visionfans的专栏

Quoting Lecun’s post: In Convolutional Nets, there is no such thing as “fully-connected layers”. There are only convolution layers with 1x1 convolution kernels and a full connection table. It’s a

Release original memory and hold new:template <typename Dtype>void Blob<Dtype>::Reshape(const vector<int>& shape) { CHECK_LE(shape.size(), kMaxBlobAxes); count_ = 1; shape_.resize(shape.size())

to finish.ReferencesDeveloping new layersSwitch LayerType to string for extensibility #1685Layer type is a string #1694

coeff should be empty or be the same size with bottomcoeff only be used with SUM operationtop and all bottoms are the same sizeoperations top[0] = PROD(bottom[:])top[0] = SUM(coeff[i]*bottom[i])

Implement the model that won the classification task of ImageNet 2013 #33choosing batch sizes and tuning sgd #218How to train imagenet with reduced memory and batch size? #430 Originally base_lr = 0

dropout in place incompatible with max pooling #117Fix Max pooling layer to use a mask #162Finish up max pooling with a mask from @sguada #448In-place computation can break gradient computation #201

看Caffe/wiki/Development，同时对照ArgMaxLayer来看！ layer的声明要写到：l  common_layers.hppl  data_layers.hppl  loss_layers.hppl  neuron_layers.hppl  vision_layers.hpp 要实现以下函数：l  LayerSetUpl  Re

by Yangqing on 14 May 2014 For a sanity check, try running with a learning rate 0 to see if any nan errors pop up (they shouldn’t, since no learning takes place). If data is not initialized well, it