Adding Loss Output to All Layers in BVLC Reference CaffeNet

[Joe Marino]

http://caffe.berkeleyvision.org/tutorial/loss.html states that any layer can contribute to the overall loss by adding the field "loss_weight: 1" to the layer definition. However, when I try adding this field to the layers in the train_val.prototxt file for the BVLC Reference CaffeNet, the network initializes layers such as conv1_conv1_0_split (which is does not initialize normally) and finally fails with the error "Duplicate blobs produced by multiple sources." How is the overall loss computed? How do I make each layer of the network contribute to the loss? Is there a better way to accomplish this without using the loss_weight parameter?

Joe

[lql2...@aliyun.com]

您好！我在使用caffe训练出了MNIST 模型但是不知道怎么使用这个模型去识别现实图片中的照片，希望能够得到您的指点。email address：lql2...@aliyun.com.期待您的回复
在 2014年10月29日星期三UTC+8上午6时34分23秒，Joe Marino写道：