关于自己训练集微调

[dsm233]

作者你好，我想用你训练好的模型在我的数据集上微调，是不是需要在配置文件的path那边修改设置pretrain_network_g？

[dsm233]

还有请问作者有没有small的测试代码，我理解samll是有两个baseblock，b有四个，large有八个？

[INVOKERer]

path需要修改设置的。small的设置应该是你理解的yml中的baseblock的list。具体的目前我身边目前没有电脑。

[dsm233]

那network_g下面的state_dict_pth_encoder和state_dict_pth_decoder不需要设置了是吧，还有学习率之类的需要调整吗

[dsm233]

用预训练的参数效果已经不错了，我想看看微调后效果是否能更好一点，但我训练后发现效果很烂还没找到原因

[INVOKERer]

建议先确认下预训练模型有没有被正确的加载。

[INVOKERer]

按照你的说法可以不需要再修改network g里的模型路径

[dsm233]

您的训练代码里如果不需要resume_state，是不是没有加载pretrain_network_g这一过程？需要我手动加载？

[INVOKERer]

嗯，如果是按照我的方式训练decoder不需要加载

[dsm233]

2025-12-01 16:12:09,184 WARNING: Params module.subdenets.0.alpha0 will not be optimized. 2025-12-01 16:12:09,184 WARNING: Params module.subdenets.0.alpha1 will not be optimized. 2025-12-01 16:12:09,185 WARNING: Params module.subdenets.0.alpha2 will not be optimized. 2025-12-01 16:12:09,185 WARNING: Params module.subdenets.0.alpha3 will not be optimized. 2025-12-01 16:12:09,185 WARNING: Params module.subdenets.0.level0.fusion.down.0.weight will not be optimized. 2025-12-01 16:12:09,185 WARNING: Params module.subdenets.0.level0.fusion.down.0.bias will not be optimized. 2025-12-01 16:12:09,185 WARNING: Params module.subdenets.0.level0.fusion.down.1.body.weight will not be optimized. 2025-12-01 16:12:09,185 WARNING: Params module.subdenets.0.level0.fusion.down.1.body.bias will not be optimized. 2025-12-01 16:12:09,185 WARNING: Params module.subdenets.0.level0.fusion.up.0.weight will not be optimized. 2025-12-01 16:12:09,186 WARNING: Params module.subdenets.0.level0.blocks.0.gamma will not be optimized. 2025-12-01 16:12:09,186 WARNING: Params module.subdenets.0.level0.blocks.0.beta will not be optimized. 2025-12-01 16:12:09,186 WARNING: Params module.subdenets.0.level0.blocks.0.conv1.weight will not be optimized. 2025-12-01 16:12:09,186 WARNING: Params module.subdenets.0.level0.blocks.0.conv1.bias will not be optimized. 2025-12-01 16:12:09,186 WARNING: Params module.subdenets.0.level0.blocks.0.conv2.weight will not be optimized. 2025-12-01 16:12:09,186 WARNING: Params module.subdenets.0.level0.blocks.0.conv2.bias will not be optimized. 2025-12-01 16:12:09,186 WARNING: Params module.subdenets.0.level0.blocks.0.conv3.weight will not be optimized. 2025-12-01 16:12:09,187 WARNING: Params module.subdenets.0.level0.blocks.0.conv3.bias will not be optimized. 2025-12-01 16:12:09,187 WARNING: Params module.subdenets.0.level0.blocks.0.sca.0.weight will not be optimized. 2025-12-01 16:12:09,187 WARNING: Params module.subdenets.0.level0.blocks.0.sca.0.bias will not be optimized. 2025-12-01 16:12:09,187 WARNING: Params module.subdenets.0.level0.blocks.0.norm1.weight will not be optimized. 2025-12-01 16:12:09,187 WARNING: Params module.subdenets.0.level0.blocks.0.norm1.bias will not be optimized. 2025-12-01 16:12:09,187 WARNING: Params module.subdenets.0.level0.blocks.0.norm2.weight will not be optimized. 2025-12-01 16:12:09,187 WARNING: Params module.subdenets.0.level0.blocks.0.norm2.bias will not be optimized. 2025-12-01 16:12:09,187 WARNING: Params module.subdenets.0.level0.blocks.0.fft_block1.complex_conv1.weight will not be optimized. 2025-12-01 16:12:09,188 WARNING: Params module.subdenets.0.level0.blocks.0.fft_block1.complex_conv1.bias will not be optimized.
为啥我的subdenets的参数也不会更新呀，这是哪里有问题

[INVOKERer]

你是不是用了我训练classifier的yml

[dsm233]

奥找到问题了，我把network_g:下面的pretrain设置为了false，导致无法更新