class Net(nn.Module):
def __init__(self, args, dev_id):
super(Net, self).__init__()
self._act = get_activation(args.model_activation)
self.encoder = GCMCLayer(args.rating_vals,
args.src_in_units,
args.dst_in_units,
args.gcn_agg_units,
args.gcn_out_units,
args.gcn_dropout,
args.gcn_agg_accum,
agg_act=self._act,
share_user_item_param=args.share_param,
device=dev_id)
if args.mix_cpu_gpu and args.use_one_hot_fea:
# if use_one_hot_fea, user and movie feature is None
# W can be extremely large, with mix_cpu_gpu W should be stored in CPU
self.encoder.partial_to(dev_id)
else:
self.encoder.to(dev_id)
self.decoder = BiDecoder(in_units=args.gcn_out_units,
num_classes=len(args.rating_vals),
num_basis=args.gen_r_num_basis_func)
self.decoder.to(dev_id)
def forward(self, compact_g, frontier, ufeat, ifeat,
possible_rating_values):
user_out, movie_out = self.encoder(frontier, ufeat, ifeat)
pred_ratings = self.decoder(compact_g, user_out, movie_out)
return pred_ratings
class Net(nn.Module):
def __init__(self, args, dev_id):
super(Net, self).__init__()
self._act = get_activation(args.model_activation)
self.encoder = nn.ModuleList()
self.encoder.append(
GCMCLayer(args.rating_vals,
args.src_in_units,
args.dst_in_units,
args.gcn_agg_units,
args.gcn_out_units,
args.gcn_dropout,
args.gcn_agg_accum,
agg_act=self._act,
share_user_item_param=args.share_param,
device=dev_id))
self.rating_vals = args.rating_vals
self.gcn_agg_accum = args.gcn_agg_accum
self.rating_vals = args.rating_vals
self.device = dev_id
self.gcn_agg_units = args.gcn_agg_units
self.src_in_units = args.src_in_units
self.batch_size = args.minibatch_size
for i in range(1, args.layers):
if args.gcn_agg_accum == 'stack':
gcn_out_units = args.gcn_out_units * len(args.rating_vals)
else:
gcn_out_units = args.gcn_out_units
self.encoder.append(
GCMCLayer(args.rating_vals,
args.gcn_out_units,
args.gcn_out_units,
gcn_out_units,
args.gcn_out_units,
args.gcn_dropout - i * 0.1,
args.gcn_agg_accum,
agg_act=self._act,
share_user_item_param=args.share_param,
ini=False,
device=dev_id))
if args.mix_cpu_gpu and args.use_one_hot_fea:
# if use_one_hot_fea, user and movie feature is None
# W can be extremely large, with mix_cpu_gpu W should be stored in CPU
self.encoder.partial_to(dev_id)
else:
self.encoder.to(dev_id)
self.decoder = BiDecoder(in_units=args.gcn_out_units,
num_classes=len(args.rating_vals),
num_basis=args.gen_r_num_basis_func)
self.decoder.to(dev_id)
def forward(self,
compact_g,
frontier,
ufeat,
ifeat,
possible_rating_values,
Two_Stage=False):
# user_out, movie_out = self.encoder(frontier, ufeat, ifeat)
'''
user_out_2, movie_out_2 = self.encoder_2(frontier, user_out, movie_out)
user_out = th.cat([user_out, user_out_2], 1)
movie_out = th.cat([movie_out, movie_out_2], 1)
'''
user_out = []
movie_out = []
for i in range(0, args.layers):
user_o, movie_o = self.encoder[i](frontier[i], ufeat, ifeat)
ufeat = user_o
ifeat = movie_o
user_out.append(user_o)
movie_out.append(movie_o)
u_size = user_o.shape[0]
m_size = movie_o.shape[0]
for i in range(0, args.layers):
if i == 0:
user_o = user_out[i][:u_size, :]
movie_o = movie_out[i][:m_size, :]
else:
user_o += user_out[i][:u_size, :] / float(i + 1)
movie_o += movie_out[i][:m_size, :] / float(i + 1)
#user_out.append(user_o)
#movie_out.append(movie_o)
# pred_ratings = self.decoder(compact_g, user_out, movie_out)
# W_r_last = None
# reg_loss = 0.0
# for rating in self.rating_vals:
# rating = to_etype_name(rating)
# if W_r_last is not None:
# reg_loss += th.sum((self.encoder.W_r[rating] - W_r_last)**2)
# W_r_last = self.encoder.W_r[rating]
# return pred_ratings, reg_loss
pred_ratings = self.decoder(compact_g, user_o, movie_o)
W_r_last = None
reg_loss = 0.0
'''
for rating in self.rating_vals:
rating = to_etype_name(rating)
if W_r_last is not None:
reg_loss += th.sum((self.encoder[0].W_r[rating] - W_r_last)**2)
W_r_last = self.encoder[0].W_r[rating]
#W_r_last_2 = self.encoder_2.W_r[rating]
'''
W = th.matmul(
self.encoder[0].att,
self.encoder[0].basis.view(self.encoder[0].basis_units, -1))
W = W.view(len(self.rating_vals), self.src_in_units, -1)
for i, rating in enumerate(self.rating_vals):
rating = to_etype_name(rating)
if i != 0:
reg_loss += -th.sum(
th.cosine_similarity(W[i, :, :], W[i - 1, :, :], dim=1))
return pred_ratings, reg_loss