typo in VAE encoder's joint projection

这个提交包含在:
Tony Zhao
2023-06-23 13:30:39 -07:00
父节点 f74f0db292
当前提交 dfe6c7f5ff

查看文件

@ -88,7 +88,7 @@ class DETRVAE(nn.Module):
if is_training: if is_training:
# project action sequence to embedding dim, and concat with a CLS token # project action sequence to embedding dim, and concat with a CLS token
action_embed = self.encoder_action_proj(actions) # (bs, seq, hidden_dim) action_embed = self.encoder_action_proj(actions) # (bs, seq, hidden_dim)
qpos_embed = self.encoder_action_proj(qpos) # (bs, hidden_dim) qpos_embed = self.encoder_joint_proj(qpos) # (bs, hidden_dim)
qpos_embed = torch.unsqueeze(qpos_embed, axis=1) # (bs, 1, hidden_dim) qpos_embed = torch.unsqueeze(qpos_embed, axis=1) # (bs, 1, hidden_dim)
cls_embed = self.cls_embed.weight # (1, hidden_dim) cls_embed = self.cls_embed.weight # (1, hidden_dim)
cls_embed = torch.unsqueeze(cls_embed, axis=0).repeat(bs, 1, 1) # (bs, 1, hidden_dim) cls_embed = torch.unsqueeze(cls_embed, axis=0).repeat(bs, 1, 1) # (bs, 1, hidden_dim)