''' Adversarial Attacks on Neural Networks for Graph Data. ICML 2018. https://arxiv.org/abs/1806.02371 Author's Implementation https://github.com/Hanjun-Dai/graph_adversarial_attack This part of code is adopted from the author's implementation (Copyright (c) 2018 Dai, Hanjun and Li, Hui and Tian, Tian and Huang, Xin and Wang, Lin and Zhu, Jun and Song, Le) but modified to be integrated into the repository. ''' import os import sys import numpy as np import torch import networkx as nx import random from torch.nn.parameter import Parameter import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from tqdm import tqdm from deeprobust.graph.rl.env import GraphNormTool class QNetNode(nn.Module): def __init__(self, node_features, node_labels, list_action_space, bilin_q=1, embed_dim=64, mlp_hidden=64, max_lv=1, gm='mean_field', device='cpu'): ''' bilin_q: bilinear q or not mlp_hidden: mlp hidden layer size mav_lv: max rounds of message passing ''' super(QNetNode, self).__init__() self.node_features = node_features self.node_labels = node_labels self.list_action_space = list_action_space self.total_nodes = len(list_action_space) self.bilin_q = bilin_q self.embed_dim = embed_dim self.mlp_hidden = mlp_hidden self.max_lv = max_lv self.gm = gm if bilin_q: last_wout = embed_dim else: last_wout = 1 self.bias_target = Parameter(torch.Tensor(1, embed_dim)) if mlp_hidden: self.linear_1 = nn.Linear(embed_dim * 2, mlp_hidden) self.linear_out = nn.Linear(mlp_hidden, last_wout) else: self.linear_out = nn.Linear(embed_dim * 2, last_wout) self.w_n2l = Parameter(torch.Tensor(node_features.size()[1], embed_dim)) self.bias_n2l = Parameter(torch.Tensor(embed_dim)) self.bias_picked = Parameter(torch.Tensor(1, embed_dim)) self.conv_params = nn.Linear(embed_dim, embed_dim) self.norm_tool = GraphNormTool(normalize=True, gm=self.gm, device=device) weights_init(self) def make_spmat(self, n_rows, n_cols, row_idx, col_idx): idxes = torch.LongTensor([[row_idx], [col_idx]]) values = torch.ones(1) sp = torch.sparse.FloatTensor(idxes, values, torch.Size([n_rows, n_cols])) if next(self.parameters()).is_cuda: sp = sp.cuda() return sp def forward(self, time_t, states, actions, greedy_acts=False, is_inference=False): if self.node_features.data.is_sparse: input_node_linear = torch.spmm(self.node_features, self.w_n2l) else: input_node_linear = torch.mm(self.node_features, self.w_n2l) input_node_linear += self.bias_n2l # TODO the number of target nodes is batch_size, it actually parallizes target_nodes, batch_graph, picked_nodes = zip(*states) list_pred = [] prefix_sum = [] for i in range(len(batch_graph)): region = self.list_action_space[target_nodes[i]] node_embed = input_node_linear.clone() if picked_nodes is not None and picked_nodes[i] is not None: with torch.set_grad_enabled(mode=not is_inference): picked_sp = self.make_spmat(self.total_nodes, 1, picked_nodes[i], 0) node_embed += torch.spmm(picked_sp, self.bias_picked) region = self.list_action_space[picked_nodes[i]] if not self.bilin_q: with torch.set_grad_enabled(mode=not is_inference): # with torch.no_grad(): target_sp = self.make_spmat(self.total_nodes, 1, target_nodes[i], 0) node_embed += torch.spmm(target_sp, self.bias_target) with torch.set_grad_enabled(mode=not is_inference): device = self.node_features.device adj = self.norm_tool.norm_extra( batch_graph[i].get_extra_adj(device)) lv = 0 input_message = node_embed node_embed = F.relu(input_message) while lv < self.max_lv: n2npool = torch.spmm(adj, node_embed) node_linear = self.conv_params( n2npool ) merged_linear = node_linear + input_message node_embed = F.relu(merged_linear) lv += 1 target_embed = node_embed[target_nodes[i], :].view(-1, 1) if region is not None: node_embed = node_embed[region] graph_embed = torch.mean(node_embed, dim=0, keepdim=True) if actions is None: graph_embed = graph_embed.repeat(node_embed.size()[0], 1) else: if region is not None: act_idx = region.index(actions[i]) else: act_idx = actions[i] node_embed = node_embed[act_idx, :].view(1, -1) embed_s_a = torch.cat((node_embed, graph_embed), dim=1) if self.mlp_hidden: embed_s_a = F.relu( self.linear_1(embed_s_a) ) raw_pred = self.linear_out(embed_s_a) if self.bilin_q: raw_pred = torch.mm(raw_pred, target_embed) list_pred.append(raw_pred) if greedy_acts: actions, _ = node_greedy_actions(target_nodes, picked_nodes, list_pred, self) return actions, list_pred class NStepQNetNode(nn.Module): def __init__(self, num_steps, node_features, node_labels, list_action_space, bilin_q=1, embed_dim=64, mlp_hidden=64, max_lv=1, gm='mean_field', device='cpu'): super(NStepQNetNode, self).__init__() self.node_features = node_features self.node_labels = node_labels self.list_action_space = list_action_space self.total_nodes = len(list_action_space) list_mod = [] for i in range(0, num_steps): # list_mod.append(QNetNode(node_features, node_labels, list_action_space)) list_mod.append(QNetNode(node_features, node_labels, list_action_space, bilin_q, embed_dim, mlp_hidden, max_lv, gm=gm, device=device)) self.list_mod = nn.ModuleList(list_mod) self.num_steps = num_steps def forward(self, time_t, states, actions, greedy_acts = False, is_inference=False): assert time_t >= 0 and time_t < self.num_steps return self.list_mod[time_t](time_t, states, actions, greedy_acts, is_inference) def glorot_uniform(t): if len(t.size()) == 2: fan_in, fan_out = t.size() elif len(t.size()) == 3: # out_ch, in_ch, kernel for Conv 1 fan_in = t.size()[1] * t.size()[2] fan_out = t.size()[0] * t.size()[2] else: fan_in = np.prod(t.size()) fan_out = np.prod(t.size()) limit = np.sqrt(6.0 / (fan_in + fan_out)) t.uniform_(-limit, limit) def _param_init(m): if isinstance(m, Parameter): glorot_uniform(m.data) elif isinstance(m, nn.Linear): m.bias.data.zero_() glorot_uniform(m.weight.data) def weights_init(m): for p in m.modules(): if isinstance(p, nn.ParameterList): for pp in p: _param_init(pp) else: _param_init(p) for name, p in m.named_parameters(): if not '.' in name: # top-level parameters _param_init(p) def node_greedy_actions(target_nodes, picked_nodes, list_q, net): assert len(target_nodes) == len(list_q) actions = [] values = [] for i in range(len(target_nodes)): region = net.list_action_space[target_nodes[i]] if picked_nodes is not None and picked_nodes[i] is not None: region = net.list_action_space[picked_nodes[i]] if region is None: assert list_q[i].size()[0] == net.total_nodes else: assert len(region) == list_q[i].size()[0] val, act = torch.max(list_q[i], dim=0) values.append(val) if region is not None: act = region[act.data.cpu().numpy()[0]] # act = Variable(torch.LongTensor([act])) act = torch.LongTensor([act]) actions.append(act) else: actions.append(act) return torch.cat(actions, dim=0).data, torch.cat(values, dim=0).data