"""
Pipeline for training and using a Flux1 model for simulation-based inference.
"""
import jax
import jax.numpy as jnp
from flax import nnx
from gensbi.models import (
Flux1,
Flux1Params,
)
import yaml
from gensbi.recipes.conditional_pipeline import (
ConditionalFlowPipeline,
ConditionalDiffusionPipeline,
)
[docs]
def parse_flux1_params(config_path: str):
"""
Parse a Flux1 configuration file.
Parameters
----------
config_path : str
Path to the configuration file.
Returns
-------
config : dict
Parsed configuration dictionary.
"""
with open(config_path, "r") as f:
config = yaml.safe_load(f)
model_params = config.get("model", {})
params_dict = dict(
in_channels=model_params.get("in_channels", 1),
vec_in_dim=model_params.get("vec_in_dim", None),
context_in_dim=model_params.get("context_in_dim", 1),
mlp_ratio=model_params.get("mlp_ratio", 4),
num_heads=model_params.get("num_heads", 6),
depth=model_params.get("depth", 8),
depth_single_blocks=model_params.get("depth_single_blocks", 16),
axes_dim=model_params.get("axes_dim", [6, 0]),
qkv_bias=model_params.get("qkv_bias", True),
theta=model_params.get("theta", -1),
id_embedding_strategy=tuple(
model_params.get("id_embedding_strategy", ("absolute", "absolute"))
),
param_dtype=getattr(jnp, model_params.get("param_dtype", "float32")),
)
return params_dict
[docs]
def parse_training_config(config_path: str):
"""
Parse a training configuration file.
Parameters
----------
config_path : str
Path to the configuration file.
Returns
-------
config : dict
Parsed configuration dictionary.
"""
with open(config_path, "r") as f:
config = yaml.safe_load(f)
# Training parameters
train_params = config.get("training", {})
multistep = train_params.get("multistep", 1)
experiment_id = train_params.get("experiment_id", 1)
early_stopping = train_params.get("early_stopping", True)
nsteps = train_params.get("nsteps", 30000) * multistep
val_every = train_params.get("val_every", 100) * multistep
# Optimizer parameters
opt_params = config.get("optimizer", {})
MAX_LR = opt_params.get("max_lr", 1e-3)
MIN_LR = opt_params.get("min_lr", 0.0)
MIN_SCALE = MIN_LR / MAX_LR if MAX_LR > 0 else 0.0
warmup_steps = opt_params.get("warmup_steps", 500)
ema_decay = opt_params.get("ema_decay", 0.999)
decay_transition = opt_params.get("decay_transition", 0.85)
training_config = {}
# overwrite the defaults with the config file values
training_config["nsteps"] = nsteps
training_config["ema_decay"] = ema_decay
training_config["decay_transition"] = decay_transition
training_config["max_lr"] = MAX_LR
training_config["min_lr"] = MIN_LR
training_config["min_scale"] = MIN_SCALE
training_config["val_every"] = val_every
training_config["early_stopping"] = early_stopping
training_config["experiment_id"] = experiment_id
training_config["multistep"] = multistep
training_config["warmup_steps"] = warmup_steps
return training_config
[docs]
class Flux1FlowPipeline(ConditionalFlowPipeline):
def __init__(
self,
train_dataset,
val_dataset,
dim_obs: int,
dim_cond: int,
ch_obs=1,
ch_cond=1,
params=None,
training_config=None,
):
"""
Flow pipeline for training and using a Flux1 model for simulation-based inference.
Parameters
----------
train_dataset : grain dataset or iterator over batches
Training dataset.
val_dataset : grain dataset or iterator over batches
Validation dataset.
dim_obs : int
Dimension of the parameter space.
dim_cond : int
Dimension of the observation space.
ch_obs : int, optional
Number of channels in the observation data. Default is 1.
ch_cond : int, optional
Number of channels in the conditional data. Default is 1.
params : Flux1Params, optional
Parameters for the Flux1 model. If None, default parameters are used.
training_config : dict, optional
Configuration for training. If None, default configuration is used.
Examples
--------
Minimal example on how to instantiate and use the Flux1FlowPipeline:
.. literalinclude:: /examples/flux1_flow_pipeline.py
:language: python
:linenos:
.. image:: /examples/flux1_flow_pipeline_marginals.png
:width: 600
.. note::
If you plan on using multiprocessing prefetching, ensure that your script is wrapped
in a ``if __name__ == "__main__":`` guard.
See https://docs.python.org/3/library/multiprocessing.html
"""
if params is not None:
ch_obs = params.in_channels
if params is not None:
ch_cond = params.context_in_dim
[docs]
self.dim_cond = dim_cond
if params is None:
params = self._get_default_params()
model = self._make_model(params)
super().__init__(
model=model,
train_dataset=train_dataset,
val_dataset=val_dataset,
dim_obs=dim_obs,
dim_cond=dim_cond,
ch_obs=ch_obs,
ch_cond=ch_cond,
params=params,
training_config=training_config,
id_embedding_strategy=params.id_embedding_strategy,
)
[docs]
self.ema_model = nnx.clone(self.model)
# TODO: check how to implement the in channels and cond channels properly, we may need to modify something here
@classmethod
[docs]
def init_pipeline_from_config(
cls,
train_dataset,
val_dataset,
dim_obs: int,
dim_cond: int,
config_path: str,
checkpoint_dir: str,
):
"""
Initialize the pipeline from a configuration file.
Parameters
----------
config_path : str
Path to the configuration file.
"""
with open(config_path, "r") as f:
config = yaml.safe_load(f)
# methodology
strategy = config.get("strategy", {})
method = strategy.get("method")
model_type = strategy.get("model")
assert method == "flow", f"Method {method} not supported in Flux1FlowPipeline."
assert (
model_type == "flux"
), f"Model type {model_type} not supported in Flux1FlowPipeline."
params_dict = parse_flux1_params(config_path)
params = Flux1Params(
rngs=nnx.Rngs(0),
dim_obs=dim_obs,
dim_cond=dim_cond,
**params_dict,
)
# Training parameters
training_config = cls.get_default_training_config()
training_config["checkpoint_dir"] = checkpoint_dir
training_config_ = parse_training_config(config_path)
for key, value in training_config_.items():
training_config[key] = value # update with config file values
pipeline = cls(
train_dataset,
val_dataset,
dim_obs,
dim_cond,
ch_obs=params.in_channels,
ch_cond=params.context_in_dim,
params=params,
training_config=training_config,
)
return pipeline
[docs]
def _make_model(self, params):
"""
Create and return the Flux1 model to be trained.
"""
model = Flux1(params)
return model
[docs]
def _get_default_params(self):
"""
Return default parameters for the Flux1 model.
"""
params = Flux1Params(
in_channels=self.ch_obs,
vec_in_dim=None,
context_in_dim=self.ch_cond,
mlp_ratio=4,
num_heads=6,
depth=8,
depth_single_blocks=16,
axes_dim=[6, 1],
qkv_bias=True,
dim_obs=self.dim_obs,
dim_cond=self.dim_cond,
theta=10 * (self.dim_obs + self.dim_cond),
id_embedding_strategy=("absolute", "absolute"),
rngs=nnx.Rngs(default=42),
param_dtype=jnp.float32,
)
return params
[docs]
class Flux1DiffusionPipeline(ConditionalDiffusionPipeline):
def __init__(
self,
train_dataset,
val_dataset,
dim_obs: int,
dim_cond: int,
ch_obs=1,
ch_cond=1,
params=None,
training_config=None,
):
"""
Diffusion pipeline for training and using a Flux1 model for simulation-based inference.
Parameters
----------
train_dataset : grain dataset or iterator over batches
Training dataset.
val_dataset : grain dataset or iterator over batches
Validation dataset.
dim_obs : int
Dimension of the parameter space.
dim_cond : int
Dimension of the observation space.
ch_obs : int, optional
Number of channels in the observation data. Default is 1.
ch_cond : int, optional
Number of channels in the conditional data. Default is 1.
params : Flux1Params, optional
Parameters for the Flux1 model. If None, default parameters are used.
training_config : dict, optional
Configuration for training. If None, default configuration is used.
Examples
--------
Minimal example on how to instantiate and use the Flux1DiffusionPipeline:
.. literalinclude:: /examples/flux1_diffusion_pipeline.py
:language: python
:linenos:
.. image:: /examples/flux1_diffusion_pipeline_marginals.png
:width: 600
.. note::
If you plan on using multiprocessing prefetching, ensure that your script is wrapped
in a ``if __name__ == "__main__":`` guard.
See https://docs.python.org/3/library/multiprocessing.html
"""
if params is not None:
ch_obs = params.in_channels
if params is not None:
ch_cond = params.context_in_dim
[docs]
self.dim_cond = dim_cond
if params is None:
params = self._get_default_params()
model = self._make_model(params)
super().__init__(
model=model,
train_dataset=train_dataset,
val_dataset=val_dataset,
dim_obs=dim_obs,
dim_cond=dim_cond,
ch_obs=ch_obs,
ch_cond=ch_cond,
params=params,
training_config=training_config,
id_embedding_strategy=params.id_embedding_strategy,
)
[docs]
self.ema_model = nnx.clone(self.model)
# TODO: need to update this too
@classmethod
[docs]
def init_pipeline_from_config(
cls,
train_dataset,
val_dataset,
dim_obs: int,
dim_cond: int,
config_path: str,
checkpoint_dir: str,
):
"""
Initialize the pipeline from a configuration file.
Parameters
----------
config_path : str
Path to the configuration file.
"""
with open(config_path, "r") as f:
config = yaml.safe_load(f)
# methodology
strategy = config.get("strategy", {})
method = strategy.get("method")
model_type = strategy.get("model")
assert (
method == "diffusion"
), f"Method {method} not supported in Flux1DiffusionPipeline."
assert (
model_type == "flux"
), f"Model type {model_type} not supported in Flux1DiffusionPipeline."
# Model parameters from config
dim_joint = dim_obs + dim_cond
params_dict = parse_flux1_params(config_path)
if params_dict["theta"] == -1:
params_dict["theta"] = 4 * dim_joint
params = Flux1Params(
rngs=nnx.Rngs(0),
dim_obs=dim_obs,
dim_cond=dim_cond,
**params_dict,
)
# Training parameters
training_config = cls.get_default_training_config()
training_config["checkpoint_dir"] = checkpoint_dir
training_config_ = parse_training_config(config_path)
for key, value in training_config_.items():
training_config[key] = value # update with config file values
pipeline = cls(
train_dataset,
val_dataset,
dim_obs,
dim_cond,
ch_obs=params.in_channels,
ch_cond=params.context_in_dim,
params=params,
training_config=training_config,
)
return pipeline
[docs]
def _make_model(self, params):
"""
Create and return the Flux1 model to be trained.
"""
model = Flux1(params)
return model
[docs]
def _get_default_params(self):
"""
Return default parameters for the Flux1 model.
"""
params = Flux1Params(
in_channels=self.ch_obs,
vec_in_dim=None,
context_in_dim=self.ch_cond,
mlp_ratio=4,
num_heads=6,
depth=8,
depth_single_blocks=16,
axes_dim=[6, 1],
qkv_bias=True,
dim_obs=self.dim_obs,
dim_cond=self.dim_cond,
theta=10 * (self.dim_obs + self.dim_cond),
id_embedding_strategy=("absolute", "absolute"),
rngs=nnx.Rngs(default=42),
param_dtype=jnp.float32,
)
return params
