moskewcz
commented
8 years ago To clarify:
- what LeNet example are you talking about? there's no LeNet example in the boda distribution that i am aware of.
- what do you mean 'run'? if you mean something like testing, it probably requires writing a testbench/wrapper specific to the particular task, although some of the existing modes for image classification (i.e. capture_classify, test_lmdb) might be pretty close to what is needed. specific issues with these modes are welcome, including enhancement requests, but no promises.
Also, there is no BodaFullConvNets directory in the boda distribution either, so it's confusing to ask about that here without explanation of what it is. But, assuming you mean the set of pre-trained models i provided to you over email that match the net definitions in the boda/nets directory:
- what do you mean by 'build files'? If you're asking how to recreate/acquire such files from scratch, this is not currently well documented. They come from various sources. Some of them i trained myself, and others are just stock pre-trained models. Many of them have been modified using boda or caffe to convert InnerProduct layers to Convolution layers. In general, acquisition of pre-trained models and/or training of such nets using caffe or other methods is outside the scope of boda (at least currently).
However, boda does nominally aim to provide a working InnerProduct->Convolution layer conversion utility, cnet_fc_to_conv: https://github.com/moskewcz/boda/blob/master/src/caffepb.cc#L842
So, if you have issues with that mode (which is likely since it is used rarely, not documented beyond it's internal help message, not too well tested, and not too robust), please submit it. I am interested in increasing the quality of such modes over the long term.
Can I run the LeNet example in boda? Then what the cli_str look like and how can I build files in the BodaFullyConvNets directory?