mifrost

mifrost is a high-performance graph encoding library for planning states and transitions. It combines C++ encoder engines (via nanobind) with Python-facing APIs. The API is native-first: encoders return BatchEncoding by default, and PyTorch Geometric (Data / HeteroData) objects are built on demand.

Documentation

• Website: https://maichmueller.github.io/mifrost/
• Tutorials: https://maichmueller.github.io/mifrost/tutorials/first-encoding/
• Encoder selection guide: https://maichmueller.github.io/mifrost/how-to/choose-an-encoder/
• API reference: https://maichmueller.github.io/mifrost/reference/api/

What it does

• Encodes planning states into graph structures for GNN pipelines.
• Supports single, batch, and stream-oriented encoding workflows.
• Exposes multiple encoder families:
  • HGraphEncoder
  • HorizonEncoder
  • TransitionHGraphEncoder / TransitionEffectsHGraphEncoder
  • FlatRelationEncoder, FlatHorizonEncoder, and both flat transition lanes
  • ColorEncoder
  • ILGEncoder
• Returns native BatchEncoding objects, with explicit helpers for:
  • PyG conversion (encode_pyg, encode_batch_pyg, as_pyg)

Requirements

• Python >= 3.12
• A working C++ toolchain
• At least one optional planner for encoder use: pymimir>=0.13.60 or pytyr==0.0.34
• For Python-side graph assembly: torch and torch-geometric
• For source builds: Conan (or CONAN_COMMAND/CONAN_CMD pointing to it)

Python dependency files

• requirements.txt: runtime Python dependency set
• requirements/base-build.txt: planner-neutral PEP 517, CMake, and Conan tooling
• requirements/build.txt: Python tooling for source builds (conan, cmake, ninja, etc.)
• requirements/test.txt: test-only Python dependencies
• requirements/perf.txt: performance-gate dependencies
• requirements/dev.txt: convenience union of build + test + perf + quality tools
• requirements/constraints-ci.txt: CI-only version constraints used by workflows
• pyproject.toml backend extras: .[pymimir], .[pytyr], .[backends]
• Development extras: .[test], .[perf], .[dev]

Installation

From PyPI

pip install "mifrost[pymimir]"

For PyTyr instead, or to use both planners in the same process:

pip install "mifrost[pytyr]"
pip install "mifrost[backends]"

From source (wheel)

git clone https://github.com/maichmueller/mifrost.git
cd mifrost
pip install ".[pymimir]"

Source builds contain both adapter modules by default. Set MIFROST_BUILD_BACKENDS=core, pymimir, pytyr, or both to build an explicit subset; for example:

MIFROST_BUILD_BACKENDS=pytyr pip install ".[pytyr]"

If Conan is not on your PATH, set:

export CONAN_COMMAND=/path/to/conan

If pymimir is installed but CMake cannot locate it, set:

export MIFROST_MIMIR_CMAKE_DIR="$(python -c 'import pymimir; print(pymimir.get_cmake_dir())')"

Editable install (development)

python -m pip install --no-build-isolation \
  --config-settings=editable.rebuild=true \
  -Cbuild-dir=build_editable \
  -e .

This enables import-triggered rebuild behavior from scikit-build-core for local development.

Reusable C++ SDK from the installed package

Installed wheels also ship the reusable native SDK alongside the Python module:

python -c 'import mifrost; print(mifrost.get_include_dir())'
python -c 'import mifrost; print(mifrost.get_cmake_dir())'
python -c 'import mifrost; print(mifrost.get_library_dir())'

Use mifrost.get_cmake_dir() in downstream CMake projects to extend CMAKE_PREFIX_PATH or pass -Dmifrost_DIR="$(python -c 'import mifrost; print(mifrost.get_cmake_dir())')".

Quick start (native-first)

import mifrost

# Pymimir domains and PyTyr planning tasks select their backend per instance.
encoder = mifrost.HGraphEncoder(domain_or_planning_task)

# state: a matching Pymimir or PyTyr state
encoding = encoder.encode(state)         # BatchEncoding
data = encoding.as_pyg()                 # HeteroData

batch_encoding = encoder.encode_batch([state1, state2, state3])  # BatchEncoding
batch = batch_encoding.as_pyg(as_batch=True)                     # HeteroDataBatch

# explicit PyG convenience helpers
data2 = encoder.encode_pyg(state)
batch2 = encoder.encode_batch_pyg([state1, state2, state3])
encoding_dict = batch_encoding.as_dict()      # dictionary form
# note: encoding_dict["tensors"] entries are DLPack-exporting values
# (consume with torch.utils.dlpack.from_dlpack(...) or mifrost.encoding_to_tensors(...))

Every public encoder family supports backend="pymimir" or backend="pytyr" for explicit selection; omitting it infers the backend from the constructor input. Instances from both planners can safely coexist. Their compatible outputs are ordinary planner-free BatchEncoding values and can be batched together:

pymimir_encoder = mifrost.HGraphEncoder(pymimir_domain)
pytyr_encoder = mifrost.HGraphEncoder(pytyr_planning_task)

mixed = mifrost.batch_encodings(
    [
        pymimir_encoder.encode(pymimir_state),
        pytyr_encoder.encode(pytyr_state),
    ],
    fast_path=True,
)
assert mixed.num_graphs == 2

See examples/encoders/backend_interchangeability_example.py for a complete same-process example that parses both planners, creates one mixed PyG batch, and runs a Torch forward/backward step.

Example output (trimmed):

type(encoding): BatchEncoding
type(data): HeteroData
node_type_count: 28
node_types_head:
  ['[+]on[g]', '_symbol_', 'clear', 'object', 'ontable',
   '[+]clear[g]', '[+]clear[g][sat]', '[+]handempty[g]']
edge_type_count: 54

type(batch_encoding): BatchEncoding
type(batch): HeteroDataBatch
num_graphs: 3
node_type_count: 28

encoding keys:
  ['node_feature_dims', 'node_names', 'num_graphs', 'object_names', 'schema', 'tensors']
schema keys:
  ['edge_tensors', 'edge_types', 'extensions', 'flags', 'graph_kind',
   'node_tensors', 'node_types', 'version']
tensor_count: 164
tensor_keys_head:
  ['_symbol_|0|object/edge_index_0', '_symbol_|0|object/edge_index_1',
   'object|0|_symbol_/edge_index_0', 'object|0|_symbol_/edge_index_1', ...]

Stream workflow:

stream = encoder.stream()
stream.append(state1)
stream.append(state2)

batch_encoding = stream.flush()  # BatchEncoding
batch = batch_encoding.as_pyg(as_batch=True)    # HeteroDataBatch

# convenience
batch2 = stream.flush_pyg(as_batch=True)

# mutable stream (supports update/remove)
mutable = encoder.mutable_stream()
sid = mutable.append(state1)
mutable.update(sid, state2)
mutable.remove(sid)

Example output:

type(batch): HeteroDataBatch
num_graphs: 2
node_type_count: 28
node_types_head:
  ['[+]clear[g]', '[+]clear[g][sat]', '[+]handempty[g]', '[+]handempty[g][sat]',
   '[+]holding[g]', '[+]holding[g][sat]', '[+]on[g]', '[+]on[g][sat]']

Encoding quick lookup

C++ (HGraphEncoderEngine)

All methods append into an existing BatchBuilder (they do not clear it). Call builder.next_graph() to commit one graph.

• encode(state, builder) / encode_state(state, builder)

  • State-only graph (objects + current facts).

• encode_step<GoalTag>(state, goals_span, actions_span, builder)

  • Convenience overload for typed goal literals (wraps into GoalInputs internally).

• encode(state, goals: GoalInputs, actions_span, builder)

  • Full step graph (state + goals + optional actions; plus optional derived relations depending on config).

• encode(state, goals, actions_span, history_subgoals, history_max_steps, builder)

  • Full step graph plus history nodes/links (dt-tagged subgoals).

C++ streaming

• HGraphStreamEncoder (append-only)

  • Direct append into one persistent builder.
  • append(...) -> id, flush(), flush_pyg(), reset().

• HGraphMutableStreamEncoder (cached/mutable, via StreamEncoderBase)

  • Supports update/remove with id stability and cache merge on flush.
  • append(...) -> id, update(id, ...), remove(id), flush(), flush_pyg(), reset(), set_reuse_removed(bool).

Python (HGraphEncoder)

• encode(state, *, ...) -> BatchEncoding

  • One graph, native encoding object.

• encode_pyg(state, *, ...) -> HeteroData

  • One graph, explicit PyG conversion path.

• encode_batch(states, *, ...) -> BatchEncoding

  • Many graphs, native batch encoding object.

• encode_batch_pyg(states, *, ...) -> HeteroDataBatch

  • Many graphs, explicit PyG conversion path.

• encode_batch argument semantics

  • states is always the batch axis.

  • Each extra batch argument is either shared (applies to all states) or a per-state sequence of length len(states) with optional None entries.

  • Batch parsing/execution is C++-backed across encoders.

  • High-level encode_batch(...) accepts wrapper/native planning inputs and converts wrapper/adapter-backed values in Python before entering the C++ batch parser.

  • Low-level _core._parse_* helpers and C++ batch internals are strict advanced-only.

  • Unknown batch kwargs raise TypeError; family-specific unsupported lanes raise a descriptive ValueError when they contain values.

  • Per-state argument length mismatches raise ValueError.

  • Example:

    • Shared goals/actions:
      • encoder.encode_batch(states, goals=goals, actions=actions)
    • Per-state goals/actions:
      • encoder.encode_batch(states, goals=[goals0, goals1], actions=[[a0], None])

  • Migration notes (hard break):

    • Old encoder-specific inference/ignoring of batch kwargs was removed.
    • Low-level _core._parse_* batch parser helpers no longer accept adapter-backed custom Python types.
      • High-level encoder encode_batch(...) now performs Python-side adapter conversion.
    • Transition*Encoder requires aligned successors and rejects explicit action/history lanes.
    • HorizonEncoder accepts per-state dags/goals/subgoal_layers and rejects explicit action/history lanes.
    • ColorEncoder rejects action payloads.

• stream() -> HGraphEncoderStream

  • Create an append-only stream encoder backed by the same C++ engine.

• mutable_stream() -> HGraphMutableEncoderStream

  • Create a mutable stream encoder supporting update/remove.

Python streaming (HGraphEncoderStream, append-only)

• append(state, *, goals=None, actions=None, subgoal_layers=None, history_subgoals=None, history_max_steps=None) -> id
• flush() -> BatchEncoding
• flush_pyg(...) -> PyG

Python mutable streaming (HGraphMutableEncoderStream)

Cached stream wrapper (ids + edits), merges on flush.

• append(state, *, goals=None, actions=None, subgoal_layers=None, history_subgoals=None, history_max_steps=None) -> id
• update(id, state, *, ...)
• remove(id)
• flush() -> BatchEncoding
• flush_pyg(...) -> PyG

Extending input types (adapter API)

Encoders are strict for native pymimir (advanced) types, but you can register explicit adapters for custom wrappers:

import mifrost

mifrost.register_state_adapter(MyStateType, lambda s: s.to_advanced_state())
mifrost.register_domain_adapter(MyDomainType, lambda d: d.to_advanced_domain())
mifrost.register_literal_adapter(MyLiteralType, lambda l: l.to_advanced_literal())
mifrost.register_action_adapter(MyActionType, lambda a: a.to_advanced_action())

Adapters are matched by exact concrete type.

Batch hard-break note:

• High-level encode_batch(...) / encode_batch_pyg(...) use adapter registries during Python-side preprocessing before calling strict C++ batch parsing.
• Direct low-level _core._parse_* batch helper calls remain advanced-only and reject adapter-backed objects.

Development

Configure and build C++ targets

python configure.py --config Release
python cbuild.py

Or use CMake presets:

cmake --preset local-release
cmake --build build/local-release

Build benchmarks:

python configure.py --config Release --build_dir build/bench-release --with_benchmarks
python cbuild.py build/bench-release --bench

Batch collation microbenchmark (batch_encodings, including pyobj collation):

python scripts/benchmark_batch_encodings.py --repeats 400 --warmup 50

Tests

Python tests:

pytest -q

C++ tests (after configure/build):

./build/<...>/src/mifrost_tests

Profiling

python scripts/profile_encoding.py --domain blocks --problem small
python scripts/profile_encoding.py --profile cprofile --include-goals
python scripts/profile_encoding.py --benchmark-pyg --no-export-node-names

Comprehensive scaling benchmark suite (diverse state batches, goals/actions combinations, transition encoders, horizon DAG-size sweeps):

python scripts/benchmark_encoder_suite.py \
  --domain blocks \
  --problem smedium \
  --max-states 256 \
  --batch-sizes 8,32,128 \
  --horizon-dag-sizes 8,32,64 \
  --horizon-batch-sizes 4,16 \
  --include-lgan-values false,true \
  --repeats 5 \
  --warmup 1 \
  --output-json /tmp/encoder_bench_suite.json

Include PyG conversion paths in the same run:

python scripts/benchmark_encoder_suite.py --benchmark-pyg

Output assembly notes

• Native-first:
  • encode(...) / encode_batch(...) return BatchEncoding.
  • BatchEncoding.as_pyg(...) converts to PyG.
• Convenience:
  • encode_pyg(...) / encode_batch_pyg(...) return PyG directly.
• BatchEncoding supports as_dict(), schema_fingerprint(), save(...), load(...).
• mifrost.batch_encodings([...]) batches single-graph encodings natively with schema checks.
• Use mifrost.encoding_to_tensors(encoding.as_dict()) when feeding a custom downstream pipeline.

Encoder architecture note

• See docs/development/architecture.md for the current Python/native layout and encoder architecture notes.

License

GPL-3.0-only