can you show me classification tutorial about SMCL

[zq1314]

@mbbrodie

[zq1314]

@mbbrodie

[mbbrodie]

Sorry for the delay, @zq1314 (out of town). I abandoned this repo nearly a year ago and shifted the topic of my dissertation. Due to the already ill-posed nature of GAN tasks, combining GANS with MCL did not produce promising results. If you're still interested in the area, however, I recommend looking into more recent MCL work, such as this recent 2018 CVPR paper: http://openaccess.thecvf.com/content_cvpr_2018/papers/Firman_DiverseNet_When_One_CVPR_2018_paper.pdf

One of the biggest hindrances to the wider adoption and acceptance of MCL is the need for multiple, expensive-to-train models. DiverseNet and some previous methods (such as TreeNets), provide more efficient methods that allow for more stable training (and require fewer computational resources. Sorry I can't provide a more detailed tutorial. If you're interested in using Caffe for your research, Stefen Lee has a CIFAR-10 MCL classification example here: https://github.com/steflee/MCL_Caffe/tree/master/examples/cifar10 (check out cifar10_quick_train_test.prototxt for the network definition).

Hope that helps and good luck.

[zq1314]

@mbbrodie thank you for your help

[zq1314]

Sorry to bother you, but their papers have no public code. @mbbrodie ，I feel helpless.

[mbbrodie]

No worries. Looks like Stefan removed the CIFAR-10 example and replaced it with an MNIST LeNet example. Check out the files that begin with 'mcl_' in this directory: https://github.com/steflee/MCL_Caffe/tree/master/examples/mnist

MNIST is a nicer starting point anyway (smaller network and easier to read through the files/see how the pieces work). If you're still struggling after that, it's not hard to reproduce the MCL layer in Python (run k models, identify the the one with the smallest loss, and only update that network). But this MNIST example will hopefully work for you. Good luck

[zq1314]

when i see the (mcl_lenet_train_test.prototxt) file,I found that it defines three identical models. but i think it should train three different models. @mbbrodie

[mbbrodie]

That architecture in your diagram is correct. All layers with 1* correspond to the first network, 2 is the second network, and 3_ is the third.

Because they have different names, Caffe treats them as separate layers/networks. At the end of mcl_lenet_train_test.prototxt, you'll see that 1_prob, 2_prob, and 3_prob feed into the MCLMultinomialLogisticLoss layer.

layer { name: "loss" type: "MCLMultinomialLogisticLoss" bottom: "1_prob" bottom: "2_prob" bottom: "3_prob" bottom: "label" top: "multiple-output loss" include { phase: TRAIN } }

That loss layer contains the code for the sMCL algorithm. If you're interested in creating a new loss layer, you can imitate the code in that file.

As a heads up, PyCaffe, PyTorch, or Tensorflow are much easier to code/debug/experiment with new MCL layers (don't have recompile / deal with header files). However, depending on your research goals, Stefan's code might be just fine.

[zq1314]

hi @mbbrodie now ,i use https://github.com/chhwang/cmcl code Successfully trained my training set, I want to see the process of error prediction accuracy in tensorboard, but it does not give the training process log in the program. I don't understand TENSORFLOW very well, can you help me look at that code, How to output the training process log?

[mbbrodie]

If you don't understand Tensorflow, go read their online documentation. They have great up-to-date tutorials for beginners, including some on Tensorboard. If you're still struggling, post a question on Stack Overflow or another forum site.