compute output stat for atomic model (deepmodeling#3642)

This PR:
- breaking change: the base atomic model is now a module. 
  - reason: the out stat is a data attribute of the base atomic model. 
- implement the `compute_or_load_output_stat` for the base atomic model.
the method computes both bias and std.
- the derived atomic models call the `compute_or_load_output_stat`
method for computing output stat.
- atomic model provides the `apply_out_stat`, the derived class may
override the method to define how the statistics is applied to an atomic
model's output. @anyangml may need.
- `out_stat` support statistics of output tensor of any shape. 

@iProzd please check if i took it correctly in
[ce7ec1f](deepmodeling@ce7ec1f)

To be done:
- atomic statistics of the bias and std. @anyangml 
- erialization and deserialization.

---------

Signed-off-by: Han Wang <[email protected]>
Co-authored-by: Han Wang <[email protected]>
Co-authored-by: Duo <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Author: 4 people

deepmd/dpmodel/atomic_model/base_atomic_model.py

@@ -27,13 +27,19 @@
 class BaseAtomicModel(BaseAtomicModel_):
     def __init__(
         self,
+        type_map: List[str],
         atom_exclude_types: List[int] = [],
         pair_exclude_types: List[Tuple[int, int]] = [],
     ):
         super().__init__()
+        self.type_map = type_map
         self.reinit_atom_exclude(atom_exclude_types)
         self.reinit_pair_exclude(pair_exclude_types)
 
+    def get_type_map(self) -> List[str]:
+        """Get the type map."""
+        return self.type_map
+
     def reinit_atom_exclude(
         self,
         exclude_types: List[int] = [],

deepmd/dpmodel/atomic_model/dp_atomic_model.py

@@ -53,7 +53,7 @@ def __init__(
         self.descriptor = descriptor
         self.fitting = fitting
         self.type_map = type_map
-        super().__init__(**kwargs)
+        super().__init__(type_map, **kwargs)
 
     def fitting_output_def(self) -> FittingOutputDef:
         """Get the output def of the fitting net."""
@@ -67,10 +67,6 @@ def get_sel(self) -> List[int]:
         """Get the neighbor selection."""
         return self.descriptor.get_sel()
 
-    def get_type_map(self) -> List[str]:
-        """Get the type map."""
-        return self.type_map
-
     def mixed_types(self) -> bool:
         """If true, the model
         1. assumes total number of atoms aligned across frames;

deepmd/dpmodel/atomic_model/linear_atomic_model.py

@@ -66,7 +66,7 @@ def __init__(
             self.mapping_list.append(self.remap_atype(tpmp, self.type_map))
         assert len(err_msg) == 0, "\n".join(err_msg)
         self.mixed_types_list = [model.mixed_types() for model in self.models]
-        super().__init__(**kwargs)
+        super().__init__(type_map, **kwargs)
 
     def mixed_types(self) -> bool:
         """If true, the model

deepmd/dpmodel/atomic_model/make_base_atomic_model.py

@@ -51,9 +51,6 @@ def atomic_output_def(self) -> FittingOutputDef:
             """
             return self.fitting_output_def()
 
-        def get_output_keys(self) -> List[str]:
-            return list(self.atomic_output_def().keys())
-
         @abstractmethod
         def get_rcut(self) -> float:
             """Get the cut-off radius."""

deepmd/dpmodel/atomic_model/pairtab_atomic_model.py

@@ -59,16 +59,20 @@ def __init__(
         rcut: float,
         sel: Union[int, List[int]],
         type_map: List[str],
+        rcond: Optional[float] = None,
+        atom_ener: Optional[List[float]] = None,
         **kwargs,
     ):
-        super().__init__()
+        super().__init__(type_map, **kwargs)
         self.tab_file = tab_file
         self.rcut = rcut
         self.type_map = type_map
 
         self.tab = PairTab(self.tab_file, rcut=rcut)
         self.type_map = type_map
         self.ntypes = len(type_map)
+        self.rcond = rcond
+        self.atom_ener = atom_ener
 
         if self.tab_file is not None:
             self.tab_info, self.tab_data = self.tab.get()

deepmd/dpmodel/output_def.py

@@ -224,6 +224,10 @@ def __init__(
             if not self.r_differentiable:
                 raise ValueError("only r_differentiable variable can calculate hessian")
 
+    @property
+    def size(self):
+        return self.output_size
+
 
 class FittingOutputDef:
     """Defines the shapes and other properties of the fitting network outputs.

deepmd/pt/model/atomic_model/base_atomic_model.py

@@ -23,6 +23,7 @@
 from deepmd.pt.utils import (
     AtomExcludeMask,
     PairExcludeMask,
+    env,
 )
 from deepmd.pt.utils.nlist import (
     extend_input_and_build_neighbor_list,
@@ -35,19 +36,88 @@
 )
 
 log = logging.getLogger(__name__)
+dtype = env.GLOBAL_PT_FLOAT_PRECISION
+device = env.DEVICE
 
 BaseAtomicModel_ = make_base_atomic_model(torch.Tensor)
 
 
-class BaseAtomicModel(BaseAtomicModel_):
+class BaseAtomicModel(torch.nn.Module, BaseAtomicModel_):
+    """The base of atomic model.
+
+    Parameters
+    ----------
+    type_map
+        Mapping atom type to the name (str) of the type.
+        For example `type_map[1]` gives the name of the type 1.
+    atom_exclude_types
+        Exclude the atomic contribution of the given types
+    pair_exclude_types
+        Exclude the pair of atoms of the given types from computing the output
+        of the atomic model. Implemented by removing the pairs from the nlist.
+    rcond : float, optional
+        The condition number for the regression of atomic energy.
+    preset_out_bias : Dict[str, List[Optional[torch.Tensor]]], optional
+        Specifying atomic energy contribution in vacuum. Given by key:value pairs.
+        The value is a list specifying the bias. the elements can be None or np.array of output shape.
+        For example: [None, [2.]] means type 0 is not set, type 1 is set to [2.]
+        The `set_davg_zero` key in the descrptor should be set.
+
+    """
+
     def __init__(
         self,
+        type_map: List[str],
         atom_exclude_types: List[int] = [],
         pair_exclude_types: List[Tuple[int, int]] = [],
+        rcond: Optional[float] = None,
+        preset_out_bias: Optional[Dict[str, torch.Tensor]] = None,
     ):
-        super().__init__()
+        torch.nn.Module.__init__(self)
+        BaseAtomicModel_.__init__(self)
+        self.type_map = type_map
         self.reinit_atom_exclude(atom_exclude_types)
         self.reinit_pair_exclude(pair_exclude_types)
+        self.rcond = rcond
+        self.preset_out_bias = preset_out_bias
+
+    def init_out_stat(self):
+        """Initialize the output bias."""
+        ntypes = self.get_ntypes()
+        self.bias_keys: List[str] = list(self.fitting_output_def().keys())
+        self.max_out_size = max(
+            [self.atomic_output_def()[kk].size for kk in self.bias_keys]
+        )
+        self.n_out = len(self.bias_keys)
+        out_bias_data = torch.zeros(
+            [self.n_out, ntypes, self.max_out_size], dtype=dtype, device=device
+        )
+        out_std_data = torch.ones(
+            [self.n_out, ntypes, self.max_out_size], dtype=dtype, device=device
+        )
+        self.register_buffer("out_bias", out_bias_data)
+        self.register_buffer("out_std", out_std_data)
+
+    def __setitem__(self, key, value):
+        if key in ["out_bias"]:
+            self.out_bias = value
+        elif key in ["out_std"]:
+            self.out_std = value
+        else:
+            raise KeyError(key)
+
+    def __getitem__(self, key):
+        if key in ["out_bias"]:
+            return self.out_bias
+        elif key in ["out_std"]:
+            return self.out_std
+        else:
+            raise KeyError(key)
+
+    @torch.jit.export
+    def get_type_map(self) -> List[str]:
+        """Get the type map."""
+        return self.type_map
 
     def reinit_atom_exclude(
         self,
@@ -165,6 +235,7 @@ def forward_common_atomic(
             fparam=fparam,
             aparam=aparam,
         )
+        ret_dict = self.apply_out_stat(ret_dict, atype)
 
         # nf x nloc
         atom_mask = ext_atom_mask[:, :nloc].to(torch.int32)
@@ -210,9 +281,60 @@ def compute_or_load_stat(
         """
         raise NotImplementedError
 
+    def compute_or_load_out_stat(
+        self,
+        merged: Union[Callable[[], List[dict]], List[dict]],
+        stat_file_path: Optional[DPPath] = None,
+    ):
+        """
+        Compute the output statistics (e.g. energy bias) for the fitting net from packed data.
+
+        Parameters
+        ----------
+        merged : Union[Callable[[], List[dict]], List[dict]]
+            - List[dict]: A list of data samples from various data systems.
+                Each element, `merged[i]`, is a data dictionary containing `keys`: `torch.Tensor`
+                originating from the `i`-th data system.
+            - Callable[[], List[dict]]: A lazy function that returns data samples in the above format
+                only when needed. Since the sampling process can be slow and memory-intensive,
+                the lazy function helps by only sampling once.
+        stat_file_path : Optional[DPPath]
+            The path to the stat file.
+
+        """
+        self.change_out_bias(
+            merged,
+            stat_file_path=stat_file_path,
+            bias_adjust_mode="set-by-statistic",
+        )
+
+    def apply_out_stat(
+        self,
+        ret: Dict[str, torch.Tensor],
+        atype: torch.Tensor,
+    ):
+        """Apply the stat to each atomic output.
+        The developer may override the method to define how the bias is applied
+        to the atomic output of the model.
+
+        Parameters
+        ----------
+        ret
+            The returned dict by the forward_atomic method
+        atype
+            The atom types. nf x nloc
+
+        """
+        out_bias, out_std = self._fetch_out_stat(self.bias_keys)
+        for kk in self.bias_keys:
+            # nf x nloc x odims, out_bias: ntypes x odims
+            ret[kk] = ret[kk] + out_bias[kk][atype]
+        return ret
+
     def change_out_bias(
         self,
         sample_merged,
+        stat_file_path: Optional[DPPath] = None,
         bias_adjust_mode="change-by-statistic",
     ) -> None:
         """Change the output bias according to the input data and the pretrained model.
@@ -231,22 +353,32 @@ def change_out_bias(
             'change-by-statistic' : perform predictions on labels of target dataset,
                     and do least square on the errors to obtain the target shift as bias.
             'set-by-statistic' : directly use the statistic output bias in the target dataset.
+        stat_file_path : Optional[DPPath]
+            The path to the stat file.
         """
         if bias_adjust_mode == "change-by-statistic":
-            delta_bias = compute_output_stats(
+            delta_bias, out_std = compute_output_stats(
                 sample_merged,
                 self.get_ntypes(),
-                keys=self.get_output_keys(),
+                keys=list(self.atomic_output_def().keys()),
+                stat_file_path=stat_file_path,
                 model_forward=self._get_forward_wrapper_func(),
-            )["energy"]
-            self.set_out_bias(delta_bias, add=True)
+                rcond=self.rcond,
+                preset_bias=self.preset_out_bias,
+            )
+            # self.set_out_bias(delta_bias, add=True)
+            self._store_out_stat(delta_bias, out_std, add=True)
         elif bias_adjust_mode == "set-by-statistic":
-            bias_atom = compute_output_stats(
+            bias_out, std_out = compute_output_stats(
                 sample_merged,
                 self.get_ntypes(),
-                keys=self.get_output_keys(),
-            )["energy"]
-            self.set_out_bias(bias_atom)
+                keys=list(self.atomic_output_def().keys()),
+                stat_file_path=stat_file_path,
+                rcond=self.rcond,
+                preset_bias=self.preset_out_bias,
+            )
+            # self.set_out_bias(bias_out)
+            self._store_out_stat(bias_out, std_out)
         else:
             raise RuntimeError("Unknown bias_adjust_mode mode: " + bias_adjust_mode)
 
@@ -279,3 +411,63 @@ def model_forward(coord, atype, box, fparam=None, aparam=None):
                 return {kk: vv.detach() for kk, vv in atomic_ret.items()}
 
         return model_forward
+
+    def _varsize(
+        self,
+        shape: List[int],
+    ) -> int:
+        output_size = 1
+        len_shape = len(shape)
+        for i in range(len_shape):
+            output_size *= shape[i]
+        return output_size
+
+    def _get_bias_index(
+        self,
+        kk: str,
+    ) -> int:
+        res: List[int] = []
+        for i, e in enumerate(self.bias_keys):
+            if e == kk:
+                res.append(i)
+        assert len(res) == 1
+        return res[0]
+
+    def _store_out_stat(
+        self,
+        out_bias: Dict[str, torch.Tensor],
+        out_std: Dict[str, torch.Tensor],
+        add: bool = False,
+    ):
+        ntypes = self.get_ntypes()
+        out_bias_data = torch.clone(self.out_bias)
+        out_std_data = torch.clone(self.out_std)
+        for kk in out_bias.keys():
+            assert kk in out_std.keys()
+            idx = self._get_bias_index(kk)
+            size = self._varsize(self.atomic_output_def()[kk].shape)
+            if not add:
+                out_bias_data[idx, :, :size] = out_bias[kk].view(ntypes, size)
+            else:
+                out_bias_data[idx, :, :size] += out_bias[kk].view(ntypes, size)
+            out_std_data[idx, :, :size] = out_std[kk].view(ntypes, size)
+        self.out_bias.copy_(out_bias_data)
+        self.out_std.copy_(out_std_data)
+
+    def _fetch_out_stat(
+        self,
+        keys: List[str],
+    ) -> Tuple[Dict[str, torch.Tensor], Dict[str, torch.Tensor]]:
+        ret_bias = {}
+        ret_std = {}
+        ntypes = self.get_ntypes()
+        for kk in keys:
+            idx = self._get_bias_index(kk)
+            isize = self._varsize(self.atomic_output_def()[kk].shape)
+            ret_bias[kk] = self.out_bias[idx, :, :isize].view(
+                [ntypes] + list(self.atomic_output_def()[kk].shape)  # noqa: RUF005
+            )
+            ret_std[kk] = self.out_std[idx, :, :isize].view(
+                [ntypes] + list(self.atomic_output_def()[kk].shape)  # noqa: RUF005
+            )
+        return ret_bias, ret_std