yifanlu0227
commented
4 months ago We use spconv 1.2.1 for training, and the checkpoint file is incompatible with spconv 2.x.
And this issue also comments on some incompatibility with spconv 2.x training, so using spconv 1.2.1 is suggested.
在尝试加载
Late_Fusion.zip中的checkpoint时遇到了问题。错误表明HeterModelLate的encoder_m3.spconv_block.conv_input.0.weight层的权重形状与checkpoint中的不匹配。错误信息如下： RuntimeError: Error(s) in loading state_dict for HeterModelLate: size mismatch for encoder_m3.spconv_block.conv_input.0.weight: copying a param with shape torch.Size([3, 3, 3, 64, 16]) from checkpoint, the shape in current model is torch.Size([16, 3, 3, 3, 64]).
在loading checkpoint的时候也会报错： ------ Loading Checkpoint ------ Missing keys from ckpt:
Extra keys from ckpt: encoder_m4
encoder_m4.camencode.up1.conv.4.bias encoder_m4.camencode.up1.conv.1.running_var encoder_m4.camencode.up1.conv.4.num_batches_tracked encoder_m4.camencode.up1.conv.1.num_batches_tracked encoder_m4.camencode.up1.conv.0.weight encoder_m4.camencode.up2.conv.4.bias encoder_m4.camencode.up2.conv.4.running_var encoder_m4.camencode.up2.conv.1.num_batches_tracked encoder_m4.camencode.up2.conv.4.num_batches_tracked encoder_m4.camencode.up1.conv.4.running_var encoder_m4.camencode.up1.conv.1.bias encoder_m4.camencode.up1.conv.1.weight encoder_m4.camencode.up2.conv.1.weight encoder_m4.camencode.up2.conv.4.weight encoder_m4.camencode.up2.conv.1.bias encoder_m4.camencode.up2.conv.1.running_mean encoder_m4.camencode.up2.conv.4.running_mean encoder_m4.camencode.up1.conv.1.running_mean encoder_m4.camencode.up2.conv.1.running_var encoder_m4.camencode.up2.conv.0.weight encoder_m4.camencode.up1.conv.4.weight encoder_m4.camencode.up1.conv.4.running_mean encoder_m4.camencode.up2.conv.3.weight encoder_m4.camencode.up1.conv.3.weight