wassname/seq2seq-time
1
0
Fork 0
mirror of https://github.com/wassname/seq2seq-time.git synced 2026-08-04 13:14:48 +08:00
Code Issues Packages Projects Releases Wiki Activity

hyperparam opt

Browse Source
This commit is contained in:
wassname
2020-11-13 15:56:16 +08:00
parent 1e69b00fee
commit 309833ed14
4 changed files with 2175 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
seq2seq_time/models/neural_process.py
+4 -2
View File
@@ -407,9 +407,11 @@ class RANP(nn.Module):
y = torch.cat([past_y, future_y], 1)
dist_post, log_var_post = self._latent_encoder(x, y)
if self.training:
z = dist_prior.rsample()
if self.training and (future_y is not None):
# USe posterior during training, is possible
z = dist_post.rsample()
else:
# During eval use the prior, also take the most probable
z = dist_prior.loc
num_targets = future_x.size(1)
z = z.unsqueeze(1).repeat(1, num_targets, 1) # [B, T_target, H]
seq2seq_time/models/transformer_process.py
+4 -3
View File
@@ -162,10 +162,11 @@ class TransformerProcess(nn.Module):
y = torch.cat([past_y, future_y], 1)
dist_post = self._latent_encoder(x, y)
if self.training:
# Sample from latent space during training
z = dist_prior.rsample()
if self.training and (future_y is not None):
# USe posterior during training, is possible
z = dist_post.rsample()
else:
# During eval use the prior, also take the most probable
z = dist_prior.loc
num_targets = future_x.size(1)
z = z.unsqueeze(1).repeat(1, num_targets, 1) # [B, S_target, H]