search

fmassa / object-detection.torch

BSD 2-Clause "Simplified" License
116 stars 40 forks source link

How did you create the "frcnn_alexnet.t7" file? #17

Closed liuchang8am closed 8 years ago

liuchang8am commented 8 years ago

I want to create a VGG16 network t7 file from pre-trained VGG16.caffemodel, how to create it, like the frcnn_alexnet.t7 you provided?

As you mention in #12 , use "loadcaffe" to load the pre-trained caffelmodel, yet loadcaffe does not support the ROIpooling layer in Fast-RCNN (-- warning: module 'roi_pool5 [type ROIPooling]' not found in generated .prototxt.lua file), how did you overcome this?

Thanks!

szagoruyko commented 8 years ago

converted models are here https://github.com/szagoruyko/fastrcnn-models.torch

liuchang8am commented 8 years ago

@szagoruyko thanks for the quick reply!

Actually, I still need to create the '.t7' model file myself, because I trained the VGG16 caffemodel task-specific. Could you please tell me how to create the file?

szagoruyko commented 8 years ago

loadcaffe will give you a sequential model that you can manually reorder too be as the ones in the link above. you will have to add inn.ROIPooling, check here https://github.com/szagoruyko/imagine-nn

liuchang8am commented 8 years ago

thanks! that's what I need!