deepmd.jax.descriptor.se_atten_v2

deepmd.jax.descriptor.se_atten_v2#

Classes#

DescrptSeAttenV2

Attention-based descriptor which is proposed in the pretrainable DPA-1[1] model.

Module Contents#

class deepmd.jax.descriptor.se_atten_v2.DescrptSeAttenV2(rcut: float, rcut_smth: float, sel: list[int] | int, ntypes: int, neuron: list[int] = [25, 50, 100], axis_neuron: int = 8, tebd_dim: int = 8, tebd_input_mode: str = 'concat', resnet_dt: bool = False, trainable: bool = True, type_one_side: bool = False, attn: int = 128, attn_layer: int = 2, attn_dotr: bool = True, attn_mask: bool = False, exclude_types: list[tuple[int, int]] = [], env_protection: float = 0.0, set_davg_zero: bool = False, activation_function: str = 'tanh', precision: str = DEFAULT_PRECISION, scaling_factor=1.0, normalize: bool = True, temperature: float | None = None, trainable_ln: bool = True, ln_eps: float | None = 1e-05, smooth_type_embedding: bool = True, concat_output_tebd: bool = True, spin: Any | None = None, stripped_type_embedding: bool | None = None, use_econf_tebd: bool = False, use_tebd_bias: bool = False, type_map: list[str] | None = None, seed: int | list[int] | None = None)[source]#

Bases: deepmd.jax.descriptor.dpa1.DescrptDPA1, deepmd.dpmodel.descriptor.se_atten_v2.DescrptSeAttenV2

Attention-based descriptor which is proposed in the pretrainable DPA-1[1] model.

This descriptor, \(\mathcal{D}^i \in \mathbb{R}^{M \times M_{<}}\), is given by

\[\mathcal{D}^i = \frac{1}{N_c^2}(\hat{\mathcal{G}}^i)^T \mathcal{R}^i (\mathcal{R}^i)^T \hat{\mathcal{G}}^i_<,\]

where \(\hat{\mathcal{G}}^i\) represents the embedding matrix:math:mathcal{G}^i after additional self-attention mechanism and \(\mathcal{R}^i\) is defined by the full case in the se_e2_a descriptor. Note that we obtain \(\mathcal{G}^i\) using the type embedding method by default in this descriptor.

To perform the self-attention mechanism, the queries \(\mathcal{Q}^{i,l} \in \mathbb{R}^{N_c\times d_k}\), keys \(\mathcal{K}^{i,l} \in \mathbb{R}^{N_c\times d_k}\), and values \(\mathcal{V}^{i,l} \in \mathbb{R}^{N_c\times d_v}\) are first obtained:

\[\left(\mathcal{Q}^{i,l}\right)_{j}=Q_{l}\left(\left(\mathcal{G}^{i,l-1}\right)_{j}\right),\]

\[\left(\mathcal{K}^{i,l}\right)_{j}=K_{l}\left(\left(\mathcal{G}^{i,l-1}\right)_{j}\right),\]

\[\left(\mathcal{V}^{i,l}\right)_{j}=V_{l}\left(\left(\mathcal{G}^{i,l-1}\right)_{j}\right),\]

where \(Q_{l}\), \(K_{l}\), \(V_{l}\) represent three trainable linear transformations that output the queries and keys of dimension \(d_k\) and values of dimension \(d_v\), and \(l\) is the index of the attention layer. The input embedding matrix to the attention layers, denoted by \(\mathcal{G}^{i,0}\), is chosen as the two-body embedding matrix.

Then the scaled dot-product attention method is adopted:

\[A(\mathcal{Q}^{i,l}, \mathcal{K}^{i,l}, \mathcal{V}^{i,l}, \mathcal{R}^{i,l})=\varphi\left(\mathcal{Q}^{i,l}, \mathcal{K}^{i,l},\mathcal{R}^{i,l}\right)\mathcal{V}^{i,l},\]

where \(\varphi\left(\mathcal{Q}^{i,l}, \mathcal{K}^{i,l},\mathcal{R}^{i,l}\right) \in \mathbb{R}^{N_c\times N_c}\) is attention weights. In the original attention method, one typically has \(\varphi\left(\mathcal{Q}^{i,l}, \mathcal{K}^{i,l}\right)=\mathrm{softmax}\left(\frac{\mathcal{Q}^{i,l} (\mathcal{K}^{i,l})^{T}}{\sqrt{d_{k}}}\right)\), with \(\sqrt{d_{k}}\) being the normalization temperature. This is slightly modified to incorporate the angular information:

\[\varphi\left(\mathcal{Q}^{i,l}, \mathcal{K}^{i,l},\mathcal{R}^{i,l}\right) = \mathrm{softmax}\left(\frac{\mathcal{Q}^{i,l} (\mathcal{K}^{i,l})^{T}}{\sqrt{d_{k}}}\right) \odot \hat{\mathcal{R}}^{i}(\hat{\mathcal{R}}^{i})^{T},\]

where \(\hat{\mathcal{R}}^{i} \in \mathbb{R}^{N_c\times 3}\) denotes normalized relative coordinates,: \(\hat{\mathcal{R}}^{i}_{j} = \frac{\boldsymbol{r}_{ij}}{\lVert \boldsymbol{r}_{ij} \lVert}\) and \(\odot\) means element-wise multiplication.
Then layer normalization is added in a residual way to finally obtain the self-attention local embedding matrix: \(\hat{\mathcal{G}}^{i} = \mathcal{G}^{i,L_a}\) after \(L_a\) attention layers:[^1]

\[\mathcal{G}^{i,l} = \mathcal{G}^{i,l-1} + \mathrm{LayerNorm}(A(\mathcal{Q}^{i,l}, \mathcal{K}^{i,l}, \mathcal{V}^{i,l}, \mathcal{R}^{i,l})).\]

Parameters:

rcut: float: The cut-off radius \(r_c\)
rcut_smth: float: From where the environment matrix should be smoothed \(r_s\)
sellist[int], int: list[int]: sel[i] specifies the maxmum number of type i atoms in the cut-off radius int: the total maxmum number of atoms in the cut-off radius
ntypesint: Number of element types
neuronlist[int]: Number of neurons in each hidden layers of the embedding net \(\mathcal{N}\)
axis_neuron: int: Number of the axis neuron \(M_2\) (number of columns of the sub-matrix of the embedding matrix)
tebd_dim: int: Dimension of the type embedding
tebd_input_mode: str: The input mode of the type embedding. Supported modes are [“concat”, “strip”]. - “concat”: Concatenate the type embedding with the smoothed radial information as the union input for the embedding network. - “strip”: Use a separated embedding network for the type embedding and combine the output with the radial embedding network output.
resnet_dt: bool: Time-step dt in the resnet construction: y = x + dt * phi (Wx + b)
trainable: bool: If the weights of this descriptors are trainable.
trainable_ln: bool: Whether to use trainable shift and scale weights in layer normalization.
ln_eps: float, Optional: The epsilon value for layer normalization.
type_one_side: bool: If ‘False’, type embeddings of both neighbor and central atoms are considered. If ‘True’, only type embeddings of neighbor atoms are considered. Default is ‘False’.
attn: int: Hidden dimension of the attention vectors
attn_layer: int: Number of attention layers
attn_dotr: bool: If dot the angular gate to the attention weights
attn_mask: bool: (Only support False to keep consistent with other backend references.) (Not used in this version. True option is not implemented.) If mask the diagonal of attention weights
exclude_typeslist[list[int]]: The excluded pairs of types which have no interaction with each other. For example, [[0, 1]] means no interaction between type 0 and type 1.
env_protection: float: Protection parameter to prevent division by zero errors during environment matrix calculations.
set_davg_zero: bool: Set the shift of embedding net input to zero.
activation_function: str: The activation function in the embedding net. Supported options are “none”, “gelu_tf”, “linear”, “relu6”, “sigmoid”, “tanh”, “gelu”, “relu”, “softplus”.
precision: str: The precision of the embedding net parameters. Supported options are “float16”, “float64”, “default”, “float32”.
scaling_factor: float: The scaling factor of normalization in calculations of attention weights. If temperature is None, the scaling of attention weights is (N_dim * scaling_factor)**0.5
normalize: bool: Whether to normalize the hidden vectors in attention weights calculation.
temperature: float: If not None, the scaling of attention weights is temperature itself.
smooth_type_embedding: bool: Whether to use smooth process in attention weights calculation.
concat_output_tebd: bool: Whether to concat type embedding at the output of the descriptor.
stripped_type_embedding: bool, Optional: (Deprecated, kept only for compatibility.) Whether to strip the type embedding into a separate embedding network. Setting this parameter to True is equivalent to setting tebd_input_mode to ‘strip’. Setting it to False is equivalent to setting tebd_input_mode to ‘concat’. The default value is None, which means the tebd_input_mode setting will be used instead.
use_econf_tebd: bool, Optional: Whether to use electronic configuration type embedding.
use_tebd_biasbool, Optional: Whether to use bias in the type embedding layer.
type_map: list[str], Optional: A list of strings. Give the name to each type of atoms.
spin: (Only support None to keep consistent with other backend references.) (Not used in this version. Not-none option is not implemented.) The old implementation of deepspin.

References

[1]

Duo Zhang, Hangrui Bi, Fu-Zhi Dai, Wanrun Jiang, Linfeng Zhang, and Han Wang. 2022. DPA-1: Pretraining of Attention-based Deep Potential Model for Molecular Simulation. arXiv preprint arXiv:2208.08236.