def forward(self, inputs):
text = inputs[0]
pos_tag = inputs[1]
neg_tag = inputs[2]
text_emb = self.text_embedding(text)
text_emb = paddle.reshape(
text_emb, shape=[-1, self.text_len, self.emb_dim])
pos_tag_emb = self.tag_embedding(pos_tag)
pos_tag_emb = paddle.reshape(pos_tag_emb, shape=[-1, self.emb_dim])
neg_tag_emb = self.tag_embedding(neg_tag)
neg_tag_emb = paddle.reshape(
neg_tag_emb, shape=[-1, self.neg_size, self.emb_dim])
conv_1d = self.conv(text_emb)
act = paddle.tanh(conv_1d)
maxpool = paddle.max(act, axis=1)
maxpool = paddle.reshape(maxpool, shape=[-1, self.hid_dim])
text_hid = self.hid_fc(maxpool)
cos_pos = F.cosine_similarity(
pos_tag_emb, text_hid, axis=1).reshape([-1, 1])
# fluid.layers.Print(cos_pos)
neg_tag_emb = paddle.max(neg_tag_emb, axis=1)
neg_tag_emb = paddle.reshape(neg_tag_emb, shape=[-1, self.emb_dim])
cos_neg = F.cosine_similarity(
neg_tag_emb, text_hid, axis=1).reshape([-1, 1])
# fluid.layers.Print(cos_neg)
return cos_pos, cos_neg
def forward(self, input_data, is_infer):
query_fc = input_data[0]
for n_layer in self._query_layers:
query_fc = n_layer(query_fc)
doc_pos_fc = input_data[1]
for n_layer in self._pos_layers:
doc_pos_fc = n_layer(doc_pos_fc)
R_Q_D_p = F.cosine_similarity(query_fc, doc_pos_fc, axis=1,
eps=0).reshape([-1, 1])
if is_infer:
return R_Q_D_p, paddle.ones(shape=[self.slice_end, 1])
R_Q_D_ns = []
for i in range(len(input_data) - 2):
doc_neg_fc_i = input_data[i + 2]
for n_layer in self._neg_layers:
doc_neg_fc_i = n_layer(doc_neg_fc_i)
R_Q_D_n = F.cosine_similarity(query_fc,
doc_neg_fc_i,
axis=1,
eps=0).reshape([-1, 1])
R_Q_D_ns.append(R_Q_D_n)
concat_Rs = paddle.concat(x=[R_Q_D_p] + R_Q_D_ns, axis=1)
prob = F.softmax(concat_Rs, axis=1)
hit_prob = paddle.slice(prob,
axes=[0, 1],
starts=[0, 0],
ends=[self.slice_end, -1])
return R_Q_D_p, hit_prob
def forward(self, inputs, is_infer=False):
self.q_slots = inputs[0]
self.pt_slots = inputs[1]
if not is_infer:
self.nt_slots = inputs[2]
q_embs = [self.embedding(query) for query in self.q_slots]
q_encodes = []
for emb in q_embs:
emb = paddle.reshape(
emb, shape=[-1, self.query_len, self.query_encode_dim])
gru = self.gru(emb)
maxpool = paddle.max(gru[0], axis=1)
maxpool = paddle.reshape(maxpool,
shape=[-1, self.query_encode_dim])
q_encodes.append(maxpool)
q_concat = paddle.concat(q_encodes, axis=1)
q_hid = self.q_fc(q_concat)
pt_embs = [self.embedding(title) for title in self.pt_slots]
pt_encodes = []
for emb in pt_embs:
emb = paddle.reshape(
emb, shape=[-1, self.pos_len, self.title_encode_dim])
gru = self.gru(emb)
maxpool = paddle.max(gru[0], axis=1)
maxpool = paddle.reshape(maxpool,
shape=[-1, self.title_encode_dim])
pt_encodes.append(maxpool)
pt_concat = paddle.concat(pt_encodes, axis=1)
pt_hid = self.t_fc(pt_concat)
cos_pos = F.cosine_similarity(q_hid, pt_hid, axis=1).reshape([-1, 1])
if is_infer:
return cos_pos, paddle.ones(shape=[1, 1])
nt_embs = [self.embedding(title) for title in self.nt_slots]
nt_encodes = []
for emb in nt_embs:
emb = paddle.reshape(
emb, shape=[-1, self.neg_len, self.title_encode_dim])
gru = self.gru(emb)
maxpool = paddle.max(gru[0], axis=1)
maxpool = paddle.reshape(maxpool,
shape=[-1, self.title_encode_dim])
nt_encodes.append(maxpool)
nt_concat = paddle.concat(nt_encodes, axis=1)
nt_hid = self.t_fc(nt_concat)
cos_neg = F.cosine_similarity(q_hid, nt_hid, axis=1).reshape([-1, 1])
return cos_pos, cos_neg
def forward(self, batch_size, user_sparse_inputs, mov_sparse_inputs,
label_input):
user_sparse_embed_seq = []
for s_input in user_sparse_inputs:
emb = self.embedding(s_input)
emb = paddle.reshape(emb, shape=[-1, self.sparse_feature_dim])
user_sparse_embed_seq.append(emb)
mov_sparse_embed_seq = []
for s_input in mov_sparse_inputs:
s_input = paddle.reshape(s_input, shape=[batch_size, -1])
emb = self.embedding(s_input)
emb = paddle.sum(emb, axis=1)
emb = paddle.reshape(emb, shape=[-1, self.sparse_feature_dim])
mov_sparse_embed_seq.append(emb)
user_features = paddle.concat(user_sparse_embed_seq, axis=1)
mov_features = paddle.concat(mov_sparse_embed_seq, axis=1)
for n_layer in self._user_layers:
user_features = n_layer(user_features)
for n_layer in self._movie_layers:
mov_features = n_layer(mov_features)
sim = F.cosine_similarity(user_features, mov_features,
axis=1).reshape([-1, 1])
predict = paddle.scale(sim, scale=5)
return predict
def forward(self, usr_var, mov_var):
# 计算用户特征和电影特征
user_features = self.get_usr_feat(usr_var)
mov_features = self.get_mov_feat(mov_var)
#使用余弦相似度算子，计算用户和电影的相似程度
sim = F.cosine_similarity(user_features, mov_features,
axis=1).reshape([-1, 1])
# 将相似度扩大范围到和电影评分相同数据范围
res = paddle.scale(sim, scale=5)
return user_features, mov_features, res
def test_dygraph_2(self):
paddle.disable_static()
shape = [12, 13]
axis = 0
eps = 1e-6
np.random.seed(1)
np_x1 = np.random.rand(*shape).astype(np.float32)
np_x2 = np.random.rand(*shape).astype(np.float32)
np_out = self._get_numpy_out(np_x1, np_x2, axis=axis, eps=eps)
tesnor_x1 = paddle.to_tensor(np_x1)
tesnor_x2 = paddle.to_tensor(np_x2)
y = F.cosine_similarity(tesnor_x1, tesnor_x2, axis=axis, eps=eps)
self.assertTrue(np.allclose(y.numpy(), np_out))
def test_dygraph_3(self):
paddle.disable_static()
shape1 = [10, 12, 10]
shape2 = [10, 1, 10]
axis = 2
eps = 1e-6
np.random.seed(1)
np_x1 = np.random.rand(*shape1).astype(np.float32)
np_x2 = np.random.rand(*shape2).astype(np.float32)
np_out = self._get_numpy_out(np_x1, np_x2, axis=axis, eps=eps)
tesnor_x1 = paddle.to_variable(np_x1)
tesnor_x2 = paddle.to_variable(np_x2)
y = F.cosine_similarity(tesnor_x1, tesnor_x2, axis=axis, eps=eps)
self.assertTrue(np.allclose(y.numpy(), np_out))
def check_static_result(self, place):
paddle.enable_static()
with program_guard(Program(), Program()):
shape = [10, 15]
axis = 1
eps = 1e-8
np.random.seed(0)
np_x1 = np.random.rand(*shape).astype(np.float32)
np_x2 = np.random.rand(*shape).astype(np.float32)
x1 = paddle.fluid.data(name="x1", shape=shape)
x2 = paddle.fluid.data(name="x2", shape=shape)
result = F.cosine_similarity(x1, x2, axis=axis, eps=eps)
exe = Executor(place)
fetches = exe.run(default_main_program(),
feed={"x1": np_x1,
"x2": np_x2},
fetch_list=[result])
np_out = self._get_numpy_out(np_x1, np_x2, axis=axis, eps=eps)
self.assertTrue(np.allclose(fetches[0], np_out))
