ROADMAP.md

ModularML Roadmap

v1.0.0 — Core ModularML Pipeline

Target Release: Q1 2026

---

Data Structures & Serialization

• Refactor FeatureSet, FeatureSubset, Batch, and related structures to use PyArrow tables
• Implement zero-copy subset & sampler views over parent FeatureSet tables
• Ensure data loads into memory only when needed for ModelGraph execution
• Make all components fully serializable (FeatureSets, ModelGraphs, Stages, Samplers, Losses, Phases)
• Support exporting Experiments as:
  • Full state (post-training, weights included)

---

Experiment Context & Tracking

• Implement automatic Experiment context binding for all defined components
• Add conflict detection for mismatched component/Experiment associations
• Store all outputs (loss curves, metrics, results, figures) linked to their source phase

---

FeatureSet / Splitting / Sampling

FeatureSet

• Fully structured feature–target–tag schema
• Per-column scaling/normalization with tracked transform pipelines

Splitting

• Ratio-based random splits
• Rule-based conditional splits (user-defined criteria)

Sampling

• Sample-wise batching
• N-Sampler-based paired sampling
• N-Sampler-based triplet sampling

---

ModelGraph

• Support sequential, branching, and merging DAGs
• Validate graph connectivity before training
• Add graph visualization utility (Graphviz/Dot/Mermaid)

---

ModelNode

• Unified wrappers for PyTorch, TensorFlow, and scikit-learn
• Built-in PyTorch models (Sequential MLP, CNN encoder)
• Backend-agnostic forward, training-step, and eval-step APIs

---

MergeNodes

• Support merging of multi ModelGraph branches
• ConcatNode for concatenating features of multiple inputs
  • Add non-concat aggregation strategies for targets and tags
  • Support padding of data with misaligned shapes
• Make merging backend-aware to prevent PyTorch auto-grad breakage

---

Experiment / TrainingPhase / EvaluationPhase

• Experiment holds static FeatureSets, splits, and ModelGraph
• Support multiple independent Training and Evaluation phases
• Each phase configurable with samplers, losses, optimizers, and trackers
• Store and version phase results in the Experiment instance

---

Unit Testing

• Add nox-based automated unit, integration, example, and doc test routines
• Increase code coverage to $\geq$ 90%

---

v1.1.0 — Multi-Experiment Container & Comparison

Target Release: Q3 2026

• Multi-input/output Samplers

  • Samplers can take in multiple FeatureSets
    • Must support sample alignment (separate from BatchSchedulingPolicy)
  • Samplers can produce multiple output streams

• Add higher-level ExperimentCollection container

• Support grouping Experiments for shared evaluation pipelines

• Provide unified comparison utilities across Experiments (metrics, plots, tables)

• Enable rapid testing of alternative ModelGraphs, architectures, or FeatureSets within the same task