lijianfei06 / mnasnet-caffe Goto Github PK

您好，我看了下您的网络结构好像没有SE模块，我自己加了SE模块训练，网络严重过拟合，请问您有试过加入SE模块吗？

请问能否提供该模型的solver.prototxt文件， 使用已有经验的base_lr， weight decay， lr_policy等参数，便于训练模型？

在cifar10数据上做了个小测验， base_lr=0.1训练60000轮后，测试精度只有58.74%， 日志显示train loss过大，应该上训练未充分， base_lr=0.0001训练60000轮后，只有28.17%的精度。
`
I1013 11:38:29.757405 3189 caffe.cpp:330] Softmax1 = 1.43176 (* 1 = 1.43176 loss)
I1013 11:38:29.757411 3189 caffe.cpp:330] accuracy = 0.5842

.... base_lr=0.001
I1013 16:11:04.564793 4153 solver.cpp:447] Snapshotting to binary proto file examples/cifar10/cifar10_MnasNet_iter_60000.caffemodel
I1013 16:11:04.612725 4153 sgd_solver.cpp:273] Snapshotting solver state to binary proto file examples/cifar10/cifar10_MnasNet_iter_60000.solverstate
I1013 16:11:04.629349 4153 solver.cpp:330] Iteration 60000, Testing net (#0)
I1013 16:11:09.132675 4173 data_layer.cpp:73] Restarting data prefetching from start.
I1013 16:11:09.322088 4153 solver.cpp:397] Test net output #0: Softmax1 = 2.14012 (* 1 = 2.14012 loss)
I1013 16:11:09.322113 4153 solver.cpp:397] Test net output #1: accuracy = 0.2817

`
caffe time命令输出MnasNet的gpu耗时在14ms以上，

`
./build/tools/caffe time -iterations=100 -gpu=0 -model=examples/cifar10/train_MnasNet.prototxt
....

I1013 12:00:47.575592 3703 caffe.cpp:409] Pooling1 forward: 0.0136704 ms.
I1013 12:00:47.575600 3703 caffe.cpp:412] Pooling1 backward: 0.014039 ms.
I1013 12:00:47.575608 3703 caffe.cpp:409] fc1 forward: 0.038953 ms.
I1013 12:00:47.575616 3703 caffe.cpp:412] fc1 backward: 0.0232038 ms.
I1013 12:00:47.575623 3703 caffe.cpp:409] Softmax1 forward: 0.101257 ms.
I1013 12:00:47.575630 3703 caffe.cpp:412] Softmax1 backward: 0.0176986 ms.
I1013 12:00:47.575664 3703 caffe.cpp:417] Average Forward pass: 14.2767 ms.
I1013 12:00:47.575672 3703 caffe.cpp:419] Average Backward pass: 36.3857 ms.
I1013 12:00:47.575690 3703 caffe.cpp:421] Average Forward-Backward: 50.905 ms.
I1013 12:00:47.575698 3703 caffe.cpp:423] Total Time: 5090.5 ms.
I1013 12:00:47.575704 3703 caffe.cpp:424] *** Benchmark ends ***

`
这个性能比darknet-19差多了， 大神能否提供一些测试的性能数据作为参考。

训练用的solver文件
`

net: "examples/cifar10/train_MnasNet.prototxt"
test_iter: 100
test_interval: 1000
base_lr: 0.0001
momentum: 0.9
weight_decay: 0.005
lr_policy: "step"
gamma: 1
stepsize: 5000
display: 100
max_iter: 160000
snapshot: 10000
snapshot_prefix: "examples/cifar10/cifar10_MnasNet"
solver_mode: GPU
`
@LiJianfei06

你好，我用自己的猫狗数据来做训练，发现精度一直上不去。 而且训练速度很慢。

我没有自己写depthwise conv， 只用了caffe提供的group 操作。

另外我的batch size 设置到了128

请问这个问题应该则怎么解决