This is neat! I hope you don't mind a bit of constructive criticism here, but early on in my research career I also thought it would be a good idea to "visualize the neural network connectome" in 3D (I implemented a very rudimentary version of your Browser-based visualizer in QT + OpenGL, no training frontend).
It turns out that while such tools seem useful at first glance, they turn out to not be that helpful to power users. For models bigger than LeNet, things get really ugly to visualize. And once you understand a high-level module and can take its training for granted, there isn't a need anymore to really look at it anymore. It can also be kind of annoying to tumble around in 3D when you just want to look at some activation maps. What does the 3D aspect of the visualization buy you here?
Tools like TensorBoard + Jupyter notebooks for inspecting weights and ad-hoc visualizations (e.g. VizDom) seem to strike the right balance.
If you want to continue pushing in this direction, I highly encourage embarking on an actual Deep Learning research project using your tool. In ML it's so important to dogfood your own software!
To kinda build on this, I think where this particular tool could be most beneficial would be for users just learning about neural networks, deep learning, CNNs and the like. That is, newbies to the arena.
Some may not need such a tool, and can easily visualise the layers, interconnections, functionality, etc. Others may be able to get by with 2D and "paper" representations.
I think, though, there may be a segment of learners who could benefit from a tool like this. For such people, the tool wouldn't have to support anything super-complex; smaller models and architectures would be fine.
I don't have a lot of experience with ANNs (just a few MOOCs and tutorials here and there), but from what I recall from those experiences, a tool like this could be beneficial, both for visualizing a complete NN graph, as well as visualizing a partial graph as it is built up, layer by layer (and to investigate "middle layer" operations and processes).
Thanks for your suggestions! TensorSpace is an in developing framework, and I acknowledge that it has some shortcomings, as you mentioned that for a big network, it will be too long in TensorSpace to visualize, and I am trying to figure out a solution to optimize it, for example, add "layer group" concept in TensorSpace in the future, which may make the long network shorter.
Honestly speaking, there are many mature tools for developers to use in model training, and everyone has own "optimum scheme". For me, TensorSpace is not designed to be a "silver bullet" to replace previous solutions or tools, I hope it can help engineers understand existing models and find areas where they can be applied.
Personally, I am interested in data visualization, and I found that neural network visualization is really a cool area. My original intention to make TensorSpace is to share this "beautiful scenery" with hackers!
Are there any good guides, tutorials, or research papers that investigate or advise how to inspect weights during training for debugging. The only things I read are to watch out for vanishing gradients, and when fine-tuning the most change in layers are seen toward the end of the network, not the beginning layers.
Yes, a recent exciting phenomena of interest to researchers is how and why the spectrum of the Hessian appears to separate into 2 parts - a "bulk" part that changes very slowly and "outliers" that change quickly. This suggests that only a few weights in the model actually change during training. If one could determine which weights these are, it might lend to faster and more efficient learning algorithms that don't have to backprop to all the parameters in a large neural network.
This looks really neat and it's definitely fun to play around with. I can't resist playing around with tools like this for a few minutes, but I've never really figured out what they're good for.
What am I supposed to learn from them? What is the actionable information? That's not really a criticism, I just feel like I'm missing something.
TensorSpace provides structure information and some basic metric for deep learning, and we use these famous network to show what TensorSpace can do, and how to use TensorSpace API to construct networks. For deep learning experts, these famous network's information is really basic, but we hope TensorSpace can help others to understand existing model, or help to present customized models.
If you're looking for ideas on what sort of information to display in the visualizations, this article is just chock full of them: https://distill.pub/2018/building-blocks/
Quite neat! Some constructive feedback - I think in addition to mouse controls, having keyboard 'spaceship' flying around is pretty important. I would also try to add model interprability tools (SmoothGrad) as an option, and perhaps contact some of those researchers to see how this can best be used for model understanding (I would guess visualizing grad flows / weights is more interesting than just layers).
Thanks so much for your suggestions. As you say, visualizing weights may be really helpful, and I get many feedbacks for adding visualizing weights. I will try to add this feature in later version of TensorSpace. Temporarily, TensorSpace integrate TrackballControl which is an plugin for Three.js to handle "space exploration", and I will try to add more keyboard event for TensorSpace.
I have viewed your blog, and it is really fun. However, it is a pity that the online demo is outdated. Could you tell me what technique you use to construct 3D scene? Thanks.
1. display weights and gradients (in addition to activations)
2. display how activations/weights/gradients change during training. For example, I should be able to point to checkpoint directory where my model is saved every epoch (or every few iterations), and then hit 'play' button to see how a particular feature map or weights filter is changing as I move between epochs/iterations.
3. display which inputs contribute the most to the activations, at every layer.
These additional features would be very useful for debugging, for model compression, for identifying information flows (e.g. in a DenseNet), for saliency analysis, and probably for other things as well.
Thanks for your feedback, I am planing to visualize weights in next version of TensorSpace. These days, I add a "live loader" for TensorSpace, user can use this API to visualize online training process.
And for others features, we are discussing how to integrate them with TensorSpace, looking forward to your further suggestions.
Really cool ideas going around in this thread. I really like thoughts around visualizing the changes during training -- not just activation.
As these tools currently stand, I agree that they are mostly for educational purposes. Taken to the nth degree, however, I think they can make a DL expert put on a VR headset when they get to work.
Thanks. Visualize Keras model is easier than other kinds of model, as TensorSpace's API is designed based on Keras, you can easily use keras-like TensorSpace API to build TensorSpace model and visualize it.
I am trying to design RNN layers, however, RNN is not the same as CNN layers. Now we use 3D to visualize CNN layer, as RNN has a time level information, it required 4D to visualize it, but I will try my best to add this attractive feature to TensorSpace!
And then I followed up with an early TensorFlow visualizer https://github.com/ericjang/tdb
It turns out that while such tools seem useful at first glance, they turn out to not be that helpful to power users. For models bigger than LeNet, things get really ugly to visualize. And once you understand a high-level module and can take its training for granted, there isn't a need anymore to really look at it anymore. It can also be kind of annoying to tumble around in 3D when you just want to look at some activation maps. What does the 3D aspect of the visualization buy you here?
Tools like TensorBoard + Jupyter notebooks for inspecting weights and ad-hoc visualizations (e.g. VizDom) seem to strike the right balance.
If you want to continue pushing in this direction, I highly encourage embarking on an actual Deep Learning research project using your tool. In ML it's so important to dogfood your own software!