import torch import torch.nn as nn import torch.nn.functional as F from base.graph_recommender import GraphRecommender from util.conf import OptionConf from util.sampler import next_batch_pairwise from base.torch_interface import TorchGraphInterface from util.loss_torch import bpr_loss, l2_reg_loss, InfoNCE # import pdb # Paper: Are graph augmentations necessary? simple graph contrastive learning for recommendation. SIGIR'22 class SimGCL(GraphRecommender): def __init__(self, conf, training_set, test_set): # training_set,test_set [[user_id, item_id, float(weight)]...] super(SimGCL, self).__init__(conf, training_set, test_set) args = OptionConf(self.config['SimGCL']) self.cl_rate = float(args['-lambda']) self.eps = float(args['-eps']) self.n_layers = int(args['-n_layer']) self.temp = float(args['-temp']) self.pool='mean' # "[concat, mean, sum, final]" self.decay = 1e-4 self.cl_weight=float(args['-cl_weight']) self.model = SimGCL_Encoder(self.data, self.emb_size, self.eps, self.n_layers) def train(self): model = self.model.cuda() optimizer = torch.optim.Adam(model.parameters(), lr=self.lRate) for epoch in range(self.maxEpoch): for n, batch in enumerate(next_batch_pairwise(self.data, self.batch_size)): user_idx, pos_idx, neg_idx = batch rec_user_emb, rec_item_emb ,rec_user_emb_all_layer,rec_item_emb_all_layer= model() user_emb, pos_item_emb = rec_user_emb[user_idx], rec_item_emb[pos_idx] user_emb_all_layer, pos_item_emb_all_layer = rec_user_emb_all_layer[user_idx], rec_item_emb_all_layer[pos_idx] # <----------------------------------------------------------------------- neg_gcn_embs_mixgcf = [] neg_gcn_embs_mixgcf.append(self.negative_sampling(rec_user_emb_all_layer, rec_item_emb_all_layer, user_idx, neg_idx, pos_idx)) neg_gcn_embs = torch.stack(neg_gcn_embs_mixgcf, dim=1) # pdb.set_trace() rec_loss, _, _= self.create_bpr_loss(user_emb_all_layer, pos_item_emb_all_layer, neg_gcn_embs) # MixGCF的BPR loss计算方式 neg_item_emb = self.pooling(neg_gcn_embs.view(-1, neg_gcn_embs.shape[2], neg_gcn_embs.shape[3])) # 这里是Mix的方法 # neg_item_emb = rec_item_emb[neg_idx[:,:1]] # 这里是SimGCL原本的方法 user_emb = self.pooling(user_emb_all_layer) pos_item_emb = self.pooling(pos_item_emb_all_layer) # <----------------------------------------------------------------------- # rec_loss = bpr_loss(user_emb, pos_item_emb, neg_item_emb) #人家原本的BPR loss计算方式 cl_loss = self.cl_rate * self.cal_cl_loss([user_idx,pos_idx]) * self.cl_weight batch_loss = rec_loss + l2_reg_loss(self.reg, user_emb, pos_item_emb,neg_item_emb) + cl_loss # neg_item_emb.shape = torch.Size([2048, 64]) # Backward and optimize optimizer.zero_grad() batch_loss.backward() optimizer.step() # if n % 100==0: # print('training:', epoch + 1, 'batch', n, 'rec_loss:', rec_loss.item(), 'cl_loss', cl_loss.item()) with torch.no_grad(): self.user_emb, self.item_emb,_,_ = self.model() self.fast_evaluation(epoch) self.user_emb, self.item_emb = self.best_user_emb, self.best_item_emb def cal_cl_loss(self, idx): u_idx = torch.unique(torch.Tensor(idx[0]).type(torch.long)).cuda() i_idx = torch.unique(torch.Tensor(idx[1]).type(torch.long)).cuda() user_view_1, item_view_1 ,_,_= self.model(perturbed=True) user_view_2, item_view_2 ,_,_= self.model(perturbed=True) user_cl_loss = InfoNCE(user_view_1[u_idx], user_view_2[u_idx], self.temp) item_cl_loss = InfoNCE(item_view_1[i_idx], item_view_2[i_idx], self.temp) return user_cl_loss + item_cl_loss def save(self): with torch.no_grad(): self.best_user_emb, self.best_item_emb ,_,_= self.model.forward() def predict(self, u): u = self.data.get_user_id(u) score = torch.matmul(self.user_emb[u], self.item_emb.transpose(0, 1)) return score.cpu().numpy() def negative_sampling(self, user_gcn_emb, item_gcn_emb, user, neg_candidates, pos_item): """user : 指的是uid neg_candidates : neg_item_id pos_item : pos_item_id """ batch_size = len(user) s_e, p_e = user_gcn_emb[user], item_gcn_emb[pos_item] # [batch_size, n_hops+1, channel] if self.pool != 'concat': s_e = self.pooling(s_e).unsqueeze(dim=1) """positive mixing""" seed = torch.rand(batch_size, 1, p_e.shape[1], 1).to(p_e.device) # (0, 1) n_e = item_gcn_emb[neg_candidates] # [batch_size, n_negs, n_hops, channel] n_e_ = seed * p_e.unsqueeze(dim=1) + (1 - seed) * n_e # mixing 往里面插入一个维度 # n_e.shape = torch.Size([1024, 64, 4, 64]) # p_e.shape = torch.Size([1024, 4, 64]) # seed.shape = torch.Size([1024, 1, 4, 1]) """hop mixing""" scores = (s_e.unsqueeze(dim=1) * n_e_).sum(dim=-1) # [batch_size, n_negs, n_hops+1] indices = torch.max(scores, dim=1)[1].detach() neg_items_emb_ = n_e_.permute([0, 2, 1, 3]) # [batch_size, n_hops+1, n_negs, channel] # [batch_size, n_hops+1, channel] return neg_items_emb_[[[i] for i in range(batch_size)], range(neg_items_emb_.shape[1]), indices, :] def pooling(self, embeddings): # [-1, n_hops, channel] if self.pool == 'mean': return embeddings.mean(dim=1) elif self.pool == 'sum': return embeddings.sum(dim=1) elif self.pool == 'concat': return embeddings.view(embeddings.shape[0], -1) else: # final return embeddings[:, -1, :] def create_bpr_loss(self, user_gcn_emb, pos_gcn_embs, neg_gcn_embs): # user_gcn_emb: [batch_size, n_hops+1, channel] = torch.Size([1024, 4, 64]) # pos_gcn_embs: [batch_size, n_hops+1, channel] = torch.Size([1024, 4, 64]) # neg_gcn_embs: [batch_size, K, n_hops+1, channel] = torch.Size([1024, 1, 4, 64]) 这里变成4维了，和其他两个tensor维度对不上了 batch_size = user_gcn_emb.shape[0] u_e = self.pooling(user_gcn_emb) # u_e.shape = torch.Size([1024, 256]) pos_e = self.pooling(pos_gcn_embs) neg_e = self.pooling(neg_gcn_embs.view(-1, neg_gcn_embs.shape[2], neg_gcn_embs.shape[3])).view(batch_size, 1, -1) # 这里把 neg_gcn_embs([2048, 1, 4, 64]) view成为([2048, 4, 64])，再 pooling ，再恢复为([2048, 1, 256]) pos_scores = torch.sum(torch.mul(u_e, pos_e), axis=1) # pos_scores.shape = torch.Size([1024]) neg_scores = torch.sum(torch.mul(u_e.unsqueeze(dim=1), neg_e), axis=-1) # [batch_size, K] mf_loss = torch.mean(torch.log(1+torch.exp(neg_scores - pos_scores.unsqueeze(dim=1)).sum(dim=1))) # cul regularizer regularize = (torch.norm(user_gcn_emb[:, 0, :]) ** 2 + torch.norm(pos_gcn_embs[:, 0, :]) ** 2 + torch.norm(neg_gcn_embs[:, :, 0, :]) ** 2) / 2 # take hop=0 emb_loss = self.decay * regularize / batch_size return mf_loss + emb_loss, mf_loss, emb_loss class SimGCL_Encoder(nn.Module): def __init__(self, data, emb_size, eps, n_layers): super(SimGCL_Encoder, self).__init__() self.data = data self.eps = eps self.emb_size = emb_size self.n_layers = n_layers self.norm_adj = data.norm_adj self.embedding_dict = self._init_model() self.sparse_norm_adj = TorchGraphInterface.convert_sparse_mat_to_tensor(self.norm_adj).cuda() def _init_model(self): initializer = nn.init.xavier_uniform_ embedding_dict = nn.ParameterDict({ 'user_emb': nn.Parameter(initializer(torch.empty(self.data.user_num, self.emb_size))), 'item_emb': nn.Parameter(initializer(torch.empty(self.data.item_num, self.emb_size))), }) return embedding_dict def forward(self, perturbed=False): ego_embeddings = torch.cat([self.embedding_dict['user_emb'], self.embedding_dict['item_emb']], 0) embs=[] # 作者说的，要跳过最原始的E_0 all_embeddings = [] for k in range(self.n_layers): ego_embeddings = torch.sparse.mm(self.sparse_norm_adj, ego_embeddings) if perturbed: random_noise = torch.rand_like(ego_embeddings).cuda() ego_embeddings += torch.sign(ego_embeddings) * F.normalize(random_noise, dim=-1) * self.eps all_embeddings.append(ego_embeddings) embs.append(ego_embeddings) all_embeddings = torch.stack(all_embeddings, dim=1) all_embeddings = torch.mean(all_embeddings, dim=1) user_all_embeddings, item_all_embeddings = torch.split(all_embeddings, [self.data.user_num, self.data.item_num]) embs = torch.stack(embs,dim=1) user_all_embeddings_all_layer, item_all_embeddings_all_layer = torch.split(embs, [self.data.user_num, self.data.item_num]) return user_all_embeddings, item_all_embeddings,user_all_embeddings_all_layer,item_all_embeddings_all_layer