camel007 / caffe-moon Goto Github PK

Hello, may I ask which version of caffe you are using? I encountered an error when I ran it. After investigation, I can confirm that it is a version problem.

Hi
I have downloaded the "custom_layers.hpp","normalize_layer.cpp" and "normalize_layer.cu" and put them in my caffe. Although I can successfully compile the code now, I can't run the test code and the error is "Creating layer norm_conv8_max Check failed: registry.count(type) == 1 (0 vs. 1) Unknown layer type Normalize". Can you help me about this problem? I guess I need a "normalize_layer.hpp" or something.

HI，
What should i do if there is a error on Normalize layers? what times you will update your repo?

Thanks

Hi
When using the prototxt and caffemodel in the caffe-windows version, I got the error as follow:
net.cpp:773] Check failed: target_blobs.size() == source_layer.blobs_size() (1 vs. 0) Incompatible number of blobs for layer norm_conv8_max

I try to find the problem, but failed. Can you give me some suggestion?
Thank you!

HI:
If i train the caffe-moon my model, i should use which caffe version is.
I think the model input is multi-lables data, is right?
thanks

Hi, thanks for your code.
I tested moon_tiny_iter_1000000.caffemodel on my computer(use test_moon_euclidean.ipynb) and got average error 12.69%. In README.txt, you noted that the accuracy is 89+%, and in paper, it says the average error is 9.06%.　I'm not sure if the above caffemodel is the final weights.

1. I do not find the mixed objective optimization，this project no relationship with the paper—“MOON”.
   2.Why use the loss function —“EuclideanLoss”？It is better use “Sigmoid Cross-Entropy” or “HingeLoss”？

HI:
i have some questions:
1 in the convert_imageset_multi_label.cpp line 164
const std::string& data_label = datum_label.data();
I think the "datum_label" maybe " datum_image".

2	       How do you prepare the samples?

3       Which a few attributes are used to predict?

4	        How to train? If join face key points

thanks

Hi,
What ratio did you use for separating train and test data?
I want to be sure that the current model is not trained on the whole CelebA dataset and the reported result is on test set, so there is no overfitting concern.
Thanks

Hi,
How should I compile and use convert_imageset_multi_label.cpp?
Thanks

In train_val.prototxt， I do not find the mixed objective optimization.

layer {
 name: "euclidean-loss"
  type: "EuclideanLoss"
  bottom: "moon-fc"
  bottom: "labels"
  top: "loss"
}

Could you explain it? Thanks.