def __init__(self,
embedding_size,
user_total,
item_total,
):
super(BPRMF, self).__init__()
self.embedding_size = embedding_size
self.user_total = user_total
self.item_total = item_total
self.is_pretrained = False
# init user and item embeddings
user_weight = torch.FloatTensor(self.user_total, self.embedding_size)
item_weight = torch.FloatTensor(self.item_total, self.embedding_size)
nn.init.xavier_uniform(user_weight)
nn.init.xavier_uniform(item_weight)
# init user and item embeddings
self.user_embeddings = nn.Embedding(self.user_total, self.embedding_size)
self.item_embeddings = nn.Embedding(self.item_total, self.embedding_size)
self.user_embeddings.weight = nn.Parameter(user_weight)
self.item_embeddings.weight = nn.Parameter(item_weight)
normalize_user_emb = F.normalize(self.user_embeddings.weight.data, p=2, dim=1)
normalize_item_emb = F.normalize(self.item_embeddings.weight.data, p=2, dim=1)
self.user_embeddings.weight.data = normalize_user_emb
self.item_embeddings.weight.data = normalize_item_emb
self.user_embeddings = to_gpu(self.user_embeddings)
self.item_embeddings = to_gpu(self.item_embeddings)
def __init__(self,
L1_flag,
embedding_size,
ent_total,
rel_total
):
super(TransEModel, self).__init__()
self.L1_flag = L1_flag
self.embedding_size = embedding_size
self.ent_total = ent_total
self.rel_total = rel_total
self.is_pretrained = False
ent_weight = torch.FloatTensor(self.ent_total, self.embedding_size)
rel_weight = torch.FloatTensor(self.rel_total, self.embedding_size)
nn.init.xavier_uniform(ent_weight)
nn.init.xavier_uniform(rel_weight)
# init user and item embeddings
self.ent_embeddings = nn.Embedding(self.ent_total, self.embedding_size)
self.rel_embeddings = nn.Embedding(self.rel_total, self.embedding_size)
self.ent_embeddings.weight = nn.Parameter(ent_weight)
self.rel_embeddings.weight = nn.Parameter(rel_weight)
normalize_ent_emb = F.normalize(self.ent_embeddings.weight.data, p=2, dim=1)
normalize_rel_emb = F.normalize(self.rel_embeddings.weight.data, p=2, dim=1)
self.ent_embeddings.weight.data = normalize_ent_emb
self.rel_embeddings.weight.data = normalize_rel_emb
self.ent_embeddings = to_gpu(self.ent_embeddings)
self.rel_embeddings = to_gpu(self.rel_embeddings)
def __init__(self, L1_flag, embedding_size, user_total, item_total,
preference_total, use_st_gumbel):
super(TransUPModel, self).__init__()
self.L1_flag = L1_flag
self.embedding_size = embedding_size
self.user_total = user_total
self.item_total = item_total
self.preference_total = preference_total
self.is_pretrained = False
self.use_st_gumbel = use_st_gumbel
user_weight = torch.FloatTensor(self.user_total, self.embedding_size)
item_weight = torch.FloatTensor(self.item_total, self.embedding_size)
pref_weight = torch.FloatTensor(self.preference_total,
self.embedding_size)
norm_weight = torch.FloatTensor(self.preference_total,
self.embedding_size)
nn.init.xavier_uniform(user_weight)
nn.init.xavier_uniform(item_weight)
nn.init.xavier_uniform(pref_weight)
nn.init.xavier_uniform(norm_weight)
# init user and item embeddings
self.user_embeddings = nn.Embedding(self.user_total,
self.embedding_size)
self.item_embeddings = nn.Embedding(self.item_total,
self.embedding_size)
self.user_embeddings.weight = nn.Parameter(user_weight)
self.item_embeddings.weight = nn.Parameter(item_weight)
normalize_user_emb = F.normalize(self.user_embeddings.weight.data,
p=2,
dim=1)
normalize_item_emb = F.normalize(self.item_embeddings.weight.data,
p=2,
dim=1)
self.user_embeddings.weight.data = normalize_user_emb
self.item_embeddings.weight.data = normalize_item_emb
# init preference parameters
self.pref_embeddings = nn.Embedding(self.preference_total,
self.embedding_size)
self.pref_norm_embeddings = nn.Embedding(self.preference_total,
self.embedding_size)
self.pref_embeddings.weight = nn.Parameter(pref_weight)
self.pref_norm_embeddings.weight = nn.Parameter(norm_weight)
normalize_pref_emb = F.normalize(self.pref_embeddings.weight.data,
p=2,
dim=1)
normalize_norm_emb = F.normalize(self.pref_norm_embeddings.weight.data,
p=2,
dim=1)
self.pref_embeddings.weight.data = normalize_pref_emb
self.pref_norm_embeddings.weight.data = normalize_norm_emb
self.user_embeddings = to_gpu(self.user_embeddings)
self.item_embeddings = to_gpu(self.item_embeddings)
self.pref_embeddings = to_gpu(self.pref_embeddings)
self.pref_norm_embeddings = to_gpu(self.pref_norm_embeddings)
def evaluateRec(FLAGS, model, eval_iter, eval_dict, all_dicts, i_map, logger, eval_descending=True, is_report=False):
# Evaluate
total_batches = len(eval_iter)
# processing bar
pbar = tqdm(total=total_batches)
pbar.set_description("Run Eval")
all_i_var = None
if FLAGS.share_embeddings:
all_i_ids = [i_map[i] for i in range(len(i_map))]
all_i_var = to_gpu(V(torch.LongTensor(all_i_ids)))
model.eval()
model.disable_grad()
results = []
for u_ids in eval_iter:
u_var = to_gpu(V(torch.LongTensor(u_ids)))
# batch * item
scores = model.evaluateRec(u_var, all_i_ids=all_i_var)
preds = zip(u_ids, scores.data.cpu().numpy())
results.extend( evalRecProcess(list(preds), eval_dict, all_dicts=all_dicts, descending=eval_descending, num_processes=FLAGS.num_processes, topn=FLAGS.topn, queue_limit=FLAGS.max_queue) )
pbar.update(1)
pbar.close()
performances = [result[:5] for result in results]
f1, p, r, hit, ndcg = np.array(performances).mean(axis=0)
logger.info("f1:{:.4f}, p:{:.4f}, r:{:.4f}, hit:{:.4f}, ndcg:{:.4f}, topn:{}.".format(f1, p, r, hit, ndcg, FLAGS.topn))
if is_report:
predict_tuples = [result[-1] for result in results]
for pred_tuple in predict_tuples:
u_id = pred_tuple[0]
top_ids = pred_tuple[1]
gold_ids = list(pred_tuple[2])
if FLAGS.model_type in ["transup", "jtransup"]:
for d in all_dicts:
gold_ids += list(d.get(u_id, set()))
gold_ids += list(eval_dict.get(u_id, set()))
u_var = to_gpu(V(torch.LongTensor([u_id])))
i_var = to_gpu(V(torch.LongTensor(gold_ids)))
probs, _, _ = model.reportPreference(u_var, i_var)
max_rel_index = torch.max(probs, 1)[1]
gold_strs = ",".join(["{}({})".format(ir[0], ir[1]) for ir in zip(gold_ids, max_rel_index.data.tolist())])
else:
gold_strs = ",".join([str(i) for i in gold_ids])
logger.info("user:{}\tgold:{}\ttop:{}".format(u_id, gold_strs, ",".join([str(i) for i in top_ids])))
model.enable_grad()
return f1, p, r, hit, ndcg
def __init__(self, L1_flag, embedding_size, user_total, item_total,
entity_total, relation_total):
super(CFKG, self).__init__()
self.L1_flag = L1_flag
self.embedding_size = embedding_size
self.user_total = user_total
self.item_total = item_total
self.ent_total = entity_total
# add buy relation between user and item
self.rel_total = relation_total + 1
self.is_pretrained = False
# init user embeddings
user_weight = torch.FloatTensor(self.user_total, self.embedding_size)
nn.init.xavier_uniform(user_weight)
self.user_embeddings = nn.Embedding(self.user_total,
self.embedding_size)
self.user_embeddings.weight = nn.Parameter(user_weight)
normalize_user_emb = F.normalize(self.user_embeddings.weight.data,
p=2,
dim=1)
self.user_embeddings.weight.data = normalize_user_emb
self.user_embeddings = to_gpu(self.user_embeddings)
# init entity and relation embeddings
ent_weight = torch.FloatTensor(self.ent_total, self.embedding_size)
rel_weight = torch.FloatTensor(self.rel_total, self.embedding_size)
nn.init.xavier_uniform(ent_weight)
nn.init.xavier_uniform(rel_weight)
self.ent_embeddings = nn.Embedding(self.ent_total, self.embedding_size)
self.rel_embeddings = nn.Embedding(self.rel_total, self.embedding_size)
self.ent_embeddings.weight = nn.Parameter(ent_weight)
self.rel_embeddings.weight = nn.Parameter(rel_weight)
normalize_ent_emb = F.normalize(self.ent_embeddings.weight.data,
p=2,
dim=1)
normalize_rel_emb = F.normalize(self.rel_embeddings.weight.data,
p=2,
dim=1)
self.ent_embeddings.weight.data = normalize_ent_emb
self.rel_embeddings.weight.data = normalize_rel_emb
self.ent_embeddings = to_gpu(self.ent_embeddings)
self.rel_embeddings = to_gpu(self.rel_embeddings)
# share embedding
self.item_embeddings = self.ent_embeddings
def forward(self, ratings, triples, is_rec=True):
if is_rec and ratings is not None:
u_ids, i_ids = ratings
e_ids = self.paddingItems(i_ids.data, self.ent_total - 1)
e_var = to_gpu(V(torch.LongTensor(e_ids)))
u_e = self.user_embeddings(u_ids)
i_e = self.item_embeddings(i_ids)
e_e = self.ent_embeddings(e_var)
ie_e = i_e + e_e
score = torch.bmm(u_e.unsqueeze(1), ie_e.unsqueeze(2)).squeeze()
elif not is_rec and triples is not None:
h, t, r = triples
h_e = self.ent_embeddings(h)
t_e = self.ent_embeddings(t)
r_e = self.rel_embeddings(r)
proj_e = self.proj_embeddings(r)
proj_h_e = projection_transR_pytorch(h_e, proj_e)
proj_t_e = projection_transR_pytorch(t_e, proj_e)
if self.L1_flag:
score = torch.sum(torch.abs(proj_h_e + r_e - proj_t_e), 1)
else:
score = torch.sum((proj_h_e + r_e - proj_t_e)**2, 1)
else:
raise NotImplementedError
return score
def forward(self, ratings, triples, is_rec=True):
if is_rec and ratings is not None:
u_ids, i_ids = ratings
batch_size = len(u_ids)
u_e = self.user_embeddings(u_ids)
i_e = self.item_embeddings(i_ids)
buy_e = self.rel_embeddings(
to_gpu(V(torch.LongTensor([self.rel_total - 1]))))
buy_e_expand = buy_e.expand(batch_size, self.embedding_size)
# L1 distance
if self.L1_flag:
score = torch.sum(torch.abs(u_e + buy_e_expand - i_e), 1)
# L2 distance
else:
score = torch.sum((u_e + buy_e_expand - i_e)**2, 1)
elif not is_rec and triples is not None:
h, t, r = triples
h_e = self.ent_embeddings(h)
t_e = self.ent_embeddings(t)
r_e = self.rel_embeddings(r)
# L1 distance
if self.L1_flag:
score = torch.sum(torch.abs(h_e + r_e - t_e), 1)
# L2 distance
else:
score = torch.sum((h_e + r_e - t_e)**2, 1)
else:
raise NotImplementedError
return score
def evaluateRec(self, u_ids, all_i_ids=None):
batch_size = len(u_ids)
all_i = self.item_embeddings(
all_i_ids
) if all_i_ids is not None and self.is_share else self.item_embeddings.weight
item_total, dim = all_i.size()
u = self.user_embeddings(u_ids)
# expand u and i to pair wise match, batch * item * dim
u_e = u.expand(item_total, batch_size, dim).permute(1, 0, 2)
i_e = all_i.expand(batch_size, item_total, dim)
e_ids = self.paddingItems(
all_i_ids.data if all_i_ids is not None else self.i_map,
self.ent_total - 1)
e_var = to_gpu(V(torch.LongTensor(e_ids)))
e_e = self.ent_embeddings(e_var).expand(batch_size, item_total, dim)
ie_e = i_e + e_e
# batch * item * dim
_, r_e, norm = self.getPreferences(u_e,
ie_e,
use_st_gumbel=self.use_st_gumbel)
proj_u_e = projection_transH_pytorch(u_e, norm)
proj_i_e = projection_transH_pytorch(ie_e, norm)
# batch * item
if self.L1_flag:
score = torch.sum(torch.abs(proj_u_e + r_e - proj_i_e), 2)
else:
score = torch.sum((proj_u_e + r_e - proj_i_e)**2, 2)
return score
def evaluateRec(self, u_ids, all_i_ids=None):
batch_size = len(u_ids)
all_i = self.item_embeddings(
all_i_ids
) if all_i_ids is not None else self.item_embeddings.weight
item_total, dim = all_i.size()
# batch * dim
u_e = self.user_embeddings(u_ids)
# batch * item * dim
u_e_expand = u_e.expand(item_total, batch_size, dim).permute(1, 0, 2)
buy_e = self.rel_embeddings(
to_gpu(V(torch.LongTensor([self.rel_total - 1]))))
buy_e_expand = buy_e.expand(batch_size, item_total, dim)
c_i_e = u_e_expand + buy_e_expand
# batch * item * dim
i_expand = all_i.expand(batch_size, item_total, dim)
if self.L1_flag:
score = torch.sum(torch.abs(c_i_e - i_expand), 2)
else:
score = torch.sum((c_i_e - i_expand)**2, 2)
return score
def __init__(self, L1_flag, embedding_size, ent_total, rel_total):
super(TransRModel, self).__init__()
self.L1_flag = L1_flag
self.embedding_size = embedding_size
self.ent_total = ent_total
self.rel_total = rel_total
self.is_pretrained = False
self.max_entity_batch = 10
ent_weight = torch.FloatTensor(self.ent_total, self.embedding_size)
rel_weight = torch.FloatTensor(self.rel_total, self.embedding_size)
proj_weight = torch.FloatTensor(
self.rel_total, self.embedding_size * self.embedding_size)
nn.init.xavier_uniform(ent_weight)
nn.init.xavier_uniform(rel_weight)
if self.is_pretrained:
nn.init.eye(proj_weight)
proj_weight = proj_weight.view(-1).expand(self.relation_total, -1)
else:
nn.init.xavier_uniform(proj_weight)
# init user and item embeddings
self.ent_embeddings = nn.Embedding(self.ent_total, self.embedding_size)
self.rel_embeddings = nn.Embedding(self.rel_total, self.embedding_size)
self.proj_embeddings = nn.Embedding(
self.rel_total, self.embedding_size * self.embedding_size)
self.ent_embeddings.weight = nn.Parameter(ent_weight)
self.rel_embeddings.weight = nn.Parameter(rel_weight)
self.proj_embeddings.weight = nn.Parameter(proj_weight)
normalize_ent_emb = F.normalize(self.ent_embeddings.weight.data,
p=2,
dim=1)
normalize_rel_emb = F.normalize(self.rel_embeddings.weight.data,
p=2,
dim=1)
# normalize_proj_emb = F.normalize(self.proj_embeddings.weight.data, p=2, dim=1)
self.ent_embeddings.weight.data = normalize_ent_emb
self.rel_embeddings.weight.data = normalize_rel_emb
# self.proj_embeddings.weight.data = normalize_proj_emb
self.ent_embeddings = to_gpu(self.ent_embeddings)
self.rel_embeddings = to_gpu(self.rel_embeddings)
self.proj_embeddings = to_gpu(self.proj_embeddings)
def reportPreference(self, u_id, i_ids):
item_num = len(i_ids)
# item * dim
u_e = self.user_embeddings(u_id.expand(item_num))
i_e = self.item_embeddings(i_ids)
e_ids = self.paddingItems(i_ids.data, self.ent_total - 1)
e_var = to_gpu(V(torch.LongTensor(e_ids)))
e_e = self.ent_embeddings(e_var)
ie_e = i_e + e_e
return self.getPreferences(u_e, ie_e, use_st_gumbel=self.use_st_gumbel)
def evaluateRec(self, u_ids, all_i_ids=None):
batch_size = len(u_ids)
i_ids = range(len(self.item_embeddings.weight))
e_ids = self.paddingItems(i_ids, self.ent_total - 1)
e_var = to_gpu(V(torch.LongTensor(e_ids)))
e_e = self.ent_embeddings(e_var)
all_ie_e = self.item_embeddings.weight + e_e
u_e = self.user_embeddings(u_ids)
return torch.matmul(u_e, all_ie_e.t())
def forward(self, ratings, triples, is_rec=True):
if is_rec and ratings is not None:
u_ids, i_ids = ratings
e_ids = self.paddingItems(i_ids.data, self.ent_total - 1)
e_var = to_gpu(V(torch.LongTensor(e_ids)))
u_e = self.user_embeddings(u_ids)
i_e = self.item_embeddings(i_ids)
e_e = self.ent_embeddings(e_var)
ie_e = i_e + e_e
_, r_e, norm = self.getPreferences(
u_e, ie_e, use_st_gumbel=self.use_st_gumbel)
proj_u_e = projection_transH_pytorch(u_e, norm)
proj_i_e = projection_transH_pytorch(ie_e, norm)
if self.L1_flag:
score = torch.sum(torch.abs(proj_u_e + r_e - proj_i_e), 1)
else:
score = torch.sum((proj_u_e + r_e - proj_i_e)**2, 1)
elif not is_rec and triples is not None:
h, t, r = triples
h_e = self.ent_embeddings(h)
t_e = self.ent_embeddings(t)
r_e = self.rel_embeddings(r)
norm_e = self.norm_embeddings(r)
proj_h_e = projection_transH_pytorch(h_e, norm_e)
proj_t_e = projection_transH_pytorch(t_e, norm_e)
if self.L1_flag:
score = torch.sum(torch.abs(proj_h_e + r_e - proj_t_e), 1)
else:
score = torch.sum((proj_h_e + r_e - proj_t_e)**2, 1)
else:
raise NotImplementedError
return score
def train_loop(FLAGS,
model,
trainer,
rating_train_dataset,
triple_train_dataset,
rating_eval_datasets,
triple_eval_datasets,
e_map,
i_map,
ikg_map,
logger,
vis=None,
is_report=False):
rating_train_iter, rating_train_total, rating_train_list, rating_train_dict = rating_train_dataset
triple_train_iter, triple_train_total, triple_train_list, head_train_dict, tail_train_dict = triple_train_dataset
all_rating_dicts = None
if FLAGS.filter_wrong_corrupted:
all_rating_dicts = [rating_train_dict] + [
tmp_data[3] for tmp_data in rating_eval_datasets
]
all_head_dicts = None
all_tail_dicts = None
if FLAGS.filter_wrong_corrupted:
all_head_dicts = [head_train_dict] + [
tmp_data[4] for tmp_data in triple_eval_datasets
]
all_tail_dicts = [tail_train_dict] + [
tmp_data[5] for tmp_data in triple_eval_datasets
]
item_total = len(i_map)
entity_total = len(e_map)
step_to_switch = 10 * FLAGS.joint_ratio
# Train.
logger.info("Training.")
# New Training Loop
pbar = None
rec_total_loss = 0.0
kg_total_loss = 0.0
model.train()
model.enable_grad()
for _ in range(trainer.step, FLAGS.training_steps):
if FLAGS.early_stopping_steps_to_wait > 0 and (
trainer.step -
trainer.best_step) > FLAGS.early_stopping_steps_to_wait:
logger.info('No improvement after ' +
str(FLAGS.early_stopping_steps_to_wait) +
' steps. Stopping training.')
if pbar is not None: pbar.close()
break
if trainer.step % FLAGS.eval_interval_steps == 0:
if pbar is not None:
pbar.close()
rec_total_loss /= (FLAGS.eval_interval_steps * FLAGS.joint_ratio)
kg_total_loss /= (FLAGS.eval_interval_steps *
(1 - FLAGS.joint_ratio))
logger.info("rec train loss:{:.4f}, kg train loss:{:.4f}!".format(
rec_total_loss, kg_total_loss))
rec_performances = []
for i, eval_data in enumerate(rating_eval_datasets):
all_eval_dicts = None
if FLAGS.filter_wrong_corrupted:
all_eval_dicts = [rating_train_dict] + [
tmp_data[3]
for j, tmp_data in enumerate(rating_eval_datasets)
if j != i
]
rec_performances.append(
evaluateRec(FLAGS,
model,
eval_data[0],
eval_data[3],
all_eval_dicts,
i_map,
logger,
eval_descending=True
if trainer.model_target == 1 else False,
is_report=is_report))
kg_performances = []
for i, eval_data in enumerate(triple_eval_datasets):
eval_head_dicts = None
eval_tail_dicts = None
if FLAGS.filter_wrong_corrupted:
eval_head_dicts = [head_train_dict] + [
tmp_data[4]
for j, tmp_data in enumerate(triple_eval_datasets)
if j != i
]
eval_tail_dicts = [tail_train_dict] + [
tmp_data[5]
for j, tmp_data in enumerate(triple_eval_datasets)
if j != i
]
kg_performances.append(
evaluateKG(FLAGS,
model,
eval_data[0],
eval_data[1],
eval_data[4],
eval_data[5],
eval_head_dicts,
eval_tail_dicts,
e_map,
logger,
eval_descending=False,
is_report=is_report))
if trainer.step > 0:
is_best = trainer.new_performance(kg_performances[0],
kg_performances)
# visuliazation
if vis is not None:
vis.plot_many_stack(
{
'Rec Train Loss': rec_total_loss,
'KG Train Loss': kg_total_loss
},
win_name="Loss Curve")
f1_dict = {}
p_dict = {}
r_dict = {}
rec_hit_dict = {}
ndcg_dict = {}
for i, performance in enumerate(rec_performances):
f1_dict['Rec Eval {} F1'.format(i)] = performance[0]
p_dict['Rec Eval {} Precision'.format(
i)] = performance[1]
r_dict['Rec Eval {} Recall'.format(i)] = performance[2]
rec_hit_dict['Rec Eval {} Hit'.format(
i)] = performance[3]
ndcg_dict['Rec Eval {} NDCG'.format(
i)] = performance[4]
kg_hit_dict = {}
meanrank_dict = {}
for i, performance in enumerate(kg_performances):
kg_hit_dict['KG Eval {} Hit'.format(
i)] = performance[0]
meanrank_dict['KG Eval {} MeanRank'.format(
i)] = performance[1]
if is_best:
log_str = [
"Best performances in {} step!".format(
trainer.best_step)
]
log_str += [
"{} : {}.".format(s, "%.5f" % f1_dict[s])
for s in f1_dict
]
log_str += [
"{} : {}.".format(s, "%.5f" % p_dict[s])
for s in p_dict
]
log_str += [
"{} : {}.".format(s, "%.5f" % r_dict[s])
for s in r_dict
]
log_str += [
"{} : {}.".format(s, "%.5f" % rec_hit_dict[s])
for s in rec_hit_dict
]
log_str += [
"{} : {}.".format(s, "%.5f" % ndcg_dict[s])
for s in ndcg_dict
]
log_str += [
"{} : {}.".format(s, "%.5f" % kg_hit_dict[s])
for s in kg_hit_dict
]
log_str += [
"{} : {}.".format(s, "%.5f" % meanrank_dict[s])
for s in meanrank_dict
]
vis.log("\n".join(log_str),
win_name="Best Performances")
vis.plot_many_stack(f1_dict,
win_name="Rec F1 Score@{}".format(
FLAGS.topn))
vis.plot_many_stack(p_dict,
win_name="Rec Precision@{}".format(
FLAGS.topn))
vis.plot_many_stack(r_dict,
win_name="Rec Recall@{}".format(
FLAGS.topn))
vis.plot_many_stack(rec_hit_dict,
win_name="Rec Hit Ratio@{}".format(
FLAGS.topn))
vis.plot_many_stack(ndcg_dict,
win_name="Rec NDCG@{}".format(
FLAGS.topn))
vis.plot_many_stack(kg_hit_dict,
win_name="KG Hit Ratio@{}".format(
FLAGS.topn))
vis.plot_many_stack(meanrank_dict, win_name="KG MeanRank")
# set model in training mode
pbar = tqdm(total=FLAGS.eval_interval_steps)
pbar.set_description("Training")
rec_total_loss = 0.0
kg_total_loss = 0.0
model.train()
model.enable_grad()
# recommendation train
if trainer.step % 10 < step_to_switch:
rating_batch = next(rating_train_iter)
u, pi, ni = getNegRatings(rating_batch,
item_total,
all_dicts=all_rating_dicts)
e_ids, i_ids = getMappedEntities(pi + ni, i_map, ikg_map)
if FLAGS.share_embeddings:
ni = [i_map[i] for i in ni]
pi = [i_map[i] for i in pi]
u_var = to_gpu(V(torch.LongTensor(u)))
pi_var = to_gpu(V(torch.LongTensor(pi)))
ni_var = to_gpu(V(torch.LongTensor(ni)))
trainer.optimizer_zero_grad()
# Run model. output: batch_size * cand_num, input: ratings, triples, is_rec=True
pos_score = model((u_var, pi_var), None, is_rec=True)
neg_score = model((u_var, ni_var), None, is_rec=True)
# Calculate loss.
losses = bprLoss(pos_score, neg_score, target=trainer.model_target)
if FLAGS.model_type in ["transup", "jtransup"]:
losses += orthogonalLoss(model.pref_embeddings.weight,
model.pref_norm_embeddings.weight)
# kg train
else:
triple_batch = next(triple_train_iter)
ph, pt, pr, nh, nt, nr = getTrainTripleBatch(
triple_batch,
entity_total,
all_head_dicts=all_head_dicts,
all_tail_dicts=all_tail_dicts)
e_ids, i_ids = getMappedItems(ph + pt + nh + nt, e_map, ikg_map)
if FLAGS.share_embeddings:
ph = [e_map[e] for e in ph]
pt = [e_map[e] for e in pt]
nh = [e_map[e] for e in nh]
nt = [e_map[e] for e in nt]
ph_var = to_gpu(V(torch.LongTensor(ph)))
pt_var = to_gpu(V(torch.LongTensor(pt)))
pr_var = to_gpu(V(torch.LongTensor(pr)))
nh_var = to_gpu(V(torch.LongTensor(nh)))
nt_var = to_gpu(V(torch.LongTensor(nt)))
nr_var = to_gpu(V(torch.LongTensor(nr)))
trainer.optimizer_zero_grad()
# Run model. output: batch_size * cand_nu, input: ratings, triples, is_rec=True
pos_score = model(None, (ph_var, pt_var, pr_var), is_rec=False)
neg_score = model(None, (nh_var, nt_var, nr_var), is_rec=False)
# Calculate loss.
# losses = nn.MarginRankingLoss(margin=FLAGS.margin).forward(pos_score, neg_score, to_gpu(torch.autograd.Variable(torch.FloatTensor([trainer.model_target]*len(ph)))))
losses = loss.marginLoss()(pos_score, neg_score, FLAGS.margin)
ent_embeddings = model.ent_embeddings(
torch.cat([ph_var, pt_var, nh_var, nt_var]))
rel_embeddings = model.rel_embeddings(torch.cat([pr_var, nr_var]))
if FLAGS.model_type in ["jtransup"]:
norm_embeddings = model.norm_embeddings(
torch.cat([pr_var, nr_var]))
losses += loss.orthogonalLoss(rel_embeddings, norm_embeddings)
losses = losses + loss.normLoss(ent_embeddings) + loss.normLoss(
rel_embeddings)
losses = FLAGS.kg_lambda * losses
# align loss if not share embeddings
if not FLAGS.share_embeddings and FLAGS.model_type not in [
'cke', 'jtransup'
]:
e_var = to_gpu(V(torch.LongTensor(e_ids)))
i_var = to_gpu(V(torch.LongTensor(i_ids)))
ent_embeddings = model.ent_embeddings(e_var)
item_embeddings = model.item_embeddings(i_var)
losses += FLAGS.norm_lambda * loss.pNormLoss(
ent_embeddings, item_embeddings, L1_flag=FLAGS.L1_flag)
# Backward pass.
losses.backward()
# for param in model.parameters():
# print(param.grad.data.sum())
# Hard Gradient Clipping
nn.utils.clip_grad_norm(
[param for name, param in model.named_parameters()],
FLAGS.clipping_max_value)
# Gradient descent step.
trainer.optimizer_step()
if trainer.step % 10 < step_to_switch:
rec_total_loss += losses.data[0]
else:
kg_total_loss += losses.data[0]
pbar.update(1)
def evaluateKG(FLAGS,
model,
eval_head_iter,
eval_tail_iter,
eval_head_dict,
eval_tail_dict,
all_head_dicts,
all_tail_dicts,
e_map,
logger,
eval_descending=True,
is_report=False):
# Evaluate
total_batches = len(eval_head_iter) + len(eval_tail_iter)
# processing bar
pbar = tqdm(total=total_batches)
pbar.set_description("Run Eval")
model.eval()
model.disable_grad()
all_e_var = None
if FLAGS.share_embeddings:
all_e_ids = [e_map[e] for e in range(len(e_map))]
all_e_var = to_gpu(V(torch.LongTensor(all_e_ids)))
# head prediction evaluation
head_results = []
for batch_trs in eval_head_iter:
t = [tr[0] for tr in batch_trs] if FLAGS.share_embeddings else [
e_map[tr[0]] for tr in batch_trs
]
r = [tr[1] for tr in batch_trs]
t_var = to_gpu(V(torch.LongTensor(t)))
r_var = to_gpu(V(torch.LongTensor(r)))
# batch * item
scores = model.evaluateHead(t_var, r_var, all_e_ids=all_e_var)
preds = zip(batch_trs, scores.data.cpu().numpy())
head_results.extend(
evalKGProcess(list(preds),
eval_head_dict,
all_dicts=all_head_dicts,
descending=eval_descending,
num_processes=FLAGS.num_processes,
topn=FLAGS.topn,
queue_limit=FLAGS.max_queue))
pbar.update(1)
# tail prediction evaluation
tail_results = []
for batch_hrs in eval_tail_iter:
h = [hr[0] for hr in batch_hrs] if FLAGS.share_embeddings else [
e_map[hr[0]] for hr in batch_hrs
]
r = [hr[1] for hr in batch_hrs]
h_var = to_gpu(V(torch.LongTensor(h)))
r_var = to_gpu(V(torch.LongTensor(r)))
# batch * item
scores = model.evaluateTail(h_var, r_var, all_e_ids=all_e_var)
preds = zip(batch_hrs, scores.data.cpu().numpy())
tail_results.extend(
evalKGProcess(list(preds),
eval_tail_dict,
all_dicts=all_tail_dicts,
descending=eval_descending,
num_processes=FLAGS.num_processes,
topn=FLAGS.topn,
queue_limit=FLAGS.max_queue))
pbar.update(1)
pbar.close()
# hit, rank
head_performances = [result[:2] for result in head_results]
tail_performances = [result[:2] for result in tail_results]
head_hit, head_mean_rank = np.array(head_performances).mean(axis=0)
tail_hit, tail_mean_rank = np.array(tail_performances).mean(axis=0)
logger.info("head hit:{:.4f}, head mean rank:{:.4f}, topn:{}.".format(
head_hit, head_mean_rank, FLAGS.topn))
logger.info("tail hit:{:.4f}, tail mean rank:{:.4f}, topn:{}.".format(
tail_hit, tail_mean_rank, FLAGS.topn))
head_num = len(head_results)
tail_num = len(tail_results)
avg_hit = float(head_hit * head_num + tail_hit * tail_num) / (head_num +
tail_num)
avg_mean_rank = float(head_mean_rank * head_num +
tail_mean_rank * tail_num) / (head_num + tail_num)
logger.info("avg hit:{:.4f}, avg mean rank:{:.4f}, topn:{}.".format(
avg_hit, avg_mean_rank, FLAGS.topn))
if is_report:
for result in head_results:
hit = result[0]
rank = result[1]
t = result[2][0]
r = result[2][1]
gold_h = result[3]
logger.info("H\t{}\t{}\t{}\t{}".format(gold_h, t, r, hit))
for result in tail_results:
hit = result[0]
rank = result[1]
h = result[2][0]
r = result[2][1]
gold_t = result[3]
logger.info("T\t{}\t{}\t{}\t{}".format(h, gold_t, r, hit))
model.enable_grad()
return avg_hit, avg_mean_rank
def __init__(self, L1_flag, embedding_size, user_total, item_total,
entity_total, relation_total, isShare):
super(coFM, self).__init__()
self.L1_flag = L1_flag
self.is_share = isShare
self.embedding_size = embedding_size
self.user_total = user_total
self.item_total = item_total
self.ent_total = entity_total
self.rel_total = relation_total
self.is_pretrained = False
# fm
user_weight = torch.FloatTensor(self.user_total, self.embedding_size)
nn.init.xavier_uniform(user_weight)
self.user_embeddings = nn.Embedding(self.user_total,
self.embedding_size)
self.user_embeddings.weight = nn.Parameter(user_weight)
normalize_user_emb = F.normalize(self.user_embeddings.weight.data,
p=2,
dim=1)
self.user_embeddings.weight.data = normalize_user_emb
self.user_embeddings = to_gpu(self.user_embeddings)
user_bias = torch.FloatTensor(self.user_total)
item_bias = torch.FloatTensor(self.item_total)
nn.init.constant(user_bias, 0)
nn.init.constant(item_bias, 0)
self.user_bias = nn.Embedding(self.user_total, 1)
self.item_bias = nn.Embedding(self.item_total, 1)
self.user_bias.weight = nn.Parameter(user_bias, 1)
self.item_bias.weight = nn.Parameter(item_bias, 1)
self.user_bias = to_gpu(self.user_bias)
self.item_bias = to_gpu(self.item_bias)
self.bias = nn.Parameter(to_gpu(torch.FloatTensor([0.0])))
# trane
rel_weight = torch.FloatTensor(self.rel_total, self.embedding_size)
nn.init.xavier_uniform(rel_weight)
self.rel_embeddings = nn.Embedding(self.rel_total, self.embedding_size)
self.rel_embeddings.weight = nn.Parameter(rel_weight)
normalize_rel_emb = F.normalize(self.rel_embeddings.weight.data,
p=2,
dim=1)
self.rel_embeddings.weight.data = normalize_rel_emb
self.rel_embeddings = to_gpu(self.rel_embeddings)
# shared embedding
ent_weight = torch.FloatTensor(self.ent_total, self.embedding_size)
nn.init.xavier_uniform(ent_weight)
self.ent_embeddings = nn.Embedding(self.ent_total, self.embedding_size)
self.ent_embeddings.weight = nn.Parameter(ent_weight)
normalize_ent_emb = F.normalize(self.ent_embeddings.weight.data,
p=2,
dim=1)
self.ent_embeddings.weight.data = normalize_ent_emb
self.ent_embeddings = to_gpu(self.ent_embeddings)
if self.is_share:
assert self.item_total == self.ent_total, "item numbers didn't match entities!"
self.item_embeddings = self.ent_embeddings
else:
item_weight = torch.FloatTensor(self.item_total,
self.embedding_size)
nn.init.xavier_uniform(item_weight)
self.item_embeddings = nn.Embedding(self.item_total,
self.embedding_size)
self.item_embeddings.weight = nn.Parameter(item_weight)
normalize_item_emb = F.normalize(self.item_embeddings.weight.data,
p=2,
dim=1)
self.item_embeddings.weight.data = normalize_item_emb
self.item_embeddings = to_gpu(self.item_embeddings)
def case_rec_evaluateRec(FLAGS,
model,
eval_iter,
eval_dict,
all_dicts,
i_map,
logger,
i,
eval_descending=True,
is_report=False):
# Evaluate
total_batches = len(eval_iter)
# processing bar
pbar = tqdm(total=total_batches)
pbar.set_description("Run Eval")
all_i_var = None
if FLAGS.share_embeddings:
all_i_ids = [i_map[i] for i in range(len(i_map))]
all_i_var = to_gpu(V(torch.LongTensor(all_i_ids)))
model.eval()
model.disable_grad()
results = []
for u_ids in eval_iter:
u_var = to_gpu(V(torch.LongTensor(u_ids)))
# batch * item
scores = model.evaluateRec(u_var, all_i_ids=all_i_var)
preds = zip(u_ids, scores.data.cpu().numpy())
results.extend(
evalRecProcess(list(preds),
eval_dict,
all_dicts=all_dicts,
descending=eval_descending,
num_processes=FLAGS.num_processes,
topn=FLAGS.topn,
queue_limit=FLAGS.max_queue))
pbar.update(1)
pbar.close()
predictions = [result[5] for result in results
] # [(pred[0], top_ids, gold), ...], gold is test
print("Saving predictions. Size: {}.".format(str(len(predictions))))
predictions_output_filepath = os.path.join(
FLAGS.log_path, FLAGS.experiment_name + '_pred.dat')
print_list = []
for triple in predictions:
u_id = triple[0]
top_ids = triple[1]
#gold = triple[2]
for i_id in top_ids:
score = 1.0 / (top_ids.index(i_id) + 1)
print_list.append((u_id, i_id, score))
WriteFile(predictions_output_filepath, data=print_list, sep='\t').write()
# Using CaseRecommender ReadFile class to read test_set from file
dataset_path = os.path.join(FLAGS.data_path, FLAGS.dataset)
eval_files = FLAGS.rec_test_files.split(':')
test_path = os.path.join(dataset_path, eval_files[i])
eval_data = ReadFile(input_file=test_path).read()
predictions_data = ReadFile(input_file=predictions_output_filepath).read()
print("Reading predictions. Size: {}.".format(
str(len(predictions_data['feedback']))))
# Creating CaseRecommender evaluator with item-recommendation parameters
evaluator = ItemRecommendationEvaluation(n_ranks=[10])
item_rec_metrics = evaluator.evaluate(predictions_data['feedback'],
eval_data)
print("From CaseRecommender evaluator: {}.".format(str(item_rec_metrics)))
logger.info("From CaseRecommender evaluator: {}.".format(
str(item_rec_metrics)))
# Creating kg-summ-rec evaluator with diversity parameters
dataset_name = os.path.basename(
os.path.dirname(os.path.dirname(FLAGS.log_path)))
tags = dataset_name.split('_')
if tags[0] == 'ml-sun':
evaluator2 = DiversityEvaluation(n_ranks=[10])
dataset_path = os.path.normpath(FLAGS.data_path + os.sep + os.pardir)
#tags = dataset_name.split('-')
#if len(tags) > 2:
# mode = dataset_name.split('-')[2]
# ratio = dataset_name.split('-')[4]
#else:
# mode = 'sv'
# ratio = '100'
dataset_path = os.path.normpath(FLAGS.data_path + os.sep + os.pardir +
os.sep + os.pardir + os.sep + tags[0] +
'_' + tags[1] + '_' + 'oKG')
mode = 'sv'
ratio = '100'
i2genre_map = read_i2genre_map(dataset_path, mode, ratio)
diversity_metrics = evaluator2.evaluate(predictions_data['feedback'],
eval_data, i2genre_map)
print("From kg-summ-rec diversity evaluator: {}.".format(
str(diversity_metrics)))
logger.info("From kg-summ-rec diversity evaluator: {}.".format(
str(diversity_metrics)))
model.enable_grad()
return item_rec_metrics
def train_loop(FLAGS,
model,
trainer,
train_dataset,
eval_datasets,
entity_total,
relation_total,
logger,
vis=None,
is_report=False):
train_iter, train_total, train_list, train_head_dict, train_tail_dict = train_dataset
all_head_dicts = None
all_tail_dicts = None
if FLAGS.filter_wrong_corrupted:
all_head_dicts = [train_head_dict
] + [tmp_data[4] for tmp_data in eval_datasets]
all_tail_dicts = [train_tail_dict
] + [tmp_data[5] for tmp_data in eval_datasets]
# Train.
logger.info("Training.")
# New Training Loop
pbar = None
total_loss = 0.0
model.enable_grad()
for _ in range(trainer.step, FLAGS.training_steps):
if FLAGS.early_stopping_steps_to_wait > 0 and (
trainer.step -
trainer.best_step) > FLAGS.early_stopping_steps_to_wait:
logger.info('No improvement after ' +
str(FLAGS.early_stopping_steps_to_wait) +
' steps. Stopping training.')
if pbar is not None: pbar.close()
break
if trainer.step % FLAGS.eval_interval_steps == 0:
if pbar is not None:
pbar.close()
total_loss /= FLAGS.eval_interval_steps
logger.info("train loss:{:.4f}!".format(total_loss))
performances = []
for i, eval_data in enumerate(eval_datasets):
eval_head_dicts = None
eval_tail_dicts = None
if FLAGS.filter_wrong_corrupted:
eval_head_dicts = [train_head_dict] + [
tmp_data[4]
for j, tmp_data in enumerate(eval_datasets) if j != i
]
eval_tail_dicts = [train_tail_dict] + [
tmp_data[5]
for j, tmp_data in enumerate(eval_datasets) if j != i
]
performances.append(
evaluate(FLAGS,
model,
entity_total,
relation_total,
eval_data[0],
eval_data[1],
eval_data[4],
eval_data[5],
eval_head_dicts,
eval_tail_dicts,
logger,
eval_descending=False,
is_report=is_report))
if trainer.step > 0:
is_best = trainer.new_performance(performances[0],
performances)
# visuliazation
if vis is not None:
vis.plot_many_stack({'KG Train Loss': total_loss},
win_name="Loss Curve")
hit_vis_dict = {}
meanrank_vis_dict = {}
for i, performance in enumerate(performances):
hit_vis_dict['KG Eval {} Hit'.format(
i)] = performance[0]
meanrank_vis_dict['KG Eval {} MeanRank'.format(
i)] = performance[1]
if is_best:
log_str = [
"Best performances in {} step!".format(
trainer.best_step)
]
log_str += [
"{} : {}.".format(s, "%.5f" % hit_vis_dict[s])
for s in hit_vis_dict
]
log_str += [
"{} : {}.".format(s, "%.5f" % meanrank_vis_dict[s])
for s in meanrank_vis_dict
]
vis.log("\n".join(log_str),
win_name="Best Performances")
vis.plot_many_stack(hit_vis_dict,
win_name="KG Hit Ratio@{}".format(
FLAGS.topn))
vis.plot_many_stack(meanrank_vis_dict,
win_name="KG MeanRank")
# set model in training mode
pbar = tqdm(total=FLAGS.eval_interval_steps)
pbar.set_description("Training")
total_loss = 0.0
model.train()
model.enable_grad()
triple_batch = next(train_iter)
ph, pt, pr, nh, nt, nr = getTrainTripleBatch(
triple_batch,
entity_total,
all_head_dicts=all_head_dicts,
all_tail_dicts=all_tail_dicts)
ph_var = to_gpu(V(torch.LongTensor(ph)))
pt_var = to_gpu(V(torch.LongTensor(pt)))
pr_var = to_gpu(V(torch.LongTensor(pr)))
nh_var = to_gpu(V(torch.LongTensor(nh)))
nt_var = to_gpu(V(torch.LongTensor(nt)))
nr_var = to_gpu(V(torch.LongTensor(nr)))
trainer.optimizer_zero_grad()
# Run model. output: batch_size * 1
pos_score = model(ph_var, pt_var, pr_var)
neg_score = model(nh_var, nt_var, nr_var)
# Calculate loss.
# losses = nn.MarginRankingLoss(margin=FLAGS.margin).forward(pos_score, neg_score, to_gpu(torch.autograd.Variable(torch.FloatTensor([trainer.model_target]*len(ph)))))
losses = loss.marginLoss()(pos_score, neg_score, FLAGS.margin)
ent_embeddings = model.ent_embeddings(
torch.cat([ph_var, pt_var, nh_var, nt_var]))
rel_embeddings = model.rel_embeddings(torch.cat([pr_var, nr_var]))
if FLAGS.model_type == "transh":
norm_embeddings = model.norm_embeddings(torch.cat([pr_var,
nr_var]))
losses += loss.orthogonalLoss(rel_embeddings, norm_embeddings)
losses = losses + loss.normLoss(ent_embeddings) + loss.normLoss(
rel_embeddings)
# Backward pass.
losses.backward()
# for param in model.parameters():
# print(param.grad.data.sum())
# Hard Gradient Clipping
nn.utils.clip_grad_norm(
[param for name, param in model.named_parameters()],
FLAGS.clipping_max_value)
# Gradient descent step.
trainer.optimizer_step()
total_loss += losses.data[0]
pbar.update(1)
def normLoss(embeddings, dim=1):
norm = torch.sum(embeddings ** 2, dim=dim, keepdim=True)
return torch.sum(torch.max(norm - to_gpu(autograd.Variable(torch.FloatTensor([1.0]))), to_gpu(autograd.Variable(torch.FloatTensor([0.0])))))
def forward(self, pos, neg, margin):
zero_tensor = to_gpu(torch.FloatTensor(pos.size()))
zero_tensor.zero_()
zero_tensor = autograd.Variable(zero_tensor)
return torch.sum(torch.max(pos - neg + margin, zero_tensor))
def train_loop(FLAGS,
model,
trainer,
train_dataset,
eval_datasets,
user_total,
item_total,
logger,
vis=None,
is_report=False):
train_iter, train_total, train_list, train_dict = train_dataset
all_dicts = None
if FLAGS.filter_wrong_corrupted:
all_dicts = [train_dict] + [tmp_data[3] for tmp_data in eval_datasets]
# Train.
logger.info("Training.")
# New Training Loop
pbar = None
total_loss = 0.0
model.train()
model.enable_grad()
for _ in range(trainer.step, FLAGS.training_steps):
# if FLAGS.early_stopping_steps_to_wait > 0 and (trainer.step - trainer.best_step) > FLAGS.early_stopping_steps_to_wait:
# logger.info('No improvement after ' +
# str(FLAGS.early_stopping_steps_to_wait) +
# ' steps. Stopping training.')
# if pbar is not None: pbar.close()
# break
if trainer.step % FLAGS.eval_interval_steps == 0:
if pbar is not None:
pbar.close()
total_loss /= FLAGS.eval_interval_steps
logger.info("train loss:{:.4f}!".format(total_loss))
# performances = []
# for i, eval_data in enumerate(eval_datasets):
# all_eval_dicts = None
# if FLAGS.filter_wrong_corrupted:
# all_eval_dicts = [train_dict] + [tmp_data[3] for j, tmp_data in enumerate(eval_datasets) if j!=i]
# performances.append( evaluate(FLAGS, model, eval_data[0], eval_data[3], all_eval_dicts, logger, eval_descending=True if trainer.model_target == 1 else False, is_report=is_report))
# if trainer.step > 0 and len(performances) > 0:
# is_best = trainer.new_performance(performances[0], performances)
# # visuliazation
# if vis is not None:
# vis.plot_many_stack({'Rec Train Loss': total_loss},
# win_name="Loss Curve")
# f1_vis_dict = {}
# p_vis_dict = {}
# r_vis_dict = {}
# hit_vis_dict = {}
# ndcg_vis_dict = {}
# for i, performance in enumerate(performances):
# f1_vis_dict['Rec Eval {} F1'.format(i)] = performance[0]
# p_vis_dict['Rec Eval {} Precision'.format(i)] = performance[1]
# r_vis_dict['Rec Eval {} Recall'.format(i)] = performance[2]
# hit_vis_dict['Rec Eval {} Hit'.format(i)] = performance[3]
# ndcg_vis_dict['Rec Eval {} NDCG'.format(i)] = performance[4]
# if is_best:
# log_str = ["Best performances in {} step!".format(trainer.best_step)]
# log_str += ["{} : {}.".format(s, "%.5f" % f1_vis_dict[s]) for s in f1_vis_dict]
# log_str += ["{} : {}.".format(s, "%.5f" % p_vis_dict[s]) for s in p_vis_dict]
# log_str += ["{} : {}.".format(s, "%.5f" % r_vis_dict[s]) for s in r_vis_dict]
# log_str += ["{} : {}.".format(s, "%.5f" % hit_vis_dict[s]) for s in hit_vis_dict]
# log_str += ["{} : {}.".format(s, "%.5f" % ndcg_vis_dict[s]) for s in ndcg_vis_dict]
# vis.log("\n".join(log_str), win_name="Best Performances")
# vis.plot_many_stack(f1_vis_dict, win_name="Rec F1 Score@{}".format(FLAGS.topn))
# vis.plot_many_stack(p_vis_dict, win_name="Rec Precision@{}".format(FLAGS.topn))
# vis.plot_many_stack(r_vis_dict, win_name="Rec Recall@{}".format(FLAGS.topn))
# vis.plot_many_stack(hit_vis_dict, win_name="Rec Hit Ratio@{}".format(FLAGS.topn))
# vis.plot_many_stack(ndcg_vis_dict, win_name="Rec NDCG@{}".format(FLAGS.topn))
# set model in training mode
pbar = tqdm(total=FLAGS.eval_interval_steps)
pbar.set_description("Training")
total_loss = 0.0
model.train()
model.enable_grad()
rating_batch = next(train_iter)
u, pi, ni = getNegRatings(rating_batch,
item_total,
all_dicts=all_dicts)
u_var = to_gpu(V(torch.LongTensor(u)))
pi_var = to_gpu(V(torch.LongTensor(pi)))
ni_var = to_gpu(V(torch.LongTensor(ni)))
trainer.optimizer_zero_grad()
# Run model. output: batch_size * cand_num
pos_score = model(u_var, pi_var)
neg_score = model(u_var, ni_var)
# Calculate loss.
losses = bprLoss(pos_score, neg_score, target=trainer.model_target)
if FLAGS.model_type in ["transup", "transupb"]:
user_embeddings = model.user_embeddings(u_var)
item_embeddings = model.item_embeddings(torch.cat([pi_var,
ni_var]))
losses += orthogonalLoss(
model.pref_embeddings.weight,
model.pref_norm_embeddings.weight) + normLoss(
user_embeddings) + normLoss(item_embeddings) + normLoss(
model.pref_embeddings.weight)
# Backward pass.
losses.backward()
# for param in model.parameters():
# print(param.grad.data.sum())
# Hard Gradient Clipping
nn.utils.clip_grad_norm(
[param for name, param in model.named_parameters()],
FLAGS.clipping_max_value)
# Gradient descent step.
trainer.optimizer_step()
total_loss += losses.item()
pbar.update(1)
def __init__(self, L1_flag, embedding_size, user_total, item_total,
entity_total, relation_total, i_map, new_map, isShare,
use_st_gumbel):
super(jTransUPModel, self).__init__()
self.L1_flag = L1_flag
self.is_share = isShare
self.use_st_gumbel = use_st_gumbel
self.embedding_size = embedding_size
self.user_total = user_total
self.item_total = item_total
# padding when item are not aligned with any entity
self.ent_total = entity_total + 1
self.rel_total = relation_total
self.is_pretrained = False
# store item to item-entity dic
self.i_map = i_map
# store item-entity to (entity, item)
self.new_map = new_map
# todo: simiplifying the init
# transup
user_weight = torch.FloatTensor(self.user_total, self.embedding_size)
item_weight = torch.FloatTensor(self.item_total, self.embedding_size)
pref_weight = torch.FloatTensor(self.rel_total, self.embedding_size)
pref_norm_weight = torch.FloatTensor(self.rel_total,
self.embedding_size)
nn.init.xavier_uniform(user_weight)
nn.init.xavier_uniform(item_weight)
nn.init.xavier_uniform(pref_weight)
nn.init.xavier_uniform(pref_norm_weight)
# init user and item embeddings
self.user_embeddings = nn.Embedding(self.user_total,
self.embedding_size)
self.item_embeddings = nn.Embedding(self.item_total,
self.embedding_size)
self.user_embeddings.weight = nn.Parameter(user_weight)
self.item_embeddings.weight = nn.Parameter(item_weight)
normalize_user_emb = F.normalize(self.user_embeddings.weight.data,
p=2,
dim=1)
normalize_item_emb = F.normalize(self.item_embeddings.weight.data,
p=2,
dim=1)
self.user_embeddings.weight.data = normalize_user_emb
self.item_embeddings.weight.data = normalize_item_emb
# init preference parameters
self.pref_embeddings = nn.Embedding(self.rel_total,
self.embedding_size)
self.pref_norm_embeddings = nn.Embedding(self.rel_total,
self.embedding_size)
self.pref_embeddings.weight = nn.Parameter(pref_weight)
self.pref_norm_embeddings.weight = nn.Parameter(pref_norm_weight)
normalize_pref_emb = F.normalize(self.pref_embeddings.weight.data,
p=2,
dim=1)
normalize_pref_norm_emb = F.normalize(
self.pref_norm_embeddings.weight.data, p=2, dim=1)
self.pref_embeddings.weight.data = normalize_pref_emb
self.pref_norm_embeddings.weight.data = normalize_pref_norm_emb
self.user_embeddings = to_gpu(self.user_embeddings)
self.item_embeddings = to_gpu(self.item_embeddings)
self.pref_embeddings = to_gpu(self.pref_embeddings)
self.pref_norm_embeddings = to_gpu(self.pref_norm_embeddings)
# transh
ent_weight = torch.FloatTensor(self.ent_total - 1, self.embedding_size)
rel_weight = torch.FloatTensor(self.rel_total, self.embedding_size)
norm_weight = torch.FloatTensor(self.rel_total, self.embedding_size)
nn.init.xavier_uniform(ent_weight)
nn.init.xavier_uniform(rel_weight)
nn.init.xavier_uniform(norm_weight)
norm_ent_weight = F.normalize(ent_weight, p=2, dim=1)
# init user and item embeddings
self.ent_embeddings = nn.Embedding(self.ent_total,
self.embedding_size,
padding_idx=self.ent_total - 1)
self.rel_embeddings = nn.Embedding(self.rel_total, self.embedding_size)
self.norm_embeddings = nn.Embedding(self.rel_total,
self.embedding_size)
self.ent_embeddings.weight = nn.Parameter(
torch.cat([norm_ent_weight,
torch.zeros(1, self.embedding_size)],
dim=0))
self.rel_embeddings.weight = nn.Parameter(rel_weight)
self.norm_embeddings.weight = nn.Parameter(norm_weight)
normalize_rel_emb = F.normalize(self.rel_embeddings.weight.data,
p=2,
dim=1)
normalize_norm_emb = F.normalize(self.norm_embeddings.weight.data,
p=2,
dim=1)
self.rel_embeddings.weight.data = normalize_rel_emb
self.norm_embeddings.weight.data = normalize_norm_emb
self.ent_embeddings = to_gpu(self.ent_embeddings)
self.rel_embeddings = to_gpu(self.rel_embeddings)
self.norm_embeddings = to_gpu(self.norm_embeddings)
def __init__(self, L1_flag, embedding_size, user_total, item_total,
entity_total, relation_total, i_map, new_map):
super(CKE, self).__init__()
self.L1_flag = L1_flag
self.embedding_size = embedding_size
self.user_total = user_total
self.item_total = item_total
# padding when item are not aligned with any entity
self.ent_total = entity_total + 1
self.rel_total = relation_total
self.is_pretrained = False
# store item to item-entity dic
self.i_map = i_map
# store item-entity to (entity, item)
self.new_map = new_map
# bprmf
# init user and item embeddings
user_weight = torch.FloatTensor(self.user_total, self.embedding_size)
item_weight = torch.FloatTensor(self.item_total, self.embedding_size)
nn.init.xavier_uniform(user_weight)
nn.init.xavier_uniform(item_weight)
self.user_embeddings = nn.Embedding(self.user_total,
self.embedding_size)
self.item_embeddings = nn.Embedding(self.item_total,
self.embedding_size)
self.user_embeddings.weight = nn.Parameter(user_weight)
self.item_embeddings.weight = nn.Parameter(item_weight)
normalize_user_emb = F.normalize(self.user_embeddings.weight.data,
p=2,
dim=1)
normalize_item_emb = F.normalize(self.item_embeddings.weight.data,
p=2,
dim=1)
self.user_embeddings.weight.data = normalize_user_emb
self.item_embeddings.weight.data = normalize_item_emb
self.user_embeddings = to_gpu(self.user_embeddings)
self.item_embeddings = to_gpu(self.item_embeddings)
# transR
ent_weight = torch.FloatTensor(self.ent_total - 1, self.embedding_size)
rel_weight = torch.FloatTensor(self.rel_total, self.embedding_size)
proj_weight = torch.FloatTensor(
self.rel_total, self.embedding_size * self.embedding_size)
nn.init.xavier_uniform(ent_weight)
nn.init.xavier_uniform(rel_weight)
norm_ent_weight = F.normalize(ent_weight, p=2, dim=1)
if self.is_pretrained:
nn.init.eye(proj_weight)
proj_weight = proj_weight.view(-1).expand(self.relation_total, -1)
else:
nn.init.xavier_uniform(proj_weight)
# init user and item embeddings
self.ent_embeddings = nn.Embedding(self.ent_total,
self.embedding_size,
padding_idx=self.ent_total - 1)
self.rel_embeddings = nn.Embedding(self.rel_total, self.embedding_size)
self.proj_embeddings = nn.Embedding(
self.rel_total, self.embedding_size * self.embedding_size)
self.ent_embeddings.weight = nn.Parameter(
torch.cat([norm_ent_weight,
torch.zeros(1, self.embedding_size)],
dim=0))
self.rel_embeddings.weight = nn.Parameter(rel_weight)
self.proj_embeddings.weight = nn.Parameter(proj_weight)
# normalize_ent_emb = F.normalize(self.ent_embeddings.weight.data, p=2, dim=1)
normalize_rel_emb = F.normalize(self.rel_embeddings.weight.data,
p=2,
dim=1)
# normalize_proj_emb = F.normalize(self.proj_embeddings.weight.data, p=2, dim=1)
# self.ent_embeddings.weight.data = normalize_ent_emb
self.rel_embeddings.weight.data = normalize_rel_emb
# self.proj_embeddings.weight.data = normalize_proj_emb
self.ent_embeddings = to_gpu(self.ent_embeddings)
self.rel_embeddings = to_gpu(self.rel_embeddings)
self.proj_embeddings = to_gpu(self.proj_embeddings)