stage2似乎没加载上stage1的权重，state_dict里的key不一样了

[JachinLin2022]

def save_knowledge_adapter(self, f): if hasattr(self, "roberta"): encoder = self.roberta.encoder else: encoder = self.encoder dict = {} for i, layer in enumerate(encoder.layer): if layer.output.kas is not None: for j, adapter in enumerate(layer.output.kas): dict[str(i) + '-' + str(j) + '-adapter'] = adapter.state_dict() if layer.output.gating is not None: dict[str(i) + '-gating'] = layer.output.gating.state_dict() torch.save(dict, f)

def load_knowledge_adapter(self, f_list, one_f=None):
    if hasattr(self, "roberta"):
        encoder = self.roberta.encoder
    else:
        encoder = self.encoder
    for n_ka, f in enumerate(f_list):
        checkpoint = torch.load(f, map_location='cuda:0')
        for i, layer in enumerate(encoder.layer):
            if str(i) + '-adapter' in checkpoint:
                layer.output.kas[n_ka].load_state_dict(checkpoint[str(i) + '-adapter'])
            # if str(i) + '-attn' in checkpoint:
                # layer.output.attn.load_state_dict(checkpoint[str(i) + '-attn'])
            if str(i) + '-gating' in checkpoint and len(f_list) == 1 and layer.output.gating is not None:
                layer.output.gating.load_state_dict(checkpoint[str(i) + '-gating'])
    if one_f is not None:
        checkpoint = torch.load(one_f, map_location='cuda:0')
        for i, layer in enumerate(encoder.layer):
            for n_ka in range(self.ka_list):
                if str(i) + '-' + str(n_ka) + '-adapter' in checkpoint:
                    layer.output.kas[n_ka].load_state_dict(
                        checkpoint[str(i) + '-' + str(n_ka) + '-adapter'])
            if str(i) + '-gating' in checkpoint:
                layer.output.gating.load_state_dict(
                    checkpoint[str(i) + '-gating'])

[tianyang-x]

你好 @JachinLin2022 ,

多谢反馈。在我们早期的实验中，我们用的是i-adapter这一格式，对应的是load_knowledge_adapter函数。在后期，我们修改了save_knowledge_adapter函数的格式，并且对load_knowledge_adapter增加了one_f参数适应新的格式。我们实验时两种格式都用过，但是由于代码整理的问题，one_f参数没有被实际使用。我们会尽快修改这一问题。

[tianyang-x]

你好 @JachinLin2022 ,

刚才的commit应该已经修复了这一问题。