File size: 4,162 Bytes
cfeea40 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 |
# Copyright (c) Microsoft Corporation.
# Licensed under the MIT License.
import os
from pathlib import Path
from typing import Literal
import fire
from mattergen.common.data.types import TargetProperty
from mattergen.common.utils.eval_utils import MatterGenCheckpointInfo
from mattergen.generator import CrystalGenerator
def main(
output_path: str,
model_path: str,
batch_size: int = 64,
num_batches: int = 1,
config_overrides: list[str] | None = None,
checkpoint_epoch: Literal["best", "last"] | int = "last",
properties_to_condition_on: TargetProperty | None = None,
sampling_config_path: str | None = None,
sampling_config_name: str = "default",
sampling_config_overrides: list[str] | None = None,
record_trajectories: bool = True,
diffusion_guidance_factor: float | None = None,
strict_checkpoint_loading: bool = True,
target_compositions: list[dict[str, int]] | None = None,
):
"""
Evaluate diffusion model against molecular metrics.
Args:
model_path: Path to DiffusionLightningModule checkpoint directory.
output_path: Path to output directory.
config_overrides: Overrides for the model config, e.g., `model.num_layers=3 model.hidden_dim=128`.
properties_to_condition_on: Property value to draw conditional sampling with respect to. When this value is an empty dictionary (default), unconditional samples are drawn.
sampling_config_path: Path to the sampling config file. (default: None, in which case we use `DEFAULT_SAMPLING_CONFIG_PATH` from explorers.common.utils.utils.py)
sampling_config_name: Name of the sampling config (corresponds to `{sampling_config_path}/{sampling_config_name}.yaml` on disk). (default: default)
sampling_config_overrides: Overrides for the sampling config, e.g., `condition_loader_partial.batch_size=32`.
load_epoch: Epoch to load from the checkpoint. If None, the best epoch is loaded. (default: None)
record: Whether to record the trajectories of the generated structures. (default: True)
strict_checkpoint_loading: Whether to raise an exception when not all parameters from the checkpoint can be matched to the model.
target_compositions: List of dictionaries with target compositions to condition on. Each dictionary should have the form `{element: number_of_atoms}`. If None, the target compositions are not conditioned on.
Only supported for models trained for crystal structure prediction (CSP) (default: None)
NOTE: When specifying dictionary values via the CLI, make sure there is no whitespace between the key and value, e.g., `--properties_to_condition_on={key1:value1}`.
"""
if not os.path.exists(output_path):
os.makedirs(output_path)
sampling_config_overrides = sampling_config_overrides or []
config_overrides = config_overrides or []
properties_to_condition_on = properties_to_condition_on or {}
target_compositions = target_compositions or []
checkpoint_info = MatterGenCheckpointInfo(
model_path=Path(model_path).resolve(),
load_epoch=checkpoint_epoch,
config_overrides=config_overrides,
strict_checkpoint_loading=strict_checkpoint_loading,
)
_sampling_config_path = Path(sampling_config_path) if sampling_config_path is not None else None
generator = CrystalGenerator(
checkpoint_info=checkpoint_info,
properties_to_condition_on=properties_to_condition_on,
batch_size=batch_size,
num_batches=num_batches,
sampling_config_name=sampling_config_name,
sampling_config_path=_sampling_config_path,
sampling_config_overrides=sampling_config_overrides,
record_trajectories=record_trajectories,
diffusion_guidance_factor=(
diffusion_guidance_factor if diffusion_guidance_factor is not None else 0.0
),
target_compositions_dict=target_compositions,
)
generator.generate(output_dir=Path(output_path))
if __name__ == "__main__":
# use fire instead of argparse to allow for the specification of dictionary values via the CLI
fire.Fire(main)
|