def __init__(self, c, h, w, name):
super(SaModel, self).__init__()
self.name = name
self.l1 = nn.Linear(w, h)
utils.init_linear(self.l1)
self.l2 = nn.Linear(h * w, c)
utils.init_linear(self.l2)
def rand_init(self):
"""
random initialization
"""
for i in range(self.num_layers):
utils.init_linear(self.trans[i])
utils.init_linear(self.gate[i])
def __init__(self,
features_num_embed=[],
embed_dim=64,
embed_drop_out=0.1):
super(FeatureEmbed, self).__init__()
self.features_embed_module = nn.ModuleList()
for num_embed in features_num_embed:
self.features_embed_module.append(
NormalEmbedding(num_embed, embed_dim, embed_drop_out))
self.linear_out = nn.Linear(
len(features_num_embed) * embed_dim, embed_dim)
utils.init_linear(self.linear_out)
def __init__(self,
url_f_embed_module=None,
user_f_embed_module=None,
embed_drop_out=0.1,
num_limit_list=[]):
super(Model6Features, self).__init__()
self.feature_model = feature_models.FeatureModel(
out_conv=True,
attend_f=False,
sa=False,
embed_drop_out=embed_drop_out,
num_limit_list=num_limit_list)
if url_f_embed_module is not None:
self.feature_model.url_features_embed = url_f_embed_module.features_embed_module
self.feature_model.url = url_f_embed_module.linear_out
if user_f_embed_module is not None:
self.feature_model.user_features_embed = user_f_embed_module.features_embed_module
self.feature_model.user = user_f_embed_module.linear_out
self.linear_out = nn.Linear(F_DIM * (2 + 2), 2)
utils.init_linear(self.linear_out)
def __init__(self,
url_f_embed_module=None,
user_f_embed_module=None,
embed_drop_out=0.1,
if_easy=True):
super(Model6Pre, self).__init__()
self.if_easy = if_easy
if if_easy:
self.user_embed = embed.FeatureEmbed([11576 + 1], F_DIM, 0.1)
self.url_embed = embed.FeatureEmbed([4732 + 1], F_DIM, 0.1)
else:
# int(uid_no), follow_cnt_bin, follow_in_cnt_bin, post_url_bin, post_cnt_bin
# ,favorite_cate_no, favorite_site_no
self.user_embed = embed.FeatureEmbed(
[11576 + 1, 10 + 1, 10 + 1, 10 + 1, 10 + 1, 20 + 1, 6 + 1],
F_DIM, 0.1)
#int(url_no), cate2idx[cate], site2idx[url_site], post_user_bin, post_freq_bin
self.url_embed = embed.FeatureEmbed(
[4732 + 1, 20 + 1, 6 + 1, 10 + 1, 10 + 1], F_DIM, 0.1)
self.linear_out = nn.Linear(F_DIM * 2, 2)
utils.init_linear(self.linear_out)
def __init__(self, word_embed_mat, dim=64):
super(TextModel, self).__init__()
### tweets and url content
if word_embed_mat is None:
self.word_embed = embed.NormalEmbedding(231172, 300, 0.1)
else:
self.word_embed = embed.FixedEmbedding(word_embed_mat, 0, False)
self.char_embed = embed.CharacterEmbedding(num_embed=70,
embed_dim=64,
out_ch=64,
kernel_size=3,
padding=1)
#self.char_embed = embed.CharacterDepthwiseEmbedding(num_embed=70, embed_dim=64, out_ch=64, kernel_size=3)
#self.char_embed2 = embed.CharacterEmbedding(num_embed=70, embed_dim=64, out_ch=64, kernel_size=5, padding=2)
#self.projection = nn.Conv1d(300+64, 128, 1)
self.hw1 = highway.hw(300 + 64, 2)
#self.hw2 = highway.hw(300+64, 2)
'''self.lstm = nn.LSTM( # if use nn.RNN(), it hardly learns
input_size=128,
hidden_size=64, # rnn hidden unit
num_layers=1, # number of rnn layer
batch_first=True, # input & output will has batch size as 1s dimension. e.g. (batch, time_step, input_size)
bidirectional=True
)'''
### attention
self.cos = torch.nn.CosineSimilarity(dim=2, eps=1e-6)
#self.linear_attention = nn.Linear(300+64, 300+64)
#utils.init_linear(self.linear_attention)
#self.linear_attention_url = nn.Linear(300+64, 300+64)
#utils.init_linear(self.linear_attention_url)
self.linear = nn.Linear(300 + 64, 64)
utils.init_linear(self.linear)
def __init__(self, out_conv=True, attend_f=True, sa=False, embed_drop_out=0.1
,num_limit_list=[CONSUM_NUM, NEIGHB_URL_NUM, FRIEND_NUM, NEIGHT_USER_NUM], F_DIM=F_DIM):
super(FeatureModel, self).__init__()
self.out_conv = out_conv
self.attend_f = attend_f
self.sa = sa
self.user = nn.Linear(7*F_DIM, F_DIM)
utils.init_linear(self.user)
self.url = nn.Linear(5*F_DIM, F_DIM)
utils.init_linear(self.url)
#self.user_url = nn.Linear(F_DIM, F_DIM)
#utils.init_linear(self.user_url)
self.cos = nn.CosineSimilarity(dim=2, eps=1e-6)
# url features: int(url_no), cate2idx[cate], site2idx[url_site], post_user_bin, post_freq_bin
self.url_features_embed = nn.ModuleList()
self.url_features_embed.append(embed.NormalEmbedding(4732+1, F_DIM, embed_drop_out))
self.url_features_embed.append(embed.NormalEmbedding(20+1, F_DIM, embed_drop_out))
self.url_features_embed.append(embed.NormalEmbedding(6+1, F_DIM, embed_drop_out))
self.url_features_embed.append(embed.NormalEmbedding(10+1, F_DIM, embed_drop_out))
self.url_features_embed.append(embed.NormalEmbedding(10+1, F_DIM, embed_drop_out))
self.url_consume = user_model.attenModel(in_c=num_limit_list[0], out_c=num_limit_list[0], name='url_consume',
attend_f=self.attend_f, out_conv=self.out_conv, sa=self.sa, h=5, w=F_DIM)
self.url_neighb = user_model.attenModel(in_c=num_limit_list[1], out_c=num_limit_list[1], name='url_neighb',
attend_f=self.attend_f, out_conv=self.out_conv, sa=self.sa, h=5, w=F_DIM)
self.url_gate = nn.Linear(F_DIM*2, F_DIM)
utils.init_linear(self.url_gate)
# user features: int(uid_no), follow_cnt_bin, follow_in_cnt_bin, post_url_bin, post_cnt_bin ,favorite_cate_no, favorite_site_no
self.user_features_embed = nn.ModuleList()
self.user_features_embed.append(embed.NormalEmbedding(11576+1, F_DIM, embed_drop_out))
self.user_features_embed.append(embed.NormalEmbedding(10+1, F_DIM, embed_drop_out))
self.user_features_embed.append(embed.NormalEmbedding(10+1, F_DIM, embed_drop_out))
self.user_features_embed.append(embed.NormalEmbedding(10+1, F_DIM, embed_drop_out))
self.user_features_embed.append(embed.NormalEmbedding(10+1, F_DIM, embed_drop_out))
self.user_features_embed.append(embed.NormalEmbedding(20+1, F_DIM, embed_drop_out))
self.user_features_embed.append(embed.NormalEmbedding(6+1, F_DIM, embed_drop_out))
self.user_friend = user_model.attenModel(in_c=num_limit_list[2], out_c=num_limit_list[2], name='user_friend',
attend_f=self.attend_f, out_conv=self.out_conv, sa=self.sa, h=7, w=F_DIM)
self.user_neighb = user_model.attenModel(in_c=num_limit_list[3], out_c=num_limit_list[3], name='user_neighb',
attend_f=self.attend_f, out_conv=self.out_conv, sa=self.sa, h=7, w=F_DIM)
self.user_gate = nn.Linear(F_DIM*2, F_DIM)
utils.init_linear(self.user_gate)
def __init__(self, x_dim=64, mem_dim=364, out_dim=64, attention_method='cos'):
super(memAttention, self).__init__()
self.method = attention_method
### x linear
self.linear_x = nn.Linear(x_dim, mem_dim)
utils.init_linear(self.linear_x)
### attention
if self.method == 'cos':
self.cos = torch.nn.CosineSimilarity(dim=2, eps=1e-6)
else:
self.linear_attention = nn.Linear(mem_dim, mem_dim)
utils.init_linear(self.linear_attention)
### out
self.linear_out = nn.Linear(mem_dim, out_dim)
utils.init_linear(self.linear_out)
def __init__(self,
url_f_embed_module=None,
user_f_embed_module=None,
embed_drop_out=0.1,
if_one_layer=False,
num_limit_list=[],
F_DIM=64):
super(Model6Main, self).__init__()
self.if_one_layer = if_one_layer
# feature model
# num_limit_list=[CONSUM_NUM, NEIGHB_URL_NUM, FRIEND_NUM, NEIGHT_USER_NUM]
self.feature_model = feature_models.FeatureModel(
out_conv=True,
attend_f=True,
sa=False,
embed_drop_out=embed_drop_out,
num_limit_list=num_limit_list,
F_DIM=F_DIM)
# int(uid_no), follow_cnt_bin, follow_in_cnt_bin, post_url_bin, post_cnt_bin
# ,favorite_cate_no, favorite_site_no
self.user_embed = embed.FeatureEmbed(
[11576 + 1, 10 + 1, 10 + 1, 10 + 1, 10 + 1, 20 + 1, 6 + 1], F_DIM,
embed_drop_out)
#int(url_no), cate2idx[cate], site2idx[url_site], post_user_bin, post_freq_bin
self.url_embed = embed.FeatureEmbed(
[4732 + 1, 20 + 1, 6 + 1, 10 + 1, 10 + 1], F_DIM, embed_drop_out)
self.cos2 = torch.nn.CosineSimilarity(dim=2, eps=1e-6)
self.cos3 = torch.nn.CosineSimilarity(dim=3, eps=1e-6)
self.linear_update2 = nn.Linear(F_DIM * 2, F_DIM)
utils.init_linear(self.linear_update2)
self.linear_update1 = nn.Linear(F_DIM * 2, F_DIM)
utils.init_linear(self.linear_update1)
self.linear_out = nn.Linear(F_DIM * (2 + 4), 2)
utils.init_linear(self.linear_out)
def __init__(self, url_f_embed_module=None, user_f_embed_module=None, embed_drop_out=0.1, if_one_layer=False):
super(Model6GCN, self).__init__()
self.if_one_layer = if_one_layer
# int(uid_no), follow_cnt_bin, follow_in_cnt_bin, post_url_bin, post_cnt_bin
# ,favorite_cate_no, favorite_site_no
self.user_embed = embed.FeatureEmbed([11576+1, 10+1, 10+1, 10+1, 10+1, 20+1, 6+1], F_DIM, embed_drop_out)
#int(url_no), cate2idx[cate], site2idx[url_site], post_user_bin, post_freq_bin
self.url_embed = embed.FeatureEmbed([4732+1, 20+1, 6+1, 10+1, 10+1], F_DIM, embed_drop_out)
if url_f_embed_module is not None:
print(self.url_embed.features_embed_module[2].embed.weight[0:2, :])
self.url_embed = url_f_embed_module
print(self.url_embed.features_embed_module[2].embed.weight[0:2, :])
'''for i in range(len(self.feature_model.url_features_embed)):
#print(i, 'url', self.feature_model.url_features_embed[i].embed.weight.requires_grad)
self.feature_model.url_features_embed[i].embed.weight.requires_grad = False
#print(i, 'url', self.feature_model.url_features_embed[i].embed.weight.requires_grad)'''
if user_f_embed_module is not None:
#print(self.user_embed.features_embed_module[2].embed.weight[0:2, :])
self.user_embed = user_f_embed_module
#print(self.user_embed.features_embed_module[2].embed.weight[0:2, :])
'''for i in range(len(self.feature_model.user_features_embed)):
#print(i, 'user', self.feature_model.user_features_embed[i].embed.weight.requires_grad)
self.feature_model.user_features_embed[i].embed.weight.requires_grad = False
#print(i, 'user', self.feature_model.user_features_embed[i].embed.weight.requires_grad)'''
self.cos2 = torch.nn.CosineSimilarity(dim=2, eps=1e-6)
self.cos3 = torch.nn.CosineSimilarity(dim=3, eps=1e-6)
self.linear_update2 = nn.Linear(F_DIM*2, F_DIM)
utils.init_linear(self.linear_update2)
self.linear_update1 = nn.Linear(F_DIM*2, F_DIM)
utils.init_linear(self.linear_update1)
self.linear_out = nn.Linear(F_DIM*2, 2)
utils.init_linear(self.linear_out)