search

yosinski / deep-visualization-toolbox

DeepVis Toolbox
http://yosinski.com/deepvis
MIT License
4.02k stars 924 forks source link

nothing about mnist's visualization #136

Open qingzew opened 6 years ago

qingzew commented 6 years ago

this is the problem https://stackoverflow.com/questions/48368441/mnists-deep-feature-visualizations-is-nothing

ZhuweiQin commented 6 years ago

Hi, could you please explain a little more about how did you visualize the LeNet on mnist? I am trying to visualize the VGG16 train on Cifar10. How should I set the label path?

qingzew commented 6 years ago

@ZhuweiQin when I have trained a mnist model with caffe, I am wondering what the feature would look like, for example, when I visualize caffenet-yos, I can see different channel focus different part of the object, but on mnist I can see nothing, why?
for label path, I think you can reference this:https://github.com/yosinski/deep-visualization-toolbox/blob/master/models/caffenet-yos/settings_local.template-caffenet-yos.py#L20

ZhuweiQin commented 6 years ago

The deconv visualization and activation (feature map) visualization explain the filters from different perspectives. Through your implementation, we can clearly see the activation of each lenet filters on mnist. It's hard to interpret the activations as the caffenet on imagenet dataset since the filter resolution is too small. Basically, the deconv projects the activation(feature map) back to the input image, showing which part of the input image activate the corresponding filter. I do not know why you can not see the deconv visualization result. The reason must be the code problem not the feature. BTW, there are some other open source projects about the deconv. Maybe you can try.