train with another dataset

[lapetite123]

I‘m interested in your great work, and I want to train with another dataset with 12 classes, I am confused at what I should change about training with differentnumber except for classed number and cmap? Can you please help me ?? thanks very much, waiting for your reply

[suhangpro]

To add a dataset properly, you will need to modify files under splatnet/semseg3d/ (or splatnet/partseg3d/ for part segmentation) accordingly. A quicker solution is to keep the original code but instead structure your data folder and files the same way as the supported datasets. Like discussed in #12 , you may still need to make some modifications to support a different set of data attribute fields.

[lapetite123]

for semantic segmentation, classification number in the last layer is not the same mine, how to change it , can you tell me in detail?

[lapetite123]

when I train with 8 classes, it reported" Cannot copy param 0 weights from layer 'conv7'; shape mismatch. Source param shape is 7 64 1 1 (448); target param shape is 8 64 1 1 (512). To learn this layer's parameters from scratch rather than copying from a saved net, rename the layer" what should I do?

[varunjampani]

Since we trained the network with 7 classes, the weights for the last layer is of shape 7x64. Since you want to train for 8 classes, the required weights would be of shape 8x64 and so you can not use the pre-trained weights of this last layer. You need to re-learn this layer parameters from scratch. As mentioned in your message, an easy way to do this would be re-naming the last classification layer in the caffe model to something else.

[lapetite123]

thanks a lot !