remdm-craftax / src /planners /model.py
MathisW78's picture
Upload COMP0258 demo bundle (code + diffusion/PPO checkpoints + ablation assets)
6140064 verified
"""Diffusion model lifecycle: construction, parameter init, checkpoint I/O, and apply closures."""
from __future__ import annotations
import json
import logging
from pathlib import Path
from typing import Any, Callable, Union
import jax
import jax.numpy as jnp
import numpy as np
import optax
import orbax.checkpoint as ocp
from flax.training.train_state import TrainState
from src.models.denoiser import DenoisingTransformer
logger = logging.getLogger(__name__)
_METADATA_FILENAME = "resume_metadata.json"
def build_model(config: dict, num_actions: int) -> DenoisingTransformer:
"""Construct a :class:`DenoisingTransformer` from a config dict.
Args:
config: Upper-cased config dict with architecture hyperparameters.
num_actions: Size of the discrete action vocabulary.
Returns:
An uninitialised :class:`DenoisingTransformer` instance.
"""
return DenoisingTransformer(
num_actions=num_actions,
plan_horizon=config["PLAN_HORIZON"],
d_model=config.get("D_MODEL", 256),
n_heads=config.get("N_HEADS", 4),
n_layers=config.get("N_LAYERS", 4),
d_ff=config.get("D_FF", 512),
obs_encoder_layers=config.get("OBS_ENCODER_LAYERS", 2),
obs_encoder_width=config.get("OBS_ENCODER_WIDTH", 512),
dropout_rate=config.get("DROPOUT_RATE", 0.1),
)
def init_params(
model: DenoisingTransformer,
rng: jax.Array,
obs_dim: int,
plan_horizon: int,
) -> Any:
"""Initialize model parameters with dummy inputs.
Args:
model: Flax module to initialise.
rng: PRNG key.
obs_dim: Observation dimensionality.
plan_horizon: Number of action steps in a plan.
Returns:
Initialised parameter pytree.
"""
return model.init(
rng,
jnp.zeros((1, obs_dim)),
jnp.zeros((1, plan_horizon), dtype=jnp.int32),
jnp.zeros((1,)),
)
def resolve_checkpoint_path(
path: str,
download_dir: str | None = None,
) -> str:
"""Resolve a checkpoint path, downloading from W&B if it is an artifact reference.
Paths prefixed with ``wandb:`` are treated as W&B artifact references
(e.g. ``wandb:entity/project/name:version``) and downloaded locally
before returning the filesystem path.
Args:
path: Local filesystem path or ``wandb:``-prefixed artifact
reference.
download_dir: Root directory for downloaded artifacts. When ``None``,
falls back to the wandb default (``./artifacts/``).
Returns:
Local filesystem path to the checkpoint directory.
"""
if not path.startswith("wandb:"):
return str(Path(path).resolve())
import wandb
artifact_ref = path.removeprefix("wandb:")
api = wandb.Api()
artifact = api.artifact(artifact_ref)
local_path = (
artifact.download(root=download_dir) if download_dir else artifact.download()
)
print(f"Downloaded W&B artifact '{artifact_ref}' -> '{local_path}'")
return local_path
def load_checkpoint(
model: DenoisingTransformer,
rng: jax.Array,
obs_dim: int,
plan_horizon: int,
path: str,
) -> Any:
"""Load diffusion model parameters from an Orbax checkpoint.
Args:
model: Flax module (used to build the abstract state structure).
rng: PRNG key for dummy initialisation.
obs_dim: Observation dimensionality.
plan_horizon: Number of action steps in a plan.
path: Path to the Orbax checkpoint directory.
Returns:
Restored parameter pytree.
Raises:
FileNotFoundError: If the checkpoint directory contains no saved steps.
"""
path = str(Path(path).resolve())
params = init_params(model, rng, obs_dim, plan_horizon)
abstract_state = create_train_state(model=model, params=params, lr=1e-4, max_grad_norm=1.0)
with ocp.CheckpointManager(path) as mgr:
step = mgr.latest_step()
if step is None:
raise FileNotFoundError(f"No checkpoint at {path}")
restored_state = mgr.restore(
step,
args=ocp.args.StandardRestore(item=abstract_state),
)
print(f"Loaded diffusion checkpoint from '{path}' (step {step})")
return restored_state.params
def create_train_state(
model: DenoisingTransformer,
params: Any,
lr: Union[float, Callable[[int], float]],
max_grad_norm: float,
) -> TrainState:
"""Create a :class:`TrainState` with gradient clipping and Adam.
Args:
model: Flax module (used only to bind ``apply_fn``).
params: Initialised parameter pytree.
lr: Constant learning rate or an optax schedule
(any callable ``step -> lr``).
max_grad_norm: Global gradient clipping threshold.
Returns:
A Flax ``TrainState`` ready for ``apply_gradients``.
"""
tx = optax.chain(optax.clip_by_global_norm(max_grad_norm), optax.adam(lr, eps=1e-5))
return TrainState.create(apply_fn=model.apply, params=params, tx=tx)
def make_apply_fns(
model: DenoisingTransformer,
) -> tuple[Callable, Callable]:
"""Return ``(apply_eval, apply_train)`` closures matching ``ModelApplyFn``.
Args:
model: Flax module.
Returns:
Tuple of ``(apply_eval, apply_train)`` where ``apply_train`` enables
dropout via ``rngs={"dropout": rng}``.
"""
def apply_eval(params: Any, obs: jnp.ndarray, z_t: jnp.ndarray, t: jnp.ndarray, _rng=None):
return model.apply(params, obs, z_t, t)
def apply_train(params: Any, obs: jnp.ndarray, z_t: jnp.ndarray, t: jnp.ndarray, rng=None):
return model.apply(
params, obs, z_t, t,
deterministic=False,
rngs={"dropout": rng} if rng is not None else {},
)
return apply_eval, apply_train
# ---------------------------------------------------------------------------
# Checkpoint metadata sidecar
# ---------------------------------------------------------------------------
class _NumpyEncoder(json.JSONEncoder):
"""JSON encoder that handles numpy scalar types."""
def default(self, o: Any) -> Any:
"""Serialize numpy scalars to native Python types.
Args:
o: Object to serialize.
Returns:
JSON-serializable object.
"""
if isinstance(o, np.integer):
return int(o)
if isinstance(o, np.floating):
return float(o)
if isinstance(o, np.ndarray):
return o.tolist()
return super().default(o)
def save_checkpoint_metadata(
checkpoint_dir: str,
mode: str,
update_step: int,
total_gradient_steps: int,
wandb_run_id: str | None,
config: dict[str, Any],
) -> None:
"""Write a JSON metadata sidecar alongside an Orbax checkpoint.
Args:
checkpoint_dir: Root directory of the Orbax checkpoint manager.
mode: Training mode (``"offline"`` or ``"online"``).
update_step: Final update step index.
total_gradient_steps: Total gradient steps completed.
wandb_run_id: Current W&B run ID, or ``None``.
config: Full training config snapshot.
"""
metadata = {
"mode": mode,
"update_step": int(update_step),
"total_gradient_steps_completed": int(total_gradient_steps),
"wandb_run_id": wandb_run_id,
"config_snapshot": config,
}
path = Path(checkpoint_dir) / _METADATA_FILENAME
with open(path, "w") as f:
json.dump(metadata, f, indent=2, cls=_NumpyEncoder)
print(f"Saved checkpoint metadata to {path}")
def load_checkpoint_metadata(
checkpoint_dir: str,
) -> dict[str, Any] | None:
"""Read the JSON metadata sidecar from a checkpoint directory.
Args:
checkpoint_dir: Root directory of the Orbax checkpoint manager.
Returns:
Parsed metadata dict, or ``None`` if the sidecar does not exist
(backward-compatible with checkpoints created before this feature).
"""
path = Path(checkpoint_dir) / _METADATA_FILENAME
if not path.exists():
return None
with open(path) as f:
return json.load(f)
def load_checkpoint_for_resume(
model: DenoisingTransformer,
rng: jax.Array,
obs_dim: int,
plan_horizon: int,
path: str,
lr_schedule: Union[float, Callable[[int], float]],
max_grad_norm: float,
) -> TrainState:
"""Load a full ``TrainState`` (params + optimizer state) for resume.
Unlike :func:`load_checkpoint` which returns only params, this function
restores the complete ``TrainState`` including Adam moments so that
training can continue seamlessly.
The ``lr_schedule`` and ``max_grad_norm`` must match the optimizer chain
structure used when the checkpoint was saved (same chain composition,
possibly different schedule values).
Args:
model: Flax module (used to build the abstract state).
rng: PRNG key for dummy initialisation.
obs_dim: Observation dimensionality.
plan_horizon: Number of action steps in a plan.
path: Path to the Orbax checkpoint directory.
lr_schedule: Learning rate or schedule matching the current run's
optimizer (must produce the same ``opt_state`` structure).
max_grad_norm: Global gradient clipping threshold.
Returns:
Restored ``TrainState`` with params, opt_state, and step from the
checkpoint. The caller should call ``.replace(step=...)`` to set the
correct LR offset for the resumed run.
Raises:
FileNotFoundError: If the checkpoint directory contains no saved steps.
"""
path = str(Path(path).resolve())
params = init_params(model, rng, obs_dim, plan_horizon)
abstract_state = create_train_state(model, params, lr_schedule, max_grad_norm)
with ocp.CheckpointManager(path) as mgr:
step = mgr.latest_step()
if step is None:
raise FileNotFoundError(
f"No checkpoint found at {path}"
)
restored_state = mgr.restore(
step,
args=ocp.args.StandardRestore(item=abstract_state),
)
print(
f"Loaded full TrainState for resume from '{path}' "
f"(step {step}, opt_state step {restored_state.step})"
)
return restored_state