apply compression for se_e2_a_tebd (deepmodeling#2841)

deepmd/descriptor/__init__.py

@@ -14,6 +14,9 @@
 from .se_a_ebd import (
     DescrptSeAEbd,
 )
+from .se_a_ebd_v2 import (
+    DescrptSeAEbdV2,
+)
 from .se_a_ef import (
     DescrptSeAEf,
     DescrptSeAEfLower,
@@ -39,6 +42,7 @@
     "DescrptHybrid",
     "DescrptLocFrame",
     "DescrptSeA",
+    "DescrptSeAEbdV2",
     "DescrptSeAEbd",
     "DescrptSeAEf",
     "DescrptSeAEfLower",

deepmd/descriptor/se_a.py

@@ -10,6 +10,7 @@
 from deepmd.common import (
     cast_precision,
     get_activation_func,
+    get_np_precision,
     get_precision,
 )
 from deepmd.env import (
@@ -30,10 +31,17 @@
 from deepmd.nvnmd.utils.config import (
     nvnmd_cfg,
 )
+from deepmd.utils.compress import (
+    get_extra_side_embedding_net_variable,
+    get_two_side_type_embedding,
+    get_type_embedding,
+    make_data,
+)
 from deepmd.utils.errors import (
     GraphWithoutTensorError,
 )
 from deepmd.utils.graph import (
+    get_pattern_nodes_from_graph_def,
     get_tensor_by_name_from_graph,
 )
 from deepmd.utils.network import (
@@ -165,6 +173,7 @@ def __init__(
         uniform_seed: bool = False,
         multi_task: bool = False,
         spin: Optional[Spin] = None,
+        stripped_type_embedding: bool = False,
         **kwargs,
     ) -> None:
         """Constructor."""
@@ -185,6 +194,7 @@ def __init__(
         self.compress_activation_fn = get_activation_func(activation_function)
         self.filter_activation_fn = get_activation_func(activation_function)
         self.filter_precision = get_precision(precision)
+        self.filter_np_precision = get_np_precision(precision)
         self.exclude_types = set()
         for tt in exclude_types:
             assert len(tt) == 2
@@ -193,6 +203,9 @@ def __init__(
         self.set_davg_zero = set_davg_zero
         self.type_one_side = type_one_side
         self.spin = spin
+        self.stripped_type_embedding = stripped_type_embedding
+        self.extra_embeeding_net_variables = None
+        self.layer_size = len(neuron)
 
         # extend sel_a for spin system
         if self.spin is not None:
@@ -463,6 +476,39 @@ def enable_compression(
                     "The size of the next layer of the neural network must be twice the size of the previous layer."
                     % ",".join([str(item) for item in self.filter_neuron])
                 )
+        if self.stripped_type_embedding:
+            ret_two_side = get_pattern_nodes_from_graph_def(
+                graph_def, f"filter_type_all{suffix}/.+_two_side_ebd"
+            )
+            ret_one_side = get_pattern_nodes_from_graph_def(
+                graph_def, f"filter_type_all{suffix}/.+_one_side_ebd"
+            )
+            if len(ret_two_side) == 0 and len(ret_one_side) == 0:
+                raise RuntimeError(
+                    "can not find variables of embedding net from graph_def, maybe it is not a compressible model."
+                )
+            elif len(ret_one_side) != 0 and len(ret_two_side) != 0:
+                raise RuntimeError(
+                    "both one side and two side embedding net varaibles are detected, it is a wrong model."
+                )
+            elif len(ret_two_side) != 0:
+                self.final_type_embedding = get_two_side_type_embedding(self, graph)
+                self.matrix = get_extra_side_embedding_net_variable(
+                    self, graph_def, "two_side", "matrix", suffix
+                )
+                self.bias = get_extra_side_embedding_net_variable(
+                    self, graph_def, "two_side", "bias", suffix
+                )
+                self.extra_embedding = make_data(self, self.final_type_embedding)
+            else:
+                self.final_type_embedding = get_type_embedding(self, graph)
+                self.matrix = get_extra_side_embedding_net_variable(
+                    self, graph_def, "one_side", "matrix", suffix
+                )
+                self.bias = get_extra_side_embedding_net_variable(
+                    self, graph_def, "one_side", "bias", suffix
+                )
+                self.extra_embedding = make_data(self, self.final_type_embedding)
 
         self.compress = True
         self.table = DPTabulate(
@@ -588,6 +634,7 @@ def build(
             coord = tf.reshape(coord_, [-1, natoms[1] * 3])
         box = tf.reshape(box_, [-1, 9])
         atype = tf.reshape(atype_, [-1, natoms[1]])
+        self.atype = atype
 
         op_descriptor = (
             build_op_descriptor() if nvnmd_cfg.enable else op_module.prod_env_mat_a
@@ -606,6 +653,10 @@ def build(
             sel_a=self.sel_a,
             sel_r=self.sel_r,
         )
+        nlist_t = tf.reshape(self.nlist + 1, [-1])
+        atype_t = tf.concat([[self.ntypes], tf.reshape(self.atype, [-1])], axis=0)
+        self.nei_type_vec = tf.nn.embedding_lookup(atype_t, nlist_t)
+
         # only used when tensorboard was set as true
         tf.summary.histogram("descrpt", self.descrpt)
         tf.summary.histogram("rij", self.rij)
@@ -692,6 +743,8 @@ def _pass_filter(
             type_embedding = input_dict.get("type_embedding", None)
         else:
             type_embedding = None
+        if self.stripped_type_embedding and type_embedding is None:
+            raise RuntimeError("type_embedding is required for se_a_tebd_v2 model.")
         start_index = 0
         inputs = tf.reshape(inputs, [-1, natoms[0], self.ndescrpt])
         output = []
@@ -901,13 +954,89 @@ def _filter_lower(
         # with (natom x nei_type_i) x 1
         xyz_scatter = tf.reshape(tf.slice(inputs_reshape, [0, 0], [-1, 1]), [-1, 1])
         if type_embedding is not None:
-            xyz_scatter = self._concat_type_embedding(
-                xyz_scatter, nframes, natoms, type_embedding
-            )
-            if self.compress:
-                raise RuntimeError(
-                    "compression of type embedded descriptor is not supported at the moment"
+            if self.stripped_type_embedding:
+                if self.type_one_side:
+                    extra_embedding_index = self.nei_type_vec
+                else:
+                    padding_ntypes = type_embedding.shape[0]
+                    atype_expand = tf.reshape(self.atype, [-1, 1])
+                    idx_i = tf.tile(atype_expand * padding_ntypes, [1, self.nnei])
+                    idx_j = tf.reshape(self.nei_type_vec, [-1, self.nnei])
+                    idx = idx_i + idx_j
+                    index_of_two_side = tf.reshape(idx, [-1])
+                    extra_embedding_index = index_of_two_side
+
+                if not self.compress:
+                    if self.type_one_side:
+                        one_side_type_embedding_suffix = "_one_side_ebd"
+                        net_output = embedding_net(
+                            type_embedding,
+                            self.filter_neuron,
+                            self.filter_precision,
+                            activation_fn=activation_fn,
+                            resnet_dt=self.filter_resnet_dt,
+                            name_suffix=one_side_type_embedding_suffix,
+                            stddev=stddev,
+                            bavg=bavg,
+                            seed=self.seed,
+                            trainable=trainable,
+                            uniform_seed=self.uniform_seed,
+                            initial_variables=self.extra_embeeding_net_variables,
+                            mixed_prec=self.mixed_prec,
+                        )
+                        net_output = tf.nn.embedding_lookup(
+                            net_output, self.nei_type_vec
+                        )
+                    else:
+                        type_embedding_nei = tf.tile(
+                            tf.reshape(type_embedding, [1, padding_ntypes, -1]),
+                            [padding_ntypes, 1, 1],
+                        )  # (ntypes) * ntypes * Y
+                        type_embedding_center = tf.tile(
+                            tf.reshape(type_embedding, [padding_ntypes, 1, -1]),
+                            [1, padding_ntypes, 1],
+                        )  # ntypes * (ntypes) * Y
+                        two_side_type_embedding = tf.concat(
+                            [type_embedding_nei, type_embedding_center], -1
+                        )  # ntypes * ntypes * (Y+Y)
+                        two_side_type_embedding = tf.reshape(
+                            two_side_type_embedding,
+                            [-1, two_side_type_embedding.shape[-1]],
+                        )
+
+                        atype_expand = tf.reshape(self.atype, [-1, 1])
+                        idx_i = tf.tile(atype_expand * padding_ntypes, [1, self.nnei])
+                        idx_j = tf.reshape(self.nei_type_vec, [-1, self.nnei])
+                        idx = idx_i + idx_j
+                        index_of_two_side = tf.reshape(idx, [-1])
+                        self.extra_embedding_index = index_of_two_side
+
+                        two_side_type_embedding_suffix = "_two_side_ebd"
+                        net_output = embedding_net(
+                            two_side_type_embedding,
+                            self.filter_neuron,
+                            self.filter_precision,
+                            activation_fn=activation_fn,
+                            resnet_dt=self.filter_resnet_dt,
+                            name_suffix=two_side_type_embedding_suffix,
+                            stddev=stddev,
+                            bavg=bavg,
+                            seed=self.seed,
+                            trainable=trainable,
+                            uniform_seed=self.uniform_seed,
+                            initial_variables=self.extra_embeeding_net_variables,
+                            mixed_prec=self.mixed_prec,
+                        )
+                        net_output = tf.nn.embedding_lookup(net_output, idx)
+                    net_output = tf.reshape(net_output, [-1, self.filter_neuron[-1]])
+            else:
+                xyz_scatter = self._concat_type_embedding(
+                    xyz_scatter, nframes, natoms, type_embedding
                 )
+                if self.compress:
+                    raise RuntimeError(
+                        "compression of type embedded descriptor is not supported when stripped_type_embedding == False"
+                    )
         # natom x 4 x outputs_size
         if nvnmd_cfg.enable:
             return filter_lower_R42GR(
@@ -929,25 +1058,48 @@ def _filter_lower(
                 self.embedding_net_variables,
             )
         if self.compress and (not is_exclude):
-            if self.type_one_side:
-                net = "filter_-1_net_" + str(type_i)
+            if self.stripped_type_embedding:
+                net_output = tf.nn.embedding_lookup(
+                    self.extra_embedding, extra_embedding_index
+                )
+                net = "filter_net"
+                info = [
+                    self.lower[net],
+                    self.upper[net],
+                    self.upper[net] * self.table_config[0],
+                    self.table_config[1],
+                    self.table_config[2],
+                    self.table_config[3],
+                ]
+                return op_module.tabulate_fusion_se_atten(
+                    tf.cast(self.table.data[net], self.filter_precision),
+                    info,
+                    xyz_scatter,
+                    tf.reshape(inputs_i, [natom, shape_i[1] // 4, 4]),
+                    net_output,
+                    last_layer_size=outputs_size[-1],
+                    is_sorted=False,
+                )
             else:
-                net = "filter_" + str(type_input) + "_net_" + str(type_i)
-            info = [
-                self.lower[net],
-                self.upper[net],
-                self.upper[net] * self.table_config[0],
-                self.table_config[1],
-                self.table_config[2],
-                self.table_config[3],
-            ]
-            return op_module.tabulate_fusion_se_a(
-                tf.cast(self.table.data[net], self.filter_precision),
-                info,
-                xyz_scatter,
-                tf.reshape(inputs_i, [natom, shape_i[1] // 4, 4]),
-                last_layer_size=outputs_size[-1],
-            )
+                if self.type_one_side:
+                    net = "filter_-1_net_" + str(type_i)
+                else:
+                    net = "filter_" + str(type_input) + "_net_" + str(type_i)
+                info = [
+                    self.lower[net],
+                    self.upper[net],
+                    self.upper[net] * self.table_config[0],
+                    self.table_config[1],
+                    self.table_config[2],
+                    self.table_config[3],
+                ]
+                return op_module.tabulate_fusion_se_a(
+                    tf.cast(self.table.data[net], self.filter_precision),
+                    info,
+                    xyz_scatter,
+                    tf.reshape(inputs_i, [natom, shape_i[1] // 4, 4]),
+                    last_layer_size=outputs_size[-1],
+                )
         else:
             if not is_exclude:
                 # with (natom x nei_type_i) x out_size
@@ -966,6 +1118,9 @@ def _filter_lower(
                     initial_variables=self.embedding_net_variables,
                     mixed_prec=self.mixed_prec,
                 )
+
+                if self.stripped_type_embedding:
+                    xyz_scatter = xyz_scatter * net_output + xyz_scatter
                 if (not self.uniform_seed) and (self.seed is not None):
                     self.seed += self.seed_shift
             else:
@@ -1179,3 +1334,10 @@ def init_variables(
                 self.dstd = new_dstd
                 if self.original_sel is None:
                     self.original_sel = sel
+
+    @property
+    def explicit_ntypes(self) -> bool:
+        """Explicit ntypes with type embedding."""
+        if self.stripped_type_embedding:
+            return True
+        return False

deepmd/descriptor/se_a_ebd_v2.py

@@ -0,0 +1,70 @@
+# SPDX-License-Identifier: LGPL-3.0-or-later
+import logging
+from typing import (
+    List,
+    Optional,
+)
+
+from deepmd.utils.spin import (
+    Spin,
+)
+
+from .descriptor import (
+    Descriptor,
+)
+from .se_a import (
+    DescrptSeA,
+)
+
+log = logging.getLogger(__name__)
+
+
+@Descriptor.register("se_a_tpe_v2")
+@Descriptor.register("se_a_ebd_v2")
+class DescrptSeAEbdV2(DescrptSeA):
+    r"""A compressible se_a_ebd model.
+
+    This model is a warpper for DescriptorSeA, which set stripped_type_embedding=True.
+    """
+
+    def __init__(
+        self,
+        rcut: float,
+        rcut_smth: float,
+        sel: List[str],
+        neuron: List[int] = [24, 48, 96],
+        axis_neuron: int = 8,
+        resnet_dt: bool = False,
+        trainable: bool = True,
+        seed: Optional[int] = None,
+        type_one_side: bool = True,
+        exclude_types: List[List[int]] = [],
+        set_davg_zero: bool = False,
+        activation_function: str = "tanh",
+        precision: str = "default",
+        uniform_seed: bool = False,
+        multi_task: bool = False,
+        spin: Optional[Spin] = None,
+        **kwargs,
+    ) -> None:
+        DescrptSeA.__init__(
+            self,
+            rcut,
+            rcut_smth,
+            sel,
+            neuron=neuron,
+            axis_neuron=axis_neuron,
+            resnet_dt=resnet_dt,
+            trainable=trainable,
+            seed=seed,
+            type_one_side=type_one_side,
+            exclude_types=exclude_types,
+            set_davg_zero=set_davg_zero,
+            activation_function=activation_function,
+            precision=precision,
+            uniform_seed=uniform_seed,
+            multi_task=multi_task,
+            spin=spin,
+            stripped_type_embedding=True,
+            **kwargs,
+        )