def iterate_epoch(self,
model,
lr,
epoch,
weight_decay=0,
warmup=0,
lr_decay_rate=1,
lr_decay_every=10,
eval_every=5,
early_stop=False):
eval_model = Evaluater(self.data_dir, model_name=self.model_name)
#es = EarlyStop(self.data_dir[0:-6] + 'early_stopping/', self.model_name, patience=6)
es = EarlyStop('../data_beauty_2core_es/early_stopping/',
self.model_name,
patience=6)
plot_loss_list = []
plot_score_list = []
for i in range(epoch):
plot_loss_list.extend(
self.iterate_train(model,
lr=lr,
weight_decay=weight_decay,
print_every=10000))
# early stop
if early_stop:
pre_model = es.early_stop(model)
if pre_model:
print('Early Stop eposh: {}'.format(i + 1))
return eval_model.topn_map(pre_model)
# lrスケジューリング
if i > warmup:
if (i - warmup) % lr_decay_every == 0:
lr = lr * lr_decay_rate
if (i + 1) % eval_every == 0:
#score = eval_model.topn_precision(model)
#print('epoch: {} precision: {}'.format(i, score))
score = eval_model.topn_map(model)
print('epoch: {} map: {}'.format(i, score))
plot_score_list.append(score)
#self._plot(plot_loss_list)
#self._plot(plot_score_list)
#return eval_model.topn_precision(model)
return eval_model.topn_map(model)
def objective(trial):
start = time.time()
# ハイパラ読み込み
# gamma = trial.suggest_loguniform('gamma', 1e-6, 1e-3)
# lin_model = trial.suggest_categorical('lin_model', ['lasso', 'elastic'])
alpha = trial.suggest_uniform('alpha', 0, 1)
beta = trial.suggest_uniform('beta', 0, 0.5)
data_dirs = [
'../data_luxury_5core/valid1/', '../data_luxury_5core/valid2/'
]
score_sum = 0
for data_dir in data_dirs:
# dataload
dataset = AmazonDataset(data_dir)
# laod model
slim_param = pickle.load(open('best_param_slim.pickle', 'rb'))
slim = train_SLIM(slim_param, data_dir, load=True)
edges = [[r[0], r[1]] for r in dataset.triplet_df.values]
# user-itemとitem-userどちらの辺も追加
for r in dataset.triplet_df.values:
if r[2] == 0:
edges.append([r[1], r[0]])
# load network
G = nx.DiGraph()
G.add_nodes_from([i for i in range(len(dataset.entity_list))])
G.add_edges_from(edges)
evaluater = Evaluater(data_dir)
ranking_mat = get_ranking_mat(G, slim, alpha, beta, dataset)
#score = evaluater.topn_precision(ranking_mat)
score = evaluater.topn_map(ranking_mat)
score_sum += score
mi, sec = time_since(time.time() - start)
print('{}m{}s'.format(mi, sec))
return -1 * score_sum / 2
def objective(trial):
start = time.time()
# hyper parameter
#gamma = trial.suggest_loguniform('gamma', 1e-6, 1e-3)
#lin_model = trial.suggest_categorical('lin_model', ['lasso', 'elastic'])
#slim = train_SLIM(lin_model, gamma)
alpha = trial.suggest_uniform('alpha', 0, 0.5)
beta = trial.suggest_uniform('beta', 0, 0.5)
gamma1 = trial.suggest_uniform('gamma1', 0, 1)
gamma2 = trial.suggest_uniform('gamma2', 0, 1)
gamma3 = trial.suggest_uniform('gamma3', 0, 1)
gamma = [gamma1, gamma2, gamma3]
data_dir = ['../data_luxury_5core/valid1', '../data_luxury_5core/valid2']
score_sum = 0
for i in range(len(data_dir)):
# dataload
dataset = AmazonDataset(data_dir[i], model_name='TransE')
edges = [[r[0], r[1]] for r in dataset.triplet_df.values]
# user-itemとitem-userどちらの辺も追加
for r in dataset.triplet_df.values:
if r[2] == 0:
edges.append([r[1], r[0]])
#user_items_test_dict = pickle.load(open('./data/user_items_test_dict.pickle', 'rb'))
# load network
G = nx.DiGraph()
G.add_nodes_from([i for i in range(len(dataset.entity_list))])
G.add_edges_from(edges)
ranking_mat = get_ranking_mat(G, dataset, model[i], gamma, alpha, beta)
#score = topn_precision(ranking_mat, user_items_test_dict)
evaluater = Evaluater(data_dir[i])
score = evaluater.topn_map(ranking_mat)
score_sum += score
mi, sec = time_since(time.time() - start)
print('{}m{}sec'.format(mi, sec))
return -1 * score_sum / 2
def objective(trial):
start = time.time()
# hyper parameter
alpha = trial.suggest_uniform('alpha', 0, 0.5)
beta = trial.suggest_uniform('beta', 0, 0.5)
gamma1 = trial.suggest_uniform('gamma1', 0, 1)
gamma2 = trial.suggest_uniform('gamma2', 0, 1)
gamma3 = trial.suggest_uniform('gamma3', 0, 1)
gamma = [gamma1, gamma2, gamma3]
data_dir = ['../' + data_path + '/valid1', '../' + data_path + '/valid2']
score_sum = 0
for i in range(len(data_dir)):
# dataload
dataset = AmazonDataset(data_dir[i], model_name='SparseTransE')
# load network
edges = [[r[0], r[1]] for r in dataset.triplet_df.values]
# user-itemとitem-userどちらの辺も追加
for r in dataset.triplet_df.values:
if r[2] == 0:
edges.append([r[1], r[0]])
G = nx.DiGraph()
G.add_nodes_from([i for i in range(len(dataset.entity_list))])
G.add_edges_from(edges)
ranking_mat = get_ranking_mat(G, dataset, model[i], gamma, alpha, beta)
#score = topn_precision(ranking_mat, user_items_test_dict)
evaluater = Evaluater(data_dir[i])
score = evaluater.topn_map(ranking_mat)
score_sum += score
mi, sec = time_since(time.time() - start)
print('{}m{}sec'.format(mi, sec))
return -1 * score_sum / 2
# load param
params = load_params()
alpha = params['alpha']
beta = params['beta']
gamma1 = params['gamma1']
gamma2 = params['gamma2']
gamma3 = params['gamma3']
gamma = [gamma1, gamma2, gamma3]
# dataload
dataset = AmazonDataset(data_dir, model_name='TransE')
# load network
edges = [[r[0], r[1]] for r in dataset.triplet_df.values]
# user-itemとitem-userどちらの辺も追加
for r in dataset.triplet_df.values:
if r[2] == 0:
edges.append([r[1], r[0]])
G = nx.DiGraph()
G.add_nodes_from([i for i in range(len(dataset.entity_list))])
G.add_edges_from(edges)
ranking_mat = get_ranking_mat(G, dataset, model, gamma, alpha, beta)
evaluater = Evaluater(data_dir)
score = evaluater.topn_map(ranking_mat)
mi, sec = time_since(time.time() - start)
print('{}m{}sec'.format(mi, sec))
np.savetxt('score_transe3.txt', np.array([score]))
class Inference():
def __init__(self, data_dir):
# 本当はAmazonDatasetクラスを渡した方が速いが、
self.evaluater = Evaluater(data_dir)
self.dataset = AmazonDataset(data_dir, model_name='TransE')
edges = [[r[0], r[1]] for r in self.dataset.triplet_df.values]
# user-itemとitem-userどちらの辺も追加
for r in self.dataset.triplet_df.values:
if r[2] == 0:
edges.append([r[1], r[0]])
# load network
self.G = nx.DiGraph()
self.G.add_nodes_from(
[i for i in range(len(self.dataset.entity_list))])
self.G.add_edges_from(edges)
def get_score(self, model, gamma, alpha, beta):
ranking_mat = self.get_ranking_mat(model, gamma, alpha, beta)
score = self.evaluater.topn_map(ranking_mat)
return score
def mk_sparse_sim_mat(self, model, gamma):
item_idx = torch.tensor([
self.dataset.entity_list.index(i) for i in self.dataset.item_list
],
dtype=torch.long,
device=device)
user_idx = torch.tensor([
self.dataset.entity_list.index(u) for u in self.dataset.user_list
],
dtype=torch.long,
device=device)
brand_idx = torch.tensor([
self.dataset.entity_list.index(b) for b in self.dataset.brand_list
],
dtype=torch.long,
device=device)
# ここもっと上手く書きたい
item_embed = model.entity_embed(item_idx)
item_sim_mat = F.relu(torch.mm(item_embed, torch.t(item_embed)))
item_sim_mat = gamma[0] * scipy.sparse.csr_matrix(
item_sim_mat.to('cpu').detach().numpy().copy())
user_embed = model.entity_embed(user_idx)
user_sim_mat = F.relu(torch.mm(user_embed, torch.t(user_embed)))
user_sim_mat = gamma[1] * scipy.sparse.csr_matrix(
user_sim_mat.to('cpu').detach().numpy().copy())
brand_embed = model.entity_embed(brand_idx)
brand_sim_mat = F.relu(torch.mm(brand_embed, torch.t(brand_embed)))
brand_sim_mat = gamma[2] * scipy.sparse.csr_matrix(
brand_sim_mat.to('cpu').detach().numpy().copy())
M = scipy.sparse.block_diag(
(item_sim_mat, user_sim_mat, brand_sim_mat))
M_ = np.array(1 - M.sum(axis=1) / np.max(M.sum(axis=1)))
M = M / np.max(M.sum(axis=1)) + scipy.sparse.diags(M_.transpose()[0])
return M
def pagerank_scipy(self,
sim_mat,
personal_vec=None,
alpha=0.85,
beta=0.01,
max_iter=100,
tol=1.0e-6,
weight='weight',
dangling=None):
N = len(self.G)
if N == 0:
return {}
nodelist = self.G.nodes()
M = nx.to_scipy_sparse_matrix(self.G,
nodelist=nodelist,
weight=weight,
dtype=float)
S = scipy.array(M.sum(axis=1)).flatten()
S[S != 0] = 1.0 / S[S != 0]
Q = scipy.sparse.spdiags(S.T, 0, *M.shape, format='csr')
M = Q * M
# 遷移行列とsim_matを統合
#sim_mat = mk_sparse_sim_mat(G, item_mat)
M = beta * M + (1 - beta) * sim_mat
# initial vector
x = scipy.repeat(1.0 / N, N)
# Personalization vector
p = personal_vec
dangling_weights = p
is_dangling = scipy.where(S == 0)[0]
#print(x.shape)
#print(M.shape)
#print(p.shape)
ppr_mat = []
for i in range(p.shape[1]):
ppr = self.power_iterate(N, M, x, p[:, i], dangling_weights[:, i],
is_dangling, alpha, max_iter, tol)
ppr_mat.append(ppr)
#if i > 100:
# print(np.array(ppr_mat).shape)
# break
return np.array(ppr_mat)
def power_iterate(self,
N,
M,
x,
p,
dangling_weights,
is_dangling,
alpha,
max_iter=500,
tol=1.0e-6):
#print(M.shape)
#print(x.shape)
#print(p.shape)
# power iteration: make up to max_iter iterations
for i in range(max_iter):
xlast = x
x = alpha * (x * M + sum(x[is_dangling]) * dangling_weights) + \
(1 - alpha) * p
# check convergence, l1 norm
x = x / x.sum()
err = scipy.absolute(x - xlast).sum()
if err < N * tol:
#return dict(zip(nodelist, map(float, x)))
#print(i)
return x
# pagerankの収束ちゃんとやっとく
#print(x.sum())
#print(err)
#print(N * tol)
#raise NetworkXError('pagerank_scipy: power iteration failed to converge '
#'in %d iterations.' % max_iter)
#return dict(zip(nodelist, map(float, x)))
return x
def item_ppr(self, sim_mat, alpha, beta):
# personal_vecを作る(eneity_size * user_size)
user_idx = [
self.dataset.entity_list.index(u) for u in self.dataset.user_list
]
personal_vec = []
for u in user_idx:
val = np.zeros(len(self.G.nodes()))
val[u] = 1
personal_vec.append(val[np.newaxis, :])
personal_vec = np.concatenate(personal_vec, axis=0).transpose()
#ppr = pagerank_torch(G, sim_mat, personal_vec, alpha, beta)
ppr = self.pagerank_scipy(sim_mat, personal_vec, alpha, beta)
item_idx = [
self.dataset.entity_list.index(i) for i in self.dataset.item_list
]
pred = ppr[:, item_idx]
#print(pred.shape)
return pred
def get_ranking_mat(self, model, gamma, alpha=0.85, beta=0.01):
ranking_mat = []
#sim_mat = reconstruct_kg(model)
sim_mat = self.mk_sparse_sim_mat(model, gamma)
pred = self.item_ppr(sim_mat, alpha, beta)
#print(pred.shape)
for i in range(len(self.dataset.user_list)):
sorted_idx = np.argsort(np.array(pred[i]))[::-1]
ranking_mat.append(sorted_idx)
#break
return ranking_mat
