Inconsistent description between training and testing models

[Ayews]

Please note for users using training code that the weights obtained from retraining need to be passed ./mlsd/mlsd_pytorch/models/mbv2_mlsd_large.py is used to load and infer, rather than ./mlsd/models/mbv2_mlsd_large.py. The latter includes some special processing for the tf model, as well as differences in the input channel (the latter has an additional alpha channel), and the weights of the two are not compatible.

[Ayews]

The input is a 4-channel pre-training weight, and the values of the 4-th channel in the first layer convolution parameters are all 0. However, even if the fourth channel is deleted and converted to input 3 channels, the correct results cannot be obtained through ./mlsd/mlsd_pytorch/models/mbv2_mlsd_large.py, due to inconsistent model descriptions