Open jackyko1991 opened 6 years ago
It's on my todo list. It's a low priority for me though.
Hey @davisking, any update on this particular issue? Thanks.
It's still on my todo list. However, it's not very hard to do, so anyone else could submit a PR ;)
Hey @davisking, any update on this particular issue? MayBe i can do a PR, could you please give me a roadmap to achieving it? Thanks.
I'm not going to work on this any time soon.
@davisking I'm working on adding this to my project. It appears to be simple enough to modify (or create another) tt::tensor_conv
and add a con3d_
layer for 3D filters. Unfortunately, 4D filters (which are required for RGB sequences) require 5D tensors as both input and the filter stack.
Are you open to expanding the dimensions of the tensor
class to 5, given it maintains a consistent interface and behavior for the existing usages?
Changing the tensor
class to have 5 dimensions is a huge change to everything. You should be able to do this without doing that though by just flattening everything along the k
dimension and making the 3D conv code smart about doing the right thing. Then no sweeping architectural changes are needed to the rest of the code in dlib.
I am currently looking for CNN training tools for windows with c++. dlib seems to be a good choice as the example clearly illustrate how ResNet is implemented.
However, I am looking for applications for 3D image segmentation which requires 3D pooling and convolution (somehow it is called 4D convolution in this link
This function is supported in caffe2, tensorflow and pytorch. Any plan to implement for a higher dimension CNN operations?