def __init__(self, args, input_size, rating_range=(1, 5)):
super(Autorec_DT, self).__init__()
self.args = args
self.time_nn = TimeNN(args, n_time_inputs=2)
self.film_time = FilmLayer(args)
self.dropout_input = nn.Dropout(model_params['dropout_input'])
self.encoder = nn.Linear(input_size, model_params['hidden_size'])
self.sig_act = nn.Sigmoid()
self.dropout_emb = nn.Dropout(model_params['dropout_emb'])
self.decoder = nn.Linear(model_params['hidden_size'], input_size)
self.limiter = nn.Hardtanh(rating_range[0], rating_range[1])
def __init__(self, args, input_size, ft_size, rating_range=(1, 5)):
super(Autorec_DFT2, self).__init__()
self.args = args
self.time_nn = TimeNN(args, n_time_inputs=3)
self.film_time = FilmLayer(args)
self.dropout_input = nn.Dropout(0.7)
self.encoder = nn.Linear(input_size, 500).to(args.device)
self.sig_act = nn.Sigmoid()
self.dropout_emb = nn.Dropout(0.5)
self.decoder = nn.Linear(500, input_size).to(args.device)
self.limiter = nn.Hardtanh(rating_range[0], rating_range[1])
self.ft_model = ContentFiltering(args, ft_size)
self.ft_comb = FeatureCombiner(args)
class Autorec_DFGT(nn.Module):
requires_time = True
requires_fts = True
requires_graph = True
def __init__(self, args, input_size, ft_size, rating_range=(1, 5)):
super(Autorec_DFGT, self).__init__()
self.args = args
self.time_nn = TimeNN(args, n_time_inputs=3)
self.film_time = FilmLayer(args)
self.dropout_input = nn.Dropout(0.7)
self.encoder = nn.Linear(input_size, 500).to(args.device)
self.sig_act = nn.Sigmoid()
self.conv = GraphConv0D(args).to(args.device)
self.dropout_emb = nn.Dropout(0.5)
self.decoder = nn.Linear(500, input_size).to(args.device)
self.limiter = nn.Hardtanh(rating_range[0], rating_range[1])
self.ft_model = ContentFiltering(args, ft_size)
self.ft_comb = FeatureCombiner(args)
def forward(self, x, time_x, ft_x, ft_n, graph):
graph = graph[0]
time_x = self.time_nn(time_x)
time_x = time_x * (x > 0)
h = self.film_time(x, time_x)
hf = self.ft_model(ft_x)
h = self.ft_comb(h, hf, ft_n)
h = self.dropout_input(h)
h = self.sig_act(self.encoder(h))
if graph['edge_index'].shape[0] == 2:
h = self.conv(h, graph['edge_index'], graph['edge_weight'])
h = self.dropout_emb(h)
p = self.decoder(h)
if not self.training:
p = self.limiter(p)
return p
def get_reg_loss(self):
reg_loss = self.args.reg / 2 * (torch.norm(self.encoder.weight)**2 +
torch.norm(self.decoder.weight)**2)
reg_loss += self.time_nn.get_reg_loss()
reg_loss += self.ft_model.get_reg_loss()
return reg_loss
class Autorec_DT(nn.Module):
requires_time = True
requires_fts = False
requires_graph = False
def __init__(self, args, input_size, rating_range=(1, 5)):
super(Autorec_DT, self).__init__()
self.args = args
self.time_nn = TimeNN(args, n_time_inputs=2)
self.film_time = FilmLayer(args)
self.dropout_input = nn.Dropout(model_params['dropout_input'])
self.encoder = nn.Linear(input_size, model_params['hidden_size'])
self.sig_act = nn.Sigmoid()
self.dropout_emb = nn.Dropout(model_params['dropout_emb'])
self.decoder = nn.Linear(model_params['hidden_size'], input_size)
self.limiter = nn.Hardtanh(rating_range[0], rating_range[1])
def forward(self, x, ft_n, time_x):
time_x = self.time_nn(time_x[..., :2])
time_x = time_x * (x > 0)
x = self.film_time(x, time_x)
x = self.dropout_input(x)
x = self.sig_act(self.encoder(x))
x = self.dropout_emb(x)
p = self.decoder(x)
if not self.training:
p = self.limiter(p)
p[torch.nonzero(ft_n[0] == 0), :] = ft_n[2]
p[:, torch.nonzero(ft_n[1] == 0)] = ft_n[2]
return p
def get_reg_loss(self):
reg_loss = self.args.reg / 2 * (torch.norm(self.encoder.weight)**2 +
torch.norm(self.decoder.weight)**2)
reg_loss += self.time_nn.get_reg_loss()
return reg_loss