gnn_base.dataset
-    padding_mode: NONE
-    shuffle_chunks: 10
-  tHjb:
-    args:
-      buffer_size: 11
-      chunks: 100
-      file_names: tHjb_NLO_inc.root
-      fold_var: Number
-      label: 1
-      name: tHjb
-      node_branch_names: *id001
-      node_branch_types: *id002
-      node_feature_scales: *id003
-      raw_dir: /global/cfs/projectdirs/atlas/joshua/root_gnn/root_gnn_dgl/data/ntuples/Hyy_pretraining/
-      save_dir: /pscratch/sd/j/joshuaho/root_gnn/root_gnn_dgl/data/processed_pretraining_multiclass_12_process/
-      tree_name: output
-      weight_var: weight
-    class: LazyDataset
-    folding: *id004
-    module: root_gnn_base.dataset
-    padding_mode: NONE
-    shuffle_chunks: 10
-  ttH:
-    args:
-      buffer_size: 11
-      chunks: 100
-      file_names: ttH_NLO_inc.root
-      fold_var: Number
-      label: 0
-      name: ttH
-      node_branch_names: *id001
-      node_branch_types: *id002
-      node_feature_scales: *id003
-      raw_dir: /global/cfs/projectdirs/atlas/joshua/root_gnn/root_gnn_dgl/data/ntuples/Hyy_pretraining/
-      save_dir: /pscratch/sd/j/joshuaho/root_gnn/root_gnn_dgl/data/processed_pretraining_multiclass_12_process/
-      tree_name: output
-      weight_var: weight
-    class: LazyDataset
-    folding: *id004
-    module: root_gnn_base.dataset
-    padding_mode: NONE
-    shuffle_chunks: 10
-  ttW:
-    args:
-      buffer_size: 11
-      chunks: 100
-      file_names: ttW.root
-      fold_var: Number
-      label: 10
-      name: ttW
-      node_branch_names: *id001
-      node_branch_types: *id002
-      node_feature_scales: *id003
-      raw_dir: /global/cfs/projectdirs/atlas/joshua/root_gnn/root_gnn_dgl/data/ntuples/Hyy_pretraining/
-      save_dir: /pscratch/sd/j/joshuaho/root_gnn/root_gnn_dgl/data/processed_pretraining_multiclass_12_process/
-      tree_name: output
-      weight_var: weight
-    class: LazyDataset
-    folding: *id004
-    module: root_gnn_base.dataset
-    padding_mode: NONE
-    shuffle_chunks: 10
-  ttbar:
-    args:
-      buffer_size: 11
-      chunks: 100
-      file_names: ttbar.root
-      fold_var: Number
-      label: 9
-      name: ttbar
-      node_branch_names: *id001
-      node_branch_types: *id002
-      node_feature_scales: *id003
-      raw_dir: /global/cfs/projectdirs/atlas/joshua/root_gnn/root_gnn_dgl/data/ntuples/Hyy_pretraining/
-      save_dir: /pscratch/sd/j/joshuaho/root_gnn/root_gnn_dgl/data/processed_pretraining_multiclass_12_process/
-      tree_name: output
-      weight_var: weight
-    class: LazyDataset
-    folding: *id004
-    module: root_gnn_base.dataset
-    padding_mode: NONE
-    shuffle_chunks: 10
-  ttt:
-    args:
-      buffer_size: 11
-      chunks: 100
-      file_names: ttt.root
-      fold_var: Number
-      label: 11
-      name: ttt
-      node_branch_names: *id001
-      node_branch_types: *id002
-      node_feature_scales: *id003
-      raw_dir: /global/cfs/projectdirs/atlas/joshua/root_gnn/root_gnn_dgl/data/ntuples/Hyy_pretraining/
-      save_dir: /pscratch/sd/j/joshuaho/root_gnn/root_gnn_dgl/data/processed_pretraining_multiclass_12_process/
-      tree_name: output
-      weight_var: weight
-    class: LazyDataset
-    folding: *id004
-    module: root_gnn_base.dataset
-    padding_mode: NONE
-    shuffle_chunks: 10
-  tttt:
-    args:
-      buffer_size: 11
-      chunks: 100
-      file_names: tttt.root
-      fold_var: Number
-      label: 7
-      name: tttt
-      node_branch_names: *id001
-      node_branch_types: *id002
-      node_feature_scales: *id003
-      raw_dir: /global/cfs/projectdirs/atlas/joshua/root_gnn/root_gnn_dgl/data/ntuples/Hyy_pretraining/
-      save_dir: /pscratch/sd/j/joshuaho/root_gnn/root_gnn_dgl/data/processed_pretraining_multiclass_12_process/
-      tree_name: output
-      weight_var: weight
-    class: LazyDataset
-    folding: *id004
-    module: root_gnn_base.dataset
-    padding_mode: NONE
-    shuffle_chunks: 10
-  ttyy:
-    args:
-      buffer_size: 11
-      chunks: 100
-      file_names: ttyy.root
-      fold_var: Number
-      label: 6
-      name: ttyy_ch
-      node_branch_names: *id001
-      node_branch_types: *id002
-      node_feature_scales: *id003
-      raw_dir: /global/cfs/projectdirs/atlas/joshua/root_gnn/root_gnn_dgl/data/ntuples/Hyy_pretraining/
-      save_dir: /pscratch/sd/j/joshuaho/root_gnn/root_gnn_dgl/data/processed_pretraining_multiclass_12_process/
-      tree_name: output
-      weight_var: weight
-    class: LazyDataset
-    folding: *id004
-    module: root_gnn_base.dataset
-    padding_mode: NONE
-    shuffle_chunks: 10
-Loss:
-  args: {}
-  class: CrossEntropyLoss
-  finish:
-    args:
-      dim: 1
-    class: Softmax
-    module: torch.nn
-  module: torch.nn
-Model:
-  args:
-    dropout: 0
-    hid_size: 64
-    in_size: 7
-    n_layers: 4
-    n_proc_steps: 4
-    out_size: 12
-  class: Edge_Network
-  module: models.GCN
-Training:
-  batch_size: 1024
-  epochs: 100
-  gamma: 0.99
-  learning_rate: 0.0001
-Training_Directory: trainings/pretraining_multiclass/multiclass_12_process/
-Training_Name: multiclass_12_process

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_0.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:f13357759bae5ee14a4dd682fe867c09381aed0842ed712b0c12770aa621dd44
-size 1707210

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_1.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:9617167a01ee1d11ba2ec03f71440bd4fa25f4fa5f32e72bd9d71f5d10cd11b2
-size 1707210

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_10.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:5224b15282d60c8f09724904e5a665f27426d81ff577b234d901260050cb97e7
-size 1707502

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_11.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:a2fd4704348456e2a6c9655ee3aa2a9143df3f8394ba0d745207d46af139b558
-size 1707502

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_12.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:6960240dabb5128be0a501077ff952a985295885fbcece663282df40bb8d4584
-size 1707502

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_13.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:c11d6492d985d8f46b3476ed13671a3f9295269c0108dc55f491d7ef8270b901
-size 1707502

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_14.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:15bfd6edcb5cb2d9176683d13c546cdd2115eda320ddbc9a04b23dba8936d989
-size 1707502

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_15.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:5697fbccabb743b9c900a9e7f1e432077879fa4f83db42385e8ad1f597844195
-size 1707502

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_16.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:dc1818c04ee5d17f5cf05e985e69dbabf55e0c1c4c4b7c2cd29939b766350b39
-size 1707502

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_17.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:945575404de1372332c67423b1ea681edd4445e5445a21774d517dec45341c34
-size 1707502

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_18.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:7f8e9f611bba22ad5682c44d158de95f76c3dfc2223654a64540f91972c12ed7
-size 1707502

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_19.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:b1957ba99bf8b1aeb6b5a19a6867433aa86e5658c3abfdb737bc16489ce3d83b
-size 1707502

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_2.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:67b72a00c515215745d693f7f62d3a02274c8b702f53d7e68c1b162b0eee44ec
-size 1707210

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_20.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:38bd863cfeeb558c586b3c90686de8faf33f08b2675ca93edd4bf8ce63ba62ee
-size 1707502

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_21.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:7e90498dffe402581e1e554870b10369c05b45aba7b98a3480422512fbd9832e
-size 1707502

legacy/root_gnn_dgl/Pretrained_GNN/multiclass_pretrained_model_12/model_epoch_22.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:a0ea74fb7cfbfa3ebf6908f75fd6d5e018ee307801f608ecdde228d29e666f81
-size 1707502