镜像自地址
https://github.com/MINT-SJTU/Evo-RL.git
已同步 2026-11-20 18:17:06 +00:00
fix(evo1): use batch_size consistently
这个提交包含在:
@ -289,7 +289,7 @@ class FlowmatchingActionHead(nn.Module):
|
||||
x = self.norm_out(x)
|
||||
|
||||
if self.horizon > 1:
|
||||
x_flat = x.reshape(B, -1)
|
||||
x_flat = x.reshape(batch_size, -1)
|
||||
|
||||
if not hasattr(self, "seq_pool_proj"):
|
||||
self.seq_pool_proj = nn.Linear(self.horizon * self.embed_dim, self.embed_dim).to(device)
|
||||
@ -370,7 +370,7 @@ class FlowmatchingActionHead(nn.Module):
|
||||
x = self.norm_out(x)
|
||||
|
||||
if self.horizon > 1:
|
||||
x_flat = x.reshape(B, -1)
|
||||
x_flat = x.reshape(batch_size, -1)
|
||||
if hasattr(self, "seq_pool_proj"):
|
||||
x_pooled = self.seq_pool_proj(x_flat)
|
||||
else:
|
||||
|
||||
在新工单中引用
屏蔽一个用户