gustavla / fractalnet Goto Github PK

when i run crox train.crox.prototxt > train.prototxt, error happen:

ERROR globals.crox:2 Unrecognized command: define-eval

why? can anybody help me?

When I use the fractalnet-34.caffemodel.h5 finetune, it crop errors: Message type "caffe.LayerParameter" has no field named "drop_path_param".
What can I do?

src/caffe/layers/fractal_join_layer.cpp: In member function ‘virtual void caffe::FractalJoinLayer::LayerSetUp(const std::vectorcaffe::Blob<Dtype_>&, const std::vectorcaffe::Blob<Dtype_>&)’:
src/caffe/layers/fractal_join_layer.cpp:11:3: error: ‘FractalJoinParameter’ was not declared in this scope
FractalJoinParameter param = this->layer_param().fractal_join_param();
^
src/caffe/layers/fractal_join_layer.cpp:11:24: error: expected ‘;’ before ‘param’
FractalJoinParameter param = this->layer_param().fractal_join_param();
^
src/caffe/layers/fractal_join_layer.cpp:14:13: error: ‘param’ was not declared in this scope
std::copy(param.drop_path_ratio().begin()

[Hi,]
when i reproduce this model on cifar10, i find that the validation error is very unstable. For the first 200 epochs,the top1 error is fluctuating between 0.15 to 0.35. Then it drops to 0.1 to 0.14, because of the smaller learningrate , during the 200 to 300 epochs .And finally ,it keeps stable at 0.08 , for the last 100 epochs. Does anyone see the similar things? how can i make the validation error smooth ?
The result as below.

net-train.pdf

drops_ dose not be reset after Forward and Backward, it still keeps the value of last Forward.
I suggest to reset drops_ at the beginning of Forward.
this is my implementation of FractalNet with global and local drop in caffe:https://github.com/KangGrandesty/fractalnet.
It may be useful for you.

Hi, I conducted some experiments on cifar-10 dataset, and I only used the local_drop_path, I've tried lots of drop_path_ratio and dropout_ratio, the full-net's result was closed to yours , but the results of single columns were Not as satisfactory as expected. Was this owing to the global_drop_path which I didn't use? If so, how can I implement this global_drop_path? I'm not quite sure about the meaning of 'tying the weight'.
Hope to hear from you soon, thx!

I used your prototxt to make a visualization of the architecture using netscope. I had to modify a few things to get it working because of the layers specific to this project, like replacing FractalJoin by Concat and removing DropPath.

Image and prototxt are available here. I'm posting it in case someone also finds it useful.