加载训练后的模型时modulated.weight权重参数不对应；The weight parameter modulated.weight does not correspond when loading the trained model #434

New Issue

claunia · 2026-01-29T21:47:42Z

claunia commented

2026-01-29 21:47:42 +00:00

Originally created by @liulipeng0538 on GitHub (Jan 26, 2024).

作者团队你们好：我编写简单的训练模型通过 torch.save(module.state_dict()) 保存模型后进行加载发现 stylegan_decoder.style_conv1.modulated_conv.weight等网络层的权重参数与保存参数存在差异，定位是只要涉及modulated_conv.weight都有差异，看源代码中的self.weight = nn.Parameter(
torch.rand(1, out_channels, in_channels, kernel_size, kernel_size) /math.sqrt(in_channels * kernel_size**2))
请问是什么原因导致的呢？多谢

Originally created by @liulipeng0538 on GitHub (Jan 26, 2024). 作者团队你们好：我编写简单的训练模型通过 torch.save(module.state_dict()) 保存模型后进行加载发现 stylegan_decoder.style_conv1.modulated_conv.weight等网络层的权重参数与保存参数存在差异，定位是只要涉及modulated_conv.weight都有差异，看源代码中的self.weight = nn.Parameter( torch.rand(1, out_channels, in_channels, kernel_size, kernel_size) /math.sqrt(in_channels * kernel_size**2)) 请问是什么原因导致的呢？多谢

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: TencentARC/GFPGAN#434