caffe,tensorflow,theano,pytorch 输入数据形状

caffe :

nsample * channels * height * width

labels start from 0, such as for 5 classes, [0,1,2,3,4]; channels: BGR
tensorflow:

nsample * height * width * channels

labels start from 0, such as for 5 classes, [0,1,2,3,4]; channels: BGR
theano:

nsample * channels * height * width 

labels start from 0, such as 5 classes ,[0,1,2,3,4];

pytorch:

nsample * channels * height * width 

labels: nsample * 1, have the same channel with Caffe: BGR

参考文献：
1.https://discuss.pytorch.org/t/convert-caffe-to-pytorch/10261