未重参数化的权重

[z-0508]

作者您好，感谢您的有趣的创意和开源工作！我们希望在您的工作上修改做下游任务，但是遇到了一些问题。

在ReadMe中已经提到您开源了重参数化后的权重，其中包含：

down_level1_block1.conv1.weight
down_level1_block1.conv1.bias
down_level1_block1.conv2.weight
down_level1_block1.conv2.bias
......

等层的权重。但是我想要在您预训练的基础上进行调整与训练，backbone_train.py下的DEANet实例化后的模型包含下列层的权重：

Layer name: down_level1_block1.conv1.conv1_1.conv.weight, Shape: torch.Size([32, 32, 3, 3])
Layer name: down_level1_block1.conv1.conv1_1.conv.bias, Shape: torch.Size([32])
Layer name: down_level1_block1.conv1.conv1_2.conv.weight, Shape: torch.Size([32, 32, 3])
Layer name: down_level1_block1.conv1.conv1_2.conv.bias, Shape: torch.Size([32])
Layer name: down_level1_block1.conv1.conv1_3.conv.weight, Shape: torch.Size([32, 32, 3])
Layer name: down_level1_block1.conv1.conv1_3.conv.bias, Shape: torch.Size([32])
Layer name: down_level1_block1.conv1.conv1_4.conv.weight, Shape: torch.Size([32, 32, 3, 3])
Layer name: down_level1_block1.conv1.conv1_4.conv.bias, Shape: torch.Size([32])
Layer name: down_level1_block1.conv1.conv1_5.weight, Shape: torch.Size([32, 32, 3, 3])
Layer name: down_level1_block1.conv1.conv1_5.bias, Shape: torch.Size([32])
Layer name: down_level1_block1.conv2.weight, Shape: torch.Size([32, 32, 3, 3])
Layer name: down_level1_block1.conv2.bias, Shape: torch.Size([32])
Layer name: down_level1_block2.conv1.conv1_1.conv.weight, Shape: torch.Size([32, 32, 3, 3])
Layer name: down_level1_block2.conv1.conv1_1.conv.bias, Shape: torch.Size([32])
Layer name: down_level1_block2.conv1.conv1_2.conv.weight, Shape: torch.Size([32, 32, 3])
Layer name: down_level1_block2.conv1.conv1_2.conv.bias, Shape: torch.Size([32])
Layer name: down_level1_block2.conv1.conv1_3.conv.weight, Shape: torch.Size([32, 32, 3])
Layer name: down_level1_block2.conv1.conv1_3.conv.bias, Shape: torch.Size([32])
Layer name: down_level1_block2.conv1.conv1_4.conv.weight, Shape: torch.Size([32, 32, 3, 3])
Layer name: down_level1_block2.conv1.conv1_4.conv.bias, Shape: torch.Size([32])
Layer name: down_level1_block2.conv1.conv1_5.weight, Shape: torch.Size([32, 32, 3, 3])
Layer name: down_level1_block2.conv1.conv1_5.bias, Shape: torch.Size([32])
Layer name: down_level1_block2.conv2.weight, Shape: torch.Size([32, 32, 3, 3])
Layer name: down_level1_block2.conv2.bias, Shape: torch.Size([32])
......

您目前开源的预训练权重不包含未重参数化的预训练权重，我们希望获得未执行reparam.py的权重，从而能在我们自己的数据集上做改进与微调。如果能提供OTS数据集下的未重参数化的预训练权重，我们将不胜感激！