loss is negative when trianing the flow model using a pointcloud autoencoder

[L-Aidan]

I found the logdet becomes 200+ after the first batchnormFlow model, is this related with the input of the flow model(the output of the autoencoder)?

[L-Aidan]

I found the loss would be negative even when training the voxel with pretrained ckpt. Does anyone else meet this problem?

[happysxpp]

the same problem