def decode(self, enc):
enc = enc.repeat(self.out_length, 1, 1)
h_dec, _ = self.dec_lstm(enc)
h_dec = h_dec.permute(1, 0, 2)
fut_pred = self.op(h_dec)
fut_pred = fut_pred.permute(1, 0, 2)
fut_pred = outputActivation(fut_pred)
return fut_pred
def decode(self, enc): if self.use_attention: # Attention builds on top of seq2seq m = enc.size(0) if self.use_cuda: h0 = torch.zeros(self.num_layers, m, self.decoder_size).cuda() # [SeqLen, Batch, decoder size] else: h0 = torch.zeros(self.num_layers, m, self.decoder_size) # [1, 128, 128] c0 = y0 = h0 enc = torch.cat((enc, self.hist_enc_back, self.grid_enc_back), 1) enc = self.proj_seq2seq(enc) # proj from [Batch, 117] to [Batch, 128] yn = enc.unsqueeze(0) # [1, Batch 128, decoder size 128] context = self.one_step_attention(self.hist_a, h0).unsqueeze(0) yn, (hn, cn) = self.dec_seq2seq(context, (h0, c0)) h_dec = yn for t in range(self.out_length - 1): context = self.one_step_attention(self.hist_a, hn).unsqueeze(0) yn, (hn, cn) = self.dec_seq2seq(context, (hn, cn)) h_dec = torch.cat((h_dec, yn), dim=0) elif self.use_seq2seq: m = enc.size(0) if self.use_cuda: self.h0 = torch.zeros(self.num_layers, m, self.decoder_size).cuda() # [SeqLen, Batch, decoder size] else: self.h0 = torch.zeros(self.num_layers, m, self.decoder_size) # [1, 128, 128] self.c0 = self.y0 = self.h0 #pdb.set_trace() enc = torch.cat((enc, self.hist_enc_back, self.grid_enc_back), 1) enc = self.proj_seq2seq(enc) # proj from [Batch, 117] to [Batch, 128] encout = enc.unsqueeze(0) # [1, Batch 128, decoder size 128] #yn, (hn, cn) = self.dec_seq2seq(yn, (self.h0, self.c0)) yn, (hn, cn) = self.dec_seq2seq(encout, (self.h0, self.c0)) #yn, (hn, cn) = self.dec_seq2seq(self.y0, (encout, encout)) h_dec = yn for t in range(self.out_length - 1): yn, (hn, cn) = self.dec_seq2seq(yn, (hn, cn)) h_dec = torch.cat((h_dec, yn), dim=0) else: # enc: [batch 128, feats 117] # we just repeat hn output, not using a_1 up to a_Tx (TODO via ATTENTION) # enc: [5sec 25, batch 128, feats 117] after repeat enc = enc.repeat(self.out_length, 1, 1) # And now we retrieve the T_y=25 outputs and discard (hn, cn) # h_dec: [25, 128, 128] via nn.LSTM(feats 117, hid 128) h_dec, _ = self.dec_lstm(enc) # h_dec: [batch 128, Ty 25, hid 128] via permute(1, 0, 2) h_dec = h_dec.permute(1, 0, 2) # fut_pred: [batch 128, Ty 25, bivariate gaussian params 5] via self.op=nn.Linear(128, 5) fut_pred = self.op(h_dec) # fut_pred: [Ty 25, batch 128, 5] via permute fut_pred = fut_pred.permute(1, 0, 2) fut_pred = outputActivation(fut_pred) # fut_pred: [Ty 25, batch 128, bivariate gaussian params 5] via outputActivation which enforces pred constraints return fut_pred
def decode(self,enc):
#print (enc.size()) # (128, 117)
enc = enc.repeat(self.out_length, 1, 1)
#print (enc.size()) # (25, 128, 117), why 25? because we want to predict future 50 timesteps histories and the downsampling rate is 2
h_dec, _ = self.dec_lstm(enc) # (25, 128, 128)
h_dec = h_dec.permute(1, 0, 2) # (128, 25, 128)
fut_pred = self.op(h_dec) # (128, 25, 5)
fut_pred = fut_pred.permute(1, 0, 2) # (25, 128, 5) 25 timesteps, 128 batchsize, 5 prediction
fut_pred = outputActivation(fut_pred)
return fut_pred
def decode_by_step(self, enc):
pre_traj = []
decoder_input = enc
for _ in range(self.out_length):
decoder_input = decoder_input.unsqueeze(0)
h_dec, _ = self.dec_lstm(decoder_input)
h_for_pred = h_dec.squeeze()
fut_pred = self.op(h_for_pred)
pre_traj.append(fut_pred.view(fut_pred.size()[0], -1))
embedding_input = fut_pred
decoder_input = self.spatial_embedding(embedding_input)
pre_traj = torch.stack(pre_traj, dim=0)
pre_traj = outputActivation(pre_traj)
return pre_traj