demo

[LXWDL]

您好，我看论文里面说三值网络可以实现16或32倍的压缩，但我跑了lenet在mnist数据集上的demo：
LOG= ./build/tools/caffe train --gpu=0 --precision=ternary --delta=7 --solver=./examples/mnist/lenet_tn_solver.prototxt --debug=no 2>&1 | tee $LOG
训练的结果模型为：lenet_tn_iter_30000.caffemodel 大小为：2.3 MB (2,339,071 字节)
并没有实现压缩，模型好像也变大了，由原来的1.7M变成了现在的2.3M

[fengfu-chris]

存储三值权重数据请将snapshot_ternary设置为true

[sainisanjay]

1. Thank you @fengfu-chris for your clarification. Actually same doubt i was having. May i know where we have to set snapshot_ternary true.
2. Secondly, i tested your network on MNIST dataset with PRECISION=ternary & binary & single but there is no difference in computation time. All three PRECISION taking same "Elapsed time from previous test: 23.372 seconds.". Can you explain this a bit why its like that.

[fengfu-chris]

Hi, @sainisanjay. 1. In the solver.prototxt file. 2. Here we only consider the performance comparison. The speed issue is not considered, yet.

[sainisanjay]

Dear @fengfu-chris , Thank you so much for your kind reply and clarification. However, as we are using binary/ternary weights hence we are eliminating multiplication operation in convolution and we know that multiplication operations are time consuming. Thus i feel that atleast network should show some affect on computation time however your network does not have any computation time difference in ternary & binary & single. Any particular reason for this????