TAAS / graphormer.py
zy414775's picture
Update graphormer.py
8e7dc56
from copy import deepcopy
from torch.nn.init import xavier_uniform_
import torch.nn.functional as F
from torch.nn import Parameter
from torch.nn.init import normal_
import torch.utils.checkpoint
from torch import Tensor, device
from .TAAS_utils import *
from transformers.modeling_utils import ModuleUtilsMixin
from fairseq import utils
from fairseq.models import (
FairseqEncoder,
FairseqEncoderModel,
register_model,
register_model_architecture,
)
from fairseq.modules import (
LayerNorm,
)
from fairseq.utils import safe_hasattr
def init_params(module, n_layers):
if isinstance(module, nn.Linear):
module.weight.data.normal_(mean=0.0, std=0.02 / math.sqrt(n_layers))
if module.bias is not None:
module.bias.data.zero_()
if isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=0.02)
@torch.jit.script
def softmax_dropout(input, dropout_prob: float, is_training: bool):
return F.dropout(F.softmax(input, -1), dropout_prob, is_training)
class SelfMultiheadAttention(nn.Module):
def __init__(
self,
embed_dim,
num_heads,
dropout=0.0,
bias=True,
scaling_factor=1,
):
super().__init__()
self.embed_dim = embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
assert (self.head_dim * num_heads == self.embed_dim), "embed_dim must be divisible by num_heads"
self.scaling = (self.head_dim * scaling_factor) ** -0.5
self.linear_q = nn.Linear(self.embed_dim, self.num_heads * self.head_dim)
self.linear_k = nn.Linear(self.embed_dim, self.num_heads * self.head_dim)
self.linear_v = nn.Linear(self.embed_dim, self.num_heads * self.head_dim)
self.out_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=bias)
def forward(
self,
query: Tensor,
attn_bias: Tensor = None,
) -> Tensor:
n_graph, n_node, embed_dim = query.size()
# q, k, v = self.in_proj(query).chunk(3, dim=-1)
_shape = (-1, n_graph * self.num_heads, self.head_dim)
q = self.linear_q(query).contiguous().view(n_graph, -1, self.num_heads, self.head_dim).transpose(1, 2) * self.scaling
k = self.linear_k(query).contiguous().view(n_graph, -1, self.num_heads, self.head_dim).transpose(1, 2)
v = self.linear_v(query).contiguous().view(n_graph, -1, self.num_heads, self.head_dim).transpose(1, 2)
attn_weights = torch.matmul(q, k.transpose(2, 3))
attn_weights = attn_weights + attn_bias
attn_probs = softmax_dropout(attn_weights, self.dropout, self.training)
attn = torch.matmul(attn_probs, v)
attn = attn.transpose(1, 2).contiguous().view(n_graph, -1, embed_dim)
attn = self.out_proj(attn)
return attn
class Graphormer3DEncoderLayer(nn.Module):
"""
Implements a Graphormer-3D Encoder Layer.
"""
def __init__(
self,
embedding_dim: int = 768,
ffn_embedding_dim: int = 3072,
num_attention_heads: int = 8,
dropout: float = 0.1,
attention_dropout: float = 0.1,
activation_dropout: float = 0.1,
) -> None:
super().__init__()
# Initialize parameters
self.embedding_dim = embedding_dim
self.num_attention_heads = num_attention_heads
self.attention_dropout = attention_dropout
self.dropout = dropout
self.activation_dropout = activation_dropout
self.self_attn = SelfMultiheadAttention(self.embedding_dim, num_attention_heads, dropout=attention_dropout)
# layer norm associated with the self attention layer
self.self_attn_layer_norm = nn.LayerNorm(self.embedding_dim)
self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim)
self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim)
self.final_layer_norm = nn.LayerNorm(self.embedding_dim)
def forward(self, x: Tensor, attn_bias: Tensor = None):
residual = x
x = self.self_attn_layer_norm(x)
x = self.self_attn(query=x, attn_bias=attn_bias)
x = F.dropout(x, p=self.dropout, training=self.training)
x = residual + x
residual = x
x = self.final_layer_norm(x)
x = F.gelu(self.fc1(x))
x = F.dropout(x, p=self.activation_dropout, training=self.training)
x = self.fc2(x)
x = F.dropout(x, p=self.dropout, training=self.training)
x = residual + x
return x
from fairseq.models import (
BaseFairseqModel,
register_model,
register_model_architecture,
)
class Graphormer3D(BaseFairseqModel):
def __init__(self):
super().__init__()
self.atom_types = 64
self.edge_types = 64 * 64
self.embed_dim = 768
self.layer_nums = 12
self.ffn_embed_dim = 768
self.blocks = 4
self.attention_heads = 48
self.input_dropout = 0.0
self.dropout = 0.1
self.attention_dropout = 0.1
self.activation_dropout = 0.0
self.node_loss_weight = 15
self.min_node_loss_weight = 1
self.eng_loss_weight = 1
self.num_kernel = 128
self.atom_encoder = nn.Embedding(self.atom_types, self.embed_dim, padding_idx=0)
self.edge_embedding = nn.Embedding(32, self.attention_heads, padding_idx=0)
self.input_dropout = nn.Dropout(0.1)
self.layers = nn.ModuleList(
[
Graphormer3DEncoderLayer(
self.embed_dim,
self.ffn_embed_dim,
num_attention_heads=self.attention_heads,
dropout=self.dropout,
attention_dropout=self.attention_dropout,
activation_dropout=self.activation_dropout,
)
for _ in range(self.layer_nums)
]
)
self.atom_encoder = nn.Embedding(512 * 9 + 1, self.embed_dim, padding_idx=0)
self.edge_encoder = nn.Embedding(512 * 3 + 1, self.attention_heads, padding_idx=0)
self.edge_type = 'multi_hop'
if self.edge_type == 'multi_hop':
self.edge_dis_encoder = nn.Embedding(16 * self.attention_heads * self.attention_heads, 1)
self.spatial_pos_encoder = nn.Embedding(512, self.attention_heads, padding_idx=0)
self.in_degree_encoder = nn.Embedding(512, self.embed_dim, padding_idx=0)
self.out_degree_encoder = nn.Embedding(512, self.embed_dim, padding_idx=0)
self.node_position_ids_encoder = nn.Embedding(10, self.embed_dim, padding_idx=0)
self.final_ln: Callable[[Tensor], Tensor] = nn.LayerNorm(self.embed_dim)
self.engergy_proj: Callable[[Tensor], Tensor] = NonLinear(self.embed_dim, 1)
self.energe_agg_factor: Callable[[Tensor], Tensor] = nn.Embedding(3, 1)
nn.init.normal_(self.energe_agg_factor.weight, 0, 0.01)
self.graph_token = nn.Embedding(1, 768)
self.graph_token_virtual_distance = nn.Embedding(1, self.attention_heads)
K = self.num_kernel
self.gbf: Callable[[Tensor, Tensor], Tensor] = GaussianLayer(K, self.edge_types)
self.bias_proj: Callable[[Tensor], Tensor] = NonLinear(K, self.attention_heads)
self.edge_proj: Callable[[Tensor], Tensor] = nn.Linear(K, self.embed_dim)
self.node_proc: Callable[[Tensor, Tensor, Tensor], Tensor] = NodeTaskHead(self.embed_dim, self.attention_heads)
def forward(self, node_feature, spatial_pos, in_degree, out_degree, edge_type_matrix, edge_input, node_position_ids):
"""
attn_bias:图中节点对之间的最短路径距离超过最短路径限制最大距离(spatial_pos_max)的位置为-inf,其余位置为0,形状为(n_graph, n_node+1, n_node+1)
spatial_pos:图中节点对之间的最短路径长度,形状为(n_graph, n_node, n_node)
x:图中节点的特征,形状为(n_graph, n_node, n_node_features)
in_degree:图中节点的入度,形状为(n_graph, n_node)
out_degree:图中节点的出度,形状为(n_graph, n_node)
edge_input:图中节点对之间的最短路径(限制最短路径最大跳数为multi_hop_max_dist)上的边的特征,形状为(n_graph, n_node, n_node, multi_hop_max_dist, n_edge_features)
attn_edge_type:图的边特征,形状为(n_graph, n_node, n_node, n_edge_features)
:param batch_data:
:return:
"""
# attn_bias, spatial_pos, x = batch_data.attn_bias, batch_data.spatial_pos, batch_data.x
# in_degree, out_degree = batch_data.in_degree, batch_data.out_degree
# edge_input, attn_edge_type = batch_data.edge_input, batch_data.attn_edge_type
# graph_attn_bias
attn_edge_type = self.edge_embedding(edge_type_matrix)
edge_input = self.edge_embedding(edge_input)#.mean(-2)
# 添加虚拟节点表示全图特征表示,之后按照图中正常节点处理
n_graph, n_node = node_feature.size()[:2]
# graph_attn_bias = attn_bias.clone()
# graph_attn_bias = graph_attn_bias.unsqueeze(1).repeat(1, self.attention_heads, 1, 1) # [n_graph, n_head, n_node+1, n_node+1]
# spatial pos
# 空间编码,节点之间最短路径长度对应的可学习标量
# [n_graph, n_node, n_node, n_head] -> [n_graph, n_head, n_node, n_node]
spatial_pos_bias = self.spatial_pos_encoder(spatial_pos).permute(0, 3, 1, 2)
# graph_attn_bias[:, :, 1:, 1:] = graph_attn_bias[:, :, 1:, 1:] + spatial_pos_bias
# graph_attn_bias = spatial_pos_bias
# reset spatial pos here
# 所有节点都和虚拟节点直接有边相连,则所有节点和虚拟节点之间的最短路径长度为1
# t = self.graph_token_virtual_distance.weight.view(1, self.attention_heads, 1)
# graph_attn_bias[:, :, 1:, 0] = graph_attn_bias[:, :, 1:, 0] + t
# graph_attn_bias[:, :, 0, :] = graph_attn_bias[:, :, 0, :] + t
# edge feature
# 每个节点对沿最短路径计算边特征和可学习嵌入点积的平均值,并作为偏置项添加到注意模块中
if self.edge_type == 'multi_hop':
spatial_pos_ = spatial_pos.clone()
spatial_pos_[spatial_pos_ == 0] = 1 # set pad to 1
# set 1 to 1, x > 1 to x - 1
spatial_pos_ = torch.where(spatial_pos_ > 1, spatial_pos_ - 1, spatial_pos_)
# if self.multi_hop_max_dist > 0:
# spatial_pos_ = spatial_pos_.clamp(0, self.multi_hop_max_dist)
# edge_input = edge_input[:, :, :, :self.multi_hop_max_dist, :]
# [n_graph, n_node, n_node, max_dist, n_head]
# edge_input = self.edge_encoder(edge_input).mean(-2)
max_dist = edge_input.size(-2)
edge_input_flat = edge_input.permute(3, 0, 1, 2, 4).reshape(max_dist, -1, self.attention_heads)
edge_input_flat = torch.bmm(edge_input_flat, self.edge_dis_encoder.weight.reshape(-1, self.attention_heads, self.attention_heads)[:max_dist, :, :])
edge_input = edge_input_flat.reshape(max_dist, n_graph, n_node, n_node, self.attention_heads).permute(1, 2, 3, 0, 4)
edge_input = (edge_input.sum(-2) / (spatial_pos_.float().unsqueeze(-1))).permute(0, 3, 1, 2)
else:
# [n_graph, n_node, n_node, n_head] -> [n_graph, n_head, n_node, n_node]
edge_input = self.edge_encoder(attn_edge_type).mean(-2).permute(0, 3, 1, 2)
# graph_attn_bias[:, :, 1:, 1:] = graph_attn_bias[:, :, 1:, 1:] + edge_input
graph_attn_bias = spatial_pos_bias + edge_input
# graph_attn_bias = graph_attn_bias + attn_bias.unsqueeze(1) # reset
# graph_attn_bias = graph_attn_bias.contiguous().view(-1, 6, 6)
# node feauture + graph token
# node_feature = x # self.atom_encoder(x).sum(dim=-2) # [n_graph, n_node, n_hidden]
# if self.flag and perturb is not None:
# node_feature += perturb
node_position_embedding = self.node_position_ids_encoder(node_position_ids)
node_position_embedding = node_position_embedding.contiguous().view(n_graph, n_node, self.embed_dim)
# print(node_position_embedding.shape)
# 根据节点的入度、出度为每个节点分配两个实值嵌入向量,添加到节点特征中作为输入
node_feature = node_feature + self.in_degree_encoder(in_degree) + \
self.out_degree_encoder(out_degree) + node_position_embedding
# print(node_feature.shape)
# graph_token_feature = self.graph_token.weight.unsqueeze(0).repeat(n_graph, 1, 1)
# graph_node_feature = torch.cat([graph_token_feature, node_feature], dim=1)
# transfomrer encoder
output = self.input_dropout(node_feature)#.permute(1, 0, 2)
for enc_layer in self.layers:
output = enc_layer(output, graph_attn_bias)
output = self.final_ln(output)
# output part
# 整个图的表示是最后一层虚拟节点的特征
# if self.dataset_name == 'PCQM4M-LSC':
# # get whole graph rep
# output = self.out_proj(output[:, 0, :])
# else:
# output = self.downstream_out_proj(output[:, 0, :])
# print(output.shape)
return output
@torch.jit.script
def gaussian(x, mean, std):
pi = 3.14159
a = (2 * pi) ** 0.5
return torch.exp(-0.5 * (((x - mean) / std) ** 2)) / (a * std)
class GaussianLayer(nn.Module):
def __init__(self, K=128, edge_types=1024):
super().__init__()
self.K = K
self.means = nn.Embedding(1, K)
self.stds = nn.Embedding(1, K)
self.mul = nn.Embedding(edge_types, 1)
self.bias = nn.Embedding(edge_types, 1)
nn.init.uniform_(self.means.weight, 0, 3)
nn.init.uniform_(self.stds.weight, 0, 3)
nn.init.constant_(self.bias.weight, 0)
nn.init.constant_(self.mul.weight, 1)
def forward(self, x, edge_types):
mul = self.mul(edge_types)
bias = self.bias(edge_types)
x = mul * x.unsqueeze(-1) + bias
x = x.expand(-1, -1, -1, self.K)
mean = self.means.weight.float().view(-1)
std = self.stds.weight.float().view(-1).abs() + 1e-5
return gaussian(x.float(), mean, std).type_as(self.means.weight)
class RBF(nn.Module):
def __init__(self, K, edge_types):
super().__init__()
self.K = K
self.means = nn.parameter.Parameter(torch.empty(K))
self.temps = nn.parameter.Parameter(torch.empty(K))
self.mul: Callable[..., Tensor] = nn.Embedding(edge_types, 1)
self.bias: Callable[..., Tensor] = nn.Embedding(edge_types, 1)
nn.init.uniform_(self.means, 0, 3)
nn.init.uniform_(self.temps, 0.1, 10)
nn.init.constant_(self.bias.weight, 0)
nn.init.constant_(self.mul.weight, 1)
def forward(self, x: Tensor, edge_types):
mul = self.mul(edge_types)
bias = self.bias(edge_types)
x = mul * x.unsqueeze(-1) + bias
mean = self.means.float()
temp = self.temps.float().abs()
return ((x - mean).square() * (-temp)).exp().type_as(self.means)
class NonLinear(nn.Module):
def __init__(self, input, output_size, hidden=None):
super(NonLinear, self).__init__()
if hidden is None:
hidden = input
self.layer1 = nn.Linear(input, hidden)
self.layer2 = nn.Linear(hidden, output_size)
def forward(self, x):
x = F.gelu(self.layer1(x))
x = self.layer2(x)
return x
class NodeTaskHead(nn.Module):
def __init__(
self,
embed_dim: int,
num_heads: int,
):
super().__init__()
self.embed_dim = embed_dim
self.q_proj: Callable[[Tensor], Tensor] = nn.Linear(embed_dim, embed_dim)
self.k_proj: Callable[[Tensor], Tensor] = nn.Linear(embed_dim, embed_dim)
self.v_proj: Callable[[Tensor], Tensor] = nn.Linear(embed_dim, embed_dim)
self.num_heads = num_heads
self.scaling = (embed_dim // num_heads) ** -0.5
self.force_proj1: Callable[[Tensor], Tensor] = nn.Linear(embed_dim, 1)
self.force_proj2: Callable[[Tensor], Tensor] = nn.Linear(embed_dim, 1)
self.force_proj3: Callable[[Tensor], Tensor] = nn.Linear(embed_dim, 1)
def forward(
self,
query: Tensor,
attn_bias: Tensor,
delta_pos: Tensor,
) -> Tensor:
bsz, n_node, _ = query.size()
q = (self.q_proj(query).view(bsz, n_node, self.num_heads, -1).transpose(1, 2) * self.scaling)
k = self.k_proj(query).view(bsz, n_node, self.num_heads, -1).transpose(1, 2)
v = self.v_proj(query).view(bsz, n_node, self.num_heads, -1).transpose(1, 2)
attn = q @ k.transpose(-1, -2) # [bsz, head, n, n]
attn_probs = softmax_dropout(attn.view(-1, n_node, n_node) + attn_bias, 0.1, self.training).view(bsz, self.num_heads, n_node, n_node)
rot_attn_probs = attn_probs.unsqueeze(-1) * delta_pos.unsqueeze(1).type_as(attn_probs) # [bsz, head, n, n, 3]
rot_attn_probs = rot_attn_probs.permute(0, 1, 4, 2, 3)
x = rot_attn_probs @ v.unsqueeze(2) # [bsz, head , 3, n, d]
x = x.permute(0, 3, 2, 1, 4).contiguous().view(bsz, n_node, 3, -1)
f1 = self.force_proj1(x[:, :, 0, :]).view(bsz, n_node, 1)
f2 = self.force_proj2(x[:, :, 1, :]).view(bsz, n_node, 1)
f3 = self.force_proj3(x[:, :, 2, :]).view(bsz, n_node, 1)
cur_force = torch.cat([f1, f2, f3], dim=-1).float()
return cur_force