def forward(self, input, hidden, return_h=False):
batch_size = input.size(1)
emb = embedded_dropout(self.encoder,
input,
dropout=self.dropoute if self.training else 0)
emb = self.lockdrop(emb, self.dropouti)
raw_output = emb
new_hidden = []
raw_outputs = []
outputs = []
for l, rnn in enumerate(self.rnns):
raw_output, new_h = rnn(raw_output, hidden[l])
new_hidden.append(new_h)
raw_outputs.append(raw_output)
hidden = new_hidden
output = self.lockdrop(raw_output, self.dropout)
outputs.append(output)
logit = self.decoder(output.view(-1, self.ninp))
log_prob = nn.functional.log_softmax(logit, dim=-1)
model_output = log_prob
model_output = model_output.view(-1, batch_size, self.ntoken)
if return_h:
return model_output, hidden, raw_outputs, outputs
return model_output, hidden
def forward(self, input, hidden, genotype, return_h=False): # 传入的hidden注意下
batch_size = input.size(1)
emb = embedded_dropout(self.encoder,
input,
dropout=self.dropout_e if self.training else 0)
emb = self.lockdrop(emb, self.dropout_i)
raw_output = emb
raw_output, new_h = self.rnn(
raw_output, hidden, genotype) # hidden states， last hidden states
hidden = new_h
output = self.lockdrop(raw_output,
self.dropout) # raw output是rnn每个t时刻的hidden
# 下面这层，实际上就是hidden->output的那个线性层
logit = self.decoder(output.view(
-1, self.n_inp)) # 计算这个batch的logit值，view一下一起计算
log_prob = nn.functional.log_softmax(logit, dim=-1)
model_output = log_prob
model_output = model_output.view(-1, batch_size, self.n_token)
if return_h:
return model_output, hidden, raw_output, output
return model_output, hidden
def forward(self, input, hidden_input_all, BN_start=0):
seq_len, batch_size = input.size()
rnnoutputs = {}
hidden_x = {}
if hidden_input_all is not None:
hidden_x.update(hidden_input_all)
else:
for x in range(len(self.RNNs)):
hidden_x['hidden%d' % x] = Variable(
torch.zeros(1, batch_size, args.hidden_size).cuda())
hidden_x['hidden_lastindrnn'] = Variable(
torch.zeros(1, batch_size, args.embed_size).cuda())
input = input.view(seq_len * batch_size)
input = embedded_dropout(
self.encoder,
input,
dropout=args.dropout_embedding if self.training else 0)
input = input.view(seq_len, batch_size, args.embed_size)
if args.dropout_words > 0:
input = dropout_overtime(input, args.dropout_words, self.training)
rnnoutputs['outlayer-1'] = input
for x in range(len(self.RNNs)):
rnnoutputs['dilayer%d' % x] = self.DIs[x](rnnoutputs['outlayer%d' %
(x - 1)])
rnnoutputs['outlayer%d' %
x], hidden_x['hidden%d' % x] = self.RNNs[x](
rnnoutputs['dilayer%d' % x], BN_start,
hidden_x['hidden%d' % x])
if args.dropout > 0:
dropout = args.dropout
if x == len(self.RNNs) - 1:
dropout = args.dropout_last
rnnoutputs['outlayer%d' % x] = dropout_overtime(
rnnoutputs['outlayer%d' % x], dropout, self.training)
rnn_out = rnnoutputs['outlayer%d' % (len(self.RNNs) - 1)]
rnn_out = self.last_fc(rnn_out)
rnn_out, hidden_x['hidden_lastindrnn'] = self.last_indrnn(
rnn_out, BN_start, hidden_x['hidden_lastindrnn'])
if args.bn_location == 'bn_before':
rnn_out = self.extra_bn(rnn_out, BN_start)
rnn_out = dropout_overtime(rnn_out, args.dropout_extrafc,
self.training)
rnn_out = rnn_out.view(seq_len * batch_size, -1)
output = self.decoder(rnn_out)
return output, hidden_x
def forward(self, input, hidden_input_all, BN_start=0):
seq_len, batch_size = input.size()
rnnoutputs = {}
hidden_x = {}
hid_ind_layer = 0
if hidden_input_all is not None:
hidden_x.update(hidden_input_all)
input = input.view(seq_len * batch_size)
input = embedded_dropout(
self.encoder,
input,
dropout=args.dropout_embedding if self.training else 0)
input = input.view(seq_len, batch_size, args.embed_size)
if args.dropout_words > 0:
input = dropout_overtime(input, args.dropout_words,
self.training) #
rnnoutputs['trans_outlayer-1'] = input
for i in range(self.num_blocks):
rnnoutputs['dense_outlayer%d' %
i], hidden_x, hid_ind_layer = self.DenseBlocks[i](
rnnoutputs['trans_outlayer%d' % (i - 1)], BN_start,
hidden_x, hid_ind_layer)
if i != self.num_blocks - 1:
rnnoutputs['trans_outlayer%d' %
i], hidden_x, hid_ind_layer = self.TransBlocks[i](
rnnoutputs['dense_outlayer%d' % i], BN_start,
hidden_x, hid_ind_layer)
rnnoutputs['trans_outlayer%d' %
(self.num_blocks -
1)], hidden_x, hid_ind_layer = self.TransBlocks[
self.num_blocks - 1](rnnoutputs['dense_outlayer%d' %
(self.num_blocks - 1)],
BN_start, hidden_x, hid_ind_layer)
rnn_out = rnnoutputs['trans_outlayer%d' % (self.num_blocks - 1)]
rnn_out = rnn_out.view(seq_len * batch_size, args.embed_size)
rnn_out = self.decoder(rnn_out)
return rnn_out, hidden_x
def forward(self, input,hidden_input_all, BN_start=0):
seq_len, batch_size = input.size()
rnnoutputs={}
hidden_x={}
if hidden_input_all is not None:
hidden_x.update(hidden_input_all)
else:
for x in range(2*len(self.resblocks)):
hidden_x['hidden%d'%x]=Variable(torch.zeros(1,batch_size,args.hidden_size).cuda())
hidden_x['hidden_resfinal']=Variable(torch.zeros(1,batch_size,args.hidden_size).cuda())
hidden_x['hidden_lastindrnn']=Variable(torch.zeros(1,batch_size,args.embed_size).cuda())
input = input.view(seq_len * batch_size)
input = embedded_dropout(self.encoder,input, dropout=args.dropout_embedding if self.training else 0)
input = input.view(seq_len, batch_size, args.embed_size)
if args.dropout_words > 0:
input = dropout_overtime(input, args.dropout_words, self.training)#
rnnoutputs['outlayer-1']=self.fc0(input)
for x in range(len(self.resblocks)):
rnnoutputs['outlayer%d'%x],hidden_x['hidden%d'%(2*x)],hidden_x['hidden%d'%(2*x+1)]= self.resblocks[x](rnnoutputs['outlayer%d'%(x-1)], BN_start, hidden_x['hidden%d'%(2*x)], hidden_x['hidden%d'%(2*x+1)])
rnn_out=rnnoutputs['outlayer%d'%(len(self.resblocks)-1)]
rnn_out,hidden_x['hidden_resfinal']=self.IndRNNwithBN_resfinal(rnn_out, BN_start, hidden_x['hidden_resfinal'])
if args.dropout_last>0:
rnn_out=dropout_overtime(rnn_out,args.dropout_last, self.training)
rnn_out=self.last_fc(rnn_out)
rnn_out1, hidden_x['hidden_lastindrnn']=self.last_indrnn(rnn_out, BN_start, hidden_x['hidden_lastindrnn'])
if args.bn_location=='bn_before':
rnn_out1=self.extra_bn(rnn_out1, BN_start)
if args.dropout_extrafc>0:
rnn_out1=dropout_overtime(rnn_out1,args.dropout_extrafc, self.training)
rnn_out1 = rnn_out1.view(seq_len * batch_size, args.embed_size)
rnn_out1=self.decoder(rnn_out1)
return rnn_out1, hidden_x