dgschwend/netscope

Neural network visualizer and analyzer

Netscope CNN Analyzer

available here: http://dgschwend.github.io/netscope

This is a CNN Analyzer tool, based on Netscope by ethereon. Netscope is a web-based tool for visualizing neural network topologies. It currently supports UC Berkeley's Caffe framework.

This fork adds analysis capabilities, enabling the computation of network complexity (number of operations) and network size (number of parameters) for easy comparison of different networks.

Documentation

• Netscope Quick Start Guide

Demo

• 🆕 Visualization of ZynqNet CNN
• Visualization of the Deep Convolutional Neural Network "SqueezeNet"

License

Released under the MIT license. All included network models provided under their respective licenses.

netscope's Issues

What does "macc" mean in the analysis of computation? Thanks

I don't know "macc" means what?
I know 'add' means addition, exp means exponential,
Really Thanks!

Concat error since last update

Hi dgschwend,

Glad to contact you again.
I've found a concat parameter error of this netscope project, since the project was updated 7days ago.
You can find it at "fire2/concat", the output channel should be 128 but was demonstrated as 192.

See the link or picture below.
http://dgschwend.github.io/netscope/#/preset/zynqnet

Cropsize gets overwritten

The value set here

netscope/src/analyzer.coffee

Line 41 in 8ae36d8

d.batchOut = 1

gets overwritten a few lines below. Bug?

Could you add support for pytorch?

Could you add support for pytorch?

activation?

I analysis my network and get that Activation is NaN, so what's this mean? Thanks.

How to calcucate the value of activation ?

Hi, Is the size of the CAFFE model the size of PARAM? what is the size of activation? (caffemodel size == PARAM size + activation size). I'am confused. thx.

Unknown Layer: power

@dgschwend

Does this recognize power layer (http://caffe.berkeleyvision.org/tutorial/layers/power.html) ?

thank you

Visualize network from repo

Can I visualize network from repository, not from gist?
For example:
https://github.com/mrgloom/kaggle-dogs-vs-cats-solution/blob/master/learning_from_scratch/Models/AlexNet/deploy.prototxt

Add Layer Visualizations

Suggested by @Grabber in DeepScale/SqueezeNet Issue #16 "regarding performance improvement for AlexNet"

@dgschwend That is a must have tool, I was exactly thinking about an oscilloscope for CNNs! Do you think your tool could generate visual representation about layers too?

On Darknet framework there is a feature called visualize that generates visual representations of the filters, layer by layer, take a look:

It would be useful to have this kind of visual representation rendered on the menu that is shown when you pass the mouse over the network layer.

ELU Layer doesn't supported

It would be nice to support ELU layer: http://caffe.berkeleyvision.org/tutorial/layers/elu.html

Like to get in touch...(no real issue)

Dear David, I have read your thesis. Very nice piece of work. I picked up your version of Netscope after I already played with ethereon's. In my version I extended the capabilities with some load/save prototxt file buttons and also zooming and panning. Please contact me directly. You find my email address here:
http://researcher.watson.ibm.com/researcher/view.php?person=us-geert

Error Encountered,Unknown Layer: interp

Hi David
when I try to visualize this model, I got Unknown Layer: interp
@dgschwend

what does macc stand for ?

Hello, Thank you for your script.
Can you elaborate what macc stands for ?

Permute problem

Hi David,

I have a question for Permute layer.
CAFFE blob : [# batch, # channel, height, width]
If Permute {dim:0, 2, 3, 1}, then the output should be [# batch, height, width, # channel]
But in this netscope CNN analysis, the output would be [# batch, width, # channel , height]
Which one is correct?

Thanks,
Tommy

NaN when input data is 5D blob

Hello, Thanks for sharing the good tool
However, I got the NaN output dimension when I use the input data shape is 5D (sample, channel, depth, height, width), instead of the normal 4D blob ( (sample, channel, height, width). Could you fix it to allow input is 5D blob? Thanks
This is my prototxt

input: "data"
input_dim: 1
input_dim: 3
input_dim: 48
input_dim: 224
input_dim: 224

layer {
  name: "conv1"
  type: "Convolution"
  bottom: "data"
  top: "conv1"
  convolution_param {
    num_output: 16
    pad: 1
    kernel_size: 3
    kernel_size: 3
    kernel_size: 3
    stride: 1
    weight_filler {
      type: "gausian"
      std : 0.1
    }
    bias_filler {
      type: "constant"
    }
  }
}

How to modify this project to adapt for self-written layers?

Hi, thanks for this work~ it is really useful.
I modified this project ( mainly in analyzer.coffee) to make it adapt for some of my own layers and parameters.
But I did not find how to re-compile this project to make it work (I konw I should use coffee but not sure how to compile the whole project ). Could you give me a hint about this?
Thanks again!

About correlation layer

Flownet and some siamese network have the correlation layer, its correlation two feature map, can you add the analysis in netscope？

How to calculate the total activation?

Hi, may I know how you calculate the total activation?
I run a simple prototxt in the netscope to check the activation value of each layer. There are only two layers in my prototxt, i.e., data layer and conv1. The netscope shows the activation value for data layer is 4 and the activation value for conv1 is 1, but the total activation is 9.

Slice layer has not been added.

Can you please add slice layer in your analyzer. ?
Slice layer is supported in BVLC caffe.
http://caffe.berkeleyvision.org/tutorial/layers/slice.html

Slice layer is required in DPN(Dual path network).

Wrong parameter count

Hello,
shouldn't there be the bias count added to the parameter count?

For example in the details of ZynqNets conv1 layer there is written 1728 (= 3*3*3*64) parameter. Shouldn't it be 1792 (= 3*3*3*64 + 64)?

Netscope - snapshot

Dear David,
I would like to download an image of the SqueezeNet architecture as it is shown in Netscope, without loosing definition (I want to avoid snapshot programs). But the architecture shown in Netscope is not an image object that can be saved in a file.
What would be the best way to do it?
Thanks in advance

PReLU Layer doesn't supported

It would be nice to have a support of this layer:
http://caffe.berkeleyvision.org/tutorial/layers/prelu.html

the total activation memory is not correct

Hi dgschwend,

I've found a error of this netscope project.
For example, the sum of activation memory in Googlenet is nearly 102.1M (input 227x227, batch 10),
but the total activation memory show that is nearly 103.73M.

The sum of activation memory in ZynqNet is nearly 8.51M (input 256x256, batch 1),
but the total activation memory show that is nearly 8.8M.

Add FPGA-accelerated ZynqNet CNN

ZynqNet is an FPGA-accelerated embedded CNN which is publicly available at https://github.com/dgschwend/zynqnet.

The ZynqNet CNN is now available as a preset in Netscope CNN Analyzer.

The ZynqNet Project Report contains additional tips + tricks for the usage of the Netscope CNN Analyzer.

Error when the network name is undefined

When the name field is not specified in .prototxt, the tool shows Untitled Network title and can't handle the input dimensions. It doesn't show any error message, so it's hard to understand why the blob dimensions are NaNs.

After the name was specified, the dimensions became normal:

@dgschwend can you fix this behavior? There are two variants: show a error message or ignore missing name and draw the network as usual.

resnet152 points to output of resnet50 on http://dgschwend.github.io/netscope/quickstart.html

resnet152 points to output of resnet50 on http://dgschwend.github.io/netscope/quickstart.html
is there a link error or hasn't resnet-152 been analyzed by your tool?

Uncaught Graph is not a DAG

caffe prototxt: https://github.com/Andrew-Qibin/DSS/blob/master/deploy.prototxt
An error "Uncaught Graph is not a DAG" occurs when I try to analyse the above network. However it works just fine on http://ethereon.github.io/netscope, without analysis.
How can I get it work?

Error encountered: Unknown layer: reshape

Hi, can you include the reshape layer into netscope? Thanks.

Slice layer

Could you please suport slice layer and some other new layers like depthwiseconvolution and shufflechannel?

Wrong number of MACC operations for Googlenet

Each convolution layer seems to have 10x the correct number of macc operations

modify analyzer.coffee

I made some modifications to analyzer.coffee but found no change in analysis. How can these modifications take effect?

How to export calculation counts for each layer?

Hello, could you please tell me how do you export the calculation counts of each layer? especially those relu and pooling layers which is not so easy to export from convolution operations...Thanks a lot~