def forward(self, pattern, pattern_len, graph, graph_len):
bsz = pattern_len.size(0)
gate = self.get_filter_gate(pattern, pattern_len, graph, graph_len)
zero_mask = (gate == 0).unsqueeze(-1) if gate is not None else None
pattern_emb, graph_emb = self.get_emb(pattern, pattern_len, graph, graph_len)
if zero_mask is not None:
graph_emb.masked_fill_(zero_mask, 0.0)
pattern_output = pattern_emb
for p_rnn in self.p_net:
o = p_rnn(pattern_output)
pattern_output = o + pattern_output
pattern_mask = (batch_convert_len_to_mask(pattern_len)==0).unsqueeze(-1)
pattern_output.masked_fill_(pattern_mask, 0.0)
graph_output = graph_emb
for g_rnn in self.g_net:
o = g_rnn(graph_output)
graph_output = o + graph_output
if zero_mask is not None:
graph_output.masked_fill_(zero_mask, 0.0)
graph_mask = (batch_convert_len_to_mask(graph_len)==0).unsqueeze(-1)
graph_output.masked_fill_(graph_mask, 0.0)
if self.add_enc:
pattern_enc, graph_enc = self.get_enc(pattern, pattern_len, graph, graph_len)
if zero_mask is not None:
graph_enc.masked_fill_(zero_mask, 0.0)
pattern_output = torch.cat([pattern_enc, pattern_output], dim=2)
graph_output = torch.cat([graph_enc, graph_output], dim=2)
pred = self.predict_net(pattern_output, pattern_len, graph_output, graph_len)
return pred
def forward(self, pattern, pattern_len, graph, graph_len):
bsz = pattern_len.size(0)
p_len, g_len = pattern.size(1), graph.size(1)
plf, glf = pattern_len.float(), graph_len.float()
inv_plf, inv_glf = 1.0 / plf, 1.0 / glf
p_mask = batch_convert_len_to_mask(pattern_len) if p_len == pattern_len.max() else None
g_mask = batch_convert_len_to_mask(graph_len) if g_len == graph_len.max() else None
p, g = pattern, graph
if g_mask is not None:
mem = list()
mem_mask = list()
for idx in gather_indices_by_lens(graph_len):
if self.mem_init.endswith("attn"):
m, mk = init_mem(g[idx, :graph_len[idx[0]]], g_mask[idx, :graph_len[idx[0]]],
mem_len=self.mem_len, mem_init=self.mem_init, attn=self.m_layer)
elif self.mem_init.endswith("lstm"):
m, mk = init_mem(g[idx, :graph_len[idx[0]]], g_mask[idx, :graph_len[idx[0]]],
mem_len=self.mem_len, mem_init=self.mem_init, lstm=self.m_layer)
else:
m, mk = init_mem(g[idx, :graph_len[idx[0]]], g_mask[idx, :graph_len[idx[0]]],
mem_len=self.mem_len, mem_init=self.mem_init, post_proj=self.m_layer)
mem.append(m)
mem_mask.append(mk)
mem = torch.cat(mem, dim=0)
mem_mask = torch.cat(mem_mask, dim=0)
else:
if self.mem_init.endswith("attn"):
mem, mem_mask = init_mem(keyvalue, None,
mem_len=self.mem_len, mem_init=self.mem_init, attn=self.m_layer)
elif self.mem_init.endswith("lstm"):
mem, mem_mask = init_mem(keyvalue, None,
mem_len=self.mem_len, mem_init=self.mem_init, lstm=self.m_layer)
else:
mem, mem_mask = init_mem(keyvalue, None,
mem_len=self.mem_len, mem_init=self.mem_init, post_proj=self.m_layer)
for i in range(self.recurrent_steps):
mem = self.p_attn(mem, p, p, p_mask)
mem = self.g_attn(mem, g, g, g_mask)
mem = mem.view(bsz, -1)
y = self.pred_layer1(torch.cat([mem, plf, glf, inv_plf, inv_glf], dim=1))
y = self.act(y)
y = self.pred_layer2(torch.cat([y, plf, glf, inv_plf, inv_glf], dim=1))
return y
def forward(self, pattern, pattern_len, graph, graph_len):
bsz = pattern_len.size(0)
p_len, g_len = pattern.size(1), graph.size(1)
plf, glf = pattern_len.float(), graph_len.float()
inv_plf, inv_glf = 1.0 / plf, 1.0 / glf
p_mask = batch_convert_len_to_mask(pattern_len) if p_len == pattern_len.max() else None
g_mask = batch_convert_len_to_mask(graph_len) if g_len == graph_len.max() else None
p, g = pattern, graph
for i in range(self.recurrent_steps):
g = self.p_attn(g, p, p_mask)
g = self.g_attn(g, g, g_mask)
p = self.drop(self.p_layer(torch.max(p, dim=1, keepdim=True)[0]))
g = self.drop(self.g_layer(g))
p = p.squeeze(1)
g = torch.max(g, dim=1)[0]
y = self.pred_layer1(torch.cat([p, g, g-p, g*p, plf, glf, inv_plf, inv_glf], dim=1))
y = self.act(y)
y = self.pred_layer2(torch.cat([y, plf, glf, inv_plf, inv_glf], dim=1))
return y
def encoder_forward(self,
enc_inp,
enc_len,
enc_txl,
enc_params,
mems=None):
qlen = enc_inp.size(1)
mlen = mems[0].size(1) if mems is not None else 0
enc_attn_mask = batch_convert_len_to_mask(enc_len + mlen,
max_seq_len=qlen + mlen)
return self._forward(enc_inp,
enc_len,
enc_txl,
enc_params,
attn_mask=enc_attn_mask,
mems=mems)
def get_filter_gate(self, pattern, pattern_len, graph, graph_len):
gate = None
if self.vl_flt is not None:
gate = self.vl_flt(
split_and_batchify_graph_feats(
pattern.ndata["label"].unsqueeze(-1), pattern_len)[0],
split_and_batchify_graph_feats(
graph.ndata["label"].unsqueeze(-1), graph_len)[0])
if gate is not None:
bsz = graph_len.size(0)
max_g_len = graph_len.max()
if bsz * max_g_len != graph.number_of_nodes():
graph_mask = batch_convert_len_to_mask(
graph_len) # bsz x max_len
gate = gate.masked_select(graph_mask.unsqueeze(-1)).view(-1, 1)
else:
gate = gate.view(-1, 1)
return gate
def forward(self, pattern, pattern_len, graph, graph_len):
# data, target, *mems
# nn.DataParallel does not allow size(0) tensors to be broadcasted.
# So, have to initialize size(0) mems inside the model forward.
# Moreover, have to return new_mems to allow nn.DataParallel to piece
# them together.
bsz = pattern_len.size(0)
gate = self.get_filter_gate(pattern, pattern_len, graph, graph_len)
zero_mask = (gate == 0).unsqueeze(-1) if gate is not None else None
pattern_emb, graph_emb = self.get_emb(pattern, pattern_len, graph,
graph_len)
if zero_mask is not None:
graph_emb.masked_fill_(zero_mask, 0.0)
pattern_emb = self.p_emb_proj(pattern_emb).mul_(self.emb_scale)
graph_emb = self.g_emb_proj(graph_emb).mul_(self.emb_scale)
pattern_segments = segment_data(pattern_emb, self.tgt_len)
pattern_seg_lens = segment_length(pattern_len, self.tgt_len)
graph_segments = segment_data(graph_emb, self.tgt_len)
graph_seg_lens = segment_length(graph_len, self.tgt_len)
pattern_outputs = list()
for i, (pattern_seg, pattern_seg_len) in enumerate(
zip(pattern_segments, pattern_seg_lens)):
if i == 0:
pattern_mems = self.init_mems(len(self.p_net), pattern_seg)
pattern_output, pattern_mems = self.encoder_forward(
pattern_seg,
pattern_seg_len,
self.p_net,
self.p_params,
mems=pattern_mems)
pattern_outputs.append(pattern_output)
pattern_output = torch.cat(pattern_outputs,
dim=1)[:, :pattern_emb.size(1)]
# some segments may only have padded elements, we need to set them as 0 manually
pattern_mask = (batch_convert_len_to_mask(
pattern_len,
max_seq_len=pattern_output.size(1)) == 0).unsqueeze(-1)
pattern_output.masked_fill_(pattern_mask, 0.0)
graph_outputs = list()
for i, (graph_seg,
graph_seg_len) in enumerate(zip(graph_segments,
graph_seg_lens)):
if i == 0:
graph_mems = self.init_mems(len(self.g_net), graph_seg)
graph_output, graph_mems = self.encoder_forward(graph_seg,
graph_seg_len,
self.g_net,
self.g_params,
mems=graph_mems)
graph_outputs.append(graph_output)
graph_output = torch.cat(graph_outputs, dim=1)[:, :graph_emb.size(1)]
# some segments may only have padded elements, we need to set them as 0 manually
graph_mask = (batch_convert_len_to_mask(
graph_len, max_seq_len=graph_output.size(1)) == 0).unsqueeze(-1)
graph_output.masked_fill_(graph_mask, 0.0)
if self.add_enc:
pattern_enc, graph_enc = self.get_enc(pattern, pattern_len, graph,
graph_len)
if zero_mask is not None:
graph_enc.masked_fill_(zero_mask, 0.0)
pattern_output = torch.cat([pattern_enc, pattern_output], dim=2)
graph_output = torch.cat([graph_enc, graph_output], dim=2)
pred = self.predict_net(pattern_output, pattern_len, graph_output,
graph_len)
return pred
