When pytorch loads the model, if some judgment is used in the model, the judgment is used as the selection execution condition, but it is also saved in the model. However, when calling, the network in the judgment condition is not selected and load_state_Dict is used will report an error. Some operators cannot find the name. For example:

if backbone == "mobilenet":
    self.backbone = mobilenet()
    flat_shape = 1024
    elif backbone == "inception_resnetv1":
    self.backbone = inception_resnet()
else:
    raise ValueError('Unsupported backbone - `{}`, Use mobilenet, inception_resnetv1.'.format(backbone))
    self.avg = nn.AdaptiveAvgPool2d((1,1))
    self.Bottleneck = nn.Linear(flat_shape, embedding_size,bias=False)
    self.last_bn = nn.BatchNorm1d(embedding_size, eps=0.001, momentum=0.1, affine=True)
    if mode == "train": # Judgment condition, test without loading full connection
        self.classifier = nn.Linear(embedding_size, num_classes)

The strict = false option can be added to avoid operators not called in the network:

model2.load_state_dict(state_dict2, strict=False)