Sample

vmc/sample/sampler

 1class Sampler:
 2
 3    def __init__(
 4        self,
 5        nqs: DDP,
 6        ele_info: ElectronInfo,
 7        eloc_param: Optional[dict],
 8        n_sample: int = 100,
 9        start_iter: int = 100,
10        start_n_sample: Optional[int] = None,
11        # therm_step: int = 2000,
12        debug_exact: bool = False,
13        seed: int = 100,
14        dtype=torch.double,
15        method_sample="AR",
16        use_same_tree: bool = False,
17        max_n_sample: Optional[int] = None,
18        max_unique_sample: Optional[int] = None,
19        only_AD: bool = False,
20        only_sample: bool = False,
21        use_sample_space: bool = False,
22        min_batch: int = 10000,
23        min_tree_height: Optional[int] = None,
24        det_lut: Optional[DetLUT] = None,
25        use_dfs_sample: bool = False,
26        use_spin_raising: bool = False,
27        spin_raising_coeff: float = 1.0,
28        given_state: Optional[Tensor] = None,
29        use_spin_flip: bool = False,
30    ) -> None:

eloc-param

 1from utils.enums import ElocMethod
 2from vmc.sample import ElocParams
 3eloc_param: ElocParams = {
 4    "method": ElocMethod.REDUCE,
 5    "use_unique": False,
 6    "use_LUT": False,
 7    "eps": 1e-2,
 8    "eps_sample": 100,
 9    # "alpha": 1.5,
10    # "max_memory": 5,
11    "batch": 1024,
12    "fp_batch": 300000,
13}

  • method: ElocMethod.SIMPLE, ElocMethod.REDUCE and ElocMethod.SAMPLE_SPACE

  • use_unique: Remove duplicate \(n^{\prime}\) , which gives a nice speedup in small systems.

  • use_LUT: Use LookUp-table to reduce \(\psi(n^{\prime})\). This must be True if method = ElocMethod.SAMPLE_SPACE.

  • eps, eps_sample: \(\epsilon, N\) see: Method. This is necessary if Method = ElocMethod.REDUCE.

  • batch, fp_batch: the nbatch of eloc and the nbatch of the forward, Default: -1. This is required if Method = ElocMethod.REDUCE or ElocMethod.SIMPLE.

  • alpha, max_memory:: the max of the memory when Method = ElocMethod.SAMPLE_SPACE.

Notes:

  • use_unique = False in the large systems(e.g. H50, STO-6G, aoa-basis).

  • use_LUT = False in the large systems or the multi-node(e.g. world-size > 16).

sample-param

 1sampler_param = {
 2    "n_sample": int(2 * 1e5),
 3    "start_n_sample": int(2 * 1.0e5),
 4    "start_iter": 200,
 5    # "max_n_sample": int(1.0e8),
 6    # "max_unique_sample": int(6 * 1.0e4),
 7    "debug_exact": False,  # exact optimization
 8    "seed": 123,
 9    "method_sample": "AR",
10    # "given_state": given_state,
11    "only_AD": False,
12    "min_batch": 80000,
13    # "det_lut": det_lut,  # only use in CI-NQS exact optimization
14    "use_same_tree": True,  # different rank-sample
15    "min_tree_height": 12,  # different rank-sample
16    "use_dfs_sample": True,
17    "eloc_param": eloc_param,
18    "use_spin_flip": False,
19}

  • n_sample: the number of the sampling.

  • start_n_sample, start_iter: the number of the sampling in the first n iteration.

  • max_n_sample, max_unique_sample: the max of the n-sample and unique-sample, which used to restrict the sampling.

  • debug_exact: exact optimization, the unique-sample is equal to the FCI-space dim.

  • seed: the random-seed of the sampling.

  • method_sample: the method of the sampling. This currently only supports AR (Auto regressive) when the world-size great 1.

  • only_AD: No sampling, random samples are selected to check the backward memory usage ratio.

  • min_batch: the batch of the sampling.

  • use_same_tree, min_tree_height: different rank-sample. There must are selected carefully if the word-size great 1.

  • use_dfs_sample: the DFS (Depth first search) or BFS (Breadth first search) sampling.

  • eloc_param: see eloc-param

  • use_spin_flip: see: Spin-flip, from utils.public_function import SpinProjection; SpinProjection.init(N=nele, S=0)

Notes:

  • min_batch, use_same_tree, min_tree_height, use_dfs_sample: These are implemented in the Ansatz(e.g. MPS-RNN, Transformer)