深度学习-Caffe 使用问题整理
1. shape mismatch in weight initialisation (transfer). "Target shape is ..., but source shape is ...".
possible solution: (1) layer name is wrong (2) 'Eltwise' or 'Concat' fusion manner, which often tends to be 2-fold.
2. training process crashes: (1) problem about training samples (a) training sample is not in proper size.
3. NVIDIA TITAN X PASCAL GPU needsCUDA 8.0 for successful compilation of caffe.
4. Install Protobuf for caffe python interface. I choose to use Anaconda for python, and we need to install protobuf based on Anaconda, so that Anaconda can find the protobuf. There are proxy problems with the Anaconda 2 version 4.2.0, while version 4.0.0 is ok in proxy. Install protobuf with "conda". The command is here: https://anaconda.org/anaconda/protobuf. The github issue is here: https://github.com/conda/conda/issues/3984