Though I am quite aware of the paper 'Visualizing and Understanding Convolutional Networks', I am not sure of how we can visualize 3D conv filters and how to interpret them.
As explained in the original paper of C3D by FacebookAI, they perform deconvolution on the intermediate conv feature maps to project them in the image space to visualize what the model is learning. But still things are not clear to me in this regard.
Take the output of the conv layer. and then perform some algorithms for dimension reduction. Then plot along with labels. You have to take three axis.
For me, the visualization of the filters may have different meanings.
If you just want to visualize the "filters" used in the network, you can extract the weights directly from the trained networks.
If you want to visualized the activation values after the layer, just take an input and do the predict for the specific layer.
If you want to analyse the behavior of the specific unit in the layer, it is interesting to do the projection over the input pixel space. In this domain, you can do the de-convolution, just like the Zelier's paper, especially to mark the position of the max-pooling value. (I remember Keras and Caffe have the plug-in for this function). Or you can do the max activation: just like training the model, but fix the trained weights, to optimize an input which can maximally activate a specific unit of the layer. For this, you can refer to the project of DeepVis Toolbox by Yosinski
- Badges
- Science topic
- Similar topics
- Computer Science
- Images