I'm trying to implement GradCam (https://arxiv.org/pdf/1610.02391.pdf) in tfjs, based on the following Keras Tutorial (http://www.hackevolve.com/where-cnn-is-looking-grad-cam/) and a simple image classification demo from tfjs, similar to (https://github.com/tensorflow/tfjs-examples/blob/master/webcam-transfer-learning/index.js) with a simple dense, fully-connected layer at the end.
However, I'm not able to retrieve the gradients needed for the gradcam computation. I tried different ways to retrieve gradients for the last sequential layer, but did not succeed, as types of tf.LayerVariable from the respective layer is not convertible to the respective type of tf.grads or tf.layerGrads.
Did anybody already succeeded to get the gradients from sequential layer to a tf.function like object?
I'm not aware of the ins and outs of the implementation, but I think something along the lines of this: http://jlin.xyz/advis/ is what you're looking for?
Source code is available here: https://github.com/jaxball/advis.js (not mine!)
This official example in the tfjs-examples repo should be close to, if not exactly, what you want:
https://github.com/tensorflow/tfjs-examples/blob/master/visualize-convnet/cam.js#L49
Related
I am learning Ruby, reading few books, tutorials, foruns and so one... so, I am brand new to this.
I am trying to develop a stock system so I can learn doing.
My questions are the following:
I created the following to store transactions: (just few parts of the code)
transactions.push type: "BUY", date: Date.strptime(date.to_s, '%d/%m/%Y'), quantity: quantity, price: price.to_money(:BRL), fees: fees.to_money(:BRL)
And one colleague here suggested to create a Transaction class to store this.
So, for the next storage information that I had, I did:
#dividends_from_stock << DividendsFromStock.new(row["Approved"], row["Value"], row["Type"], row["Last Day With"], row["Payment Day"])
Now, FIRST question: which way is better? Hash in Array or Object in Array? And why?
This #dividends_from_stock is returned by the method 'dividends'.
I want to find all the dividends that were paid above a specific date:
puts ciel3.dividends.find_all {|dividend| Date.parse(dividend.last_day_with) > Date.parse('12/05/2014')}
I get the following:
#<DividendsFromStock:0x2785e60>
#<DividendsFromStock:0x2785410>
#<DividendsFromStock:0x2784a68>
#<DividendsFromStock:0x27840c0>
#<DividendsFromStock:0x1ec91f8>
#<DividendsFromStock:0x2797ce0>
#<DividendsFromStock:0x2797338>
#<DividendsFromStock:0x2796990>
Ok with this I am able to spot (I think) all the objects that has date higher than the 12/05/2014. But (SECOND question) how can I get the information regarding the 'value' (or other information) stored inside the objects?
Generally it is always better to define classes. Classes have names. They will help you understand what is going on when your program gets big. You can always see the class of each variable like this: var.class. If you use hashes everywhere, you will be confused because these calls will always return Hash. But if you define classes for things, you will see your class names.
Define methods in your classes that return the information you need. If you define a method called to_s, Ruby will call it behind the scenes on the object when you print it or use it in an interpolation (puts "Some #{var} here").
You probably want a first-class model of some kind to represent the concept of a trade/transaction and a list of transactions that serves as a ledger.
I'd advise steering closer to a database for this instead of manipulating toy objects in memory. Sequel can be a pretty simple ORM if used minimally, but ActiveRecord is often a lot more beginner friendly and has fewer sharp edges.
Using naked hashes or arrays is good for prototyping and seeing if something works in principle. Beyond that it's important to give things proper classes so you can relate them properly and start to refine how these things fit together.
I'd even start with TransactionHistory being a class derived from Array where you get all that functionality for free, then can go and add on custom things as necessary.
For example, you have a pretty gnarly interface to DividendsFromStock which could be cleaned up by having that format of row be accepted to the initialize function as-is.
Don't forget to write a to_s or inspect method for any custom classes you want to be able to print or have a look at. These are usually super simple to write and come in very handy when debugging.
thank you!
I will answer my question, based on the information provided by tadman and Ilya Vassilevsky (and also B. Seven).
1- It is better to create a class, and the objects. It will help me organize my code, and debug. Localize who is who and doing what. Also seems better to use with DB.
2- I am a little bit shamed with my question after figure out the solution. It is far simpler than I was thinking. Just needed two steps:
willpay = ciel3.dividends.find_all {|dividend| Date.parse(dividend.last_day_with) > Date.parse('10/09/2015')}
willpay.each do |dividend|
puts "#{ciel3.code} has approved #{dividend.type} on #{dividend.approved} and will pay by #{dividend.payment_day} the value of #{dividend.value.format} per share, for those that had the asset on #{dividend.last_day_with}"
puts
end
I'm performing gaussian mixture model classification, and based on that, used "mvnpdf" function in MATLAB.
As far as I know the function returns a multi variate probability density for the data points or elements passed to it.
However I'm trying to recreate it on C and I assumed that mvnpdf is the regular Gaussian distribution (clearly it is not) because the results don't match.
Does anyone know how "mvnpdf" works ? Because I haven't been able to find documentation on it .
The documentation for mvnpdf is here
if you are looking for the exact code just put a break point at the point where you call it and see how it works
Okay I actually found a decent link that explains in detail what's happening inside .
This might be a better link to look at - http://octave.sourceforge.net/statistics/function/mvnpdf.html
Need to use c for a project and i saw this screenshot in a pdf which gave me the idea
http://i983.photobucket.com/albums/ae313/edmoney777/Screenshotfrom2013-11-10015540_zps3f09b5aa.png
It say's you can treat each pixel of an image as a graph node(or vertex i guess) so i was wondering how
i would do this using OpenCV and the CvGraph set of functions. Im trying to do this to learn about and how
to use graphs in computer vision and i think this would be a good starting point.
I know i can add a vetex to a graph with
int cvGraphAddVtx(CvGraph* graph, const CvGraphVtx* vtx=NULL, CvGraphVtx** inserted_vtx=NULL )
and the documentation says for the above functions vtx parameter
"Optional input argument used to initialize the added vertex (only user-defined fields beyond sizeof(CvGraphVtx) are copied)"
is this how i would represent a pixel as a graph vertex or am i barking up the wrong tree...I would love to learn more about
graphs so if someone could help me by maybe posting code, links, or good ol' fashioned advice...Id be grateful=)
http://vision.csd.uwo.ca/code has an implementation on Mulit-label optimization. GCoptimization.cpp file has a GCoptimizationGridGraph class, which I guess is what you need. I am not a C++ expert, so can't still figure out how it works. I am also looking for some simpler solution.
I'm interested in different algorithms people use to visualise millions of particles in a box. I know you can use Cloud-In-Cell, adaptive mesh, Kernel smoothing, nearest grid point methods etc to reduce the load in memory but there is very little documentation on how to do these things online.
i.e. I have array with:
x,y,z
1,2,3
4,5,6
6,7,8
xi,yi,zi
for i = 100 million for example. I don't want a package like Mayavi/Paraview to do it, I want to code this myself then load the decomposed matrix into Mayavi (rather than on-the-fly rendering) My poor 8Gb Macbook explodes if I try and use the particle positions. Any tutorials would be appreciated.
Analysing and creating visualisations for complex multi-dimensional data is complex. The best visualisation almost always depends on what the data is, and what relationships exists within the data. Of course, you are probably wanting to create visualisation of the data to show and explore relationships. Ultimately, this comes down to trying different posibilities.
My advice is to think about the data, and try to find sensible ways to slice up the dimensions. 3D plots, like surface plots or voxel renderings may be what you want. Personally, I prefer trying to find 2D representations, because they are easier to understand and to communicate to other people. Contour plots are great because they show 3D information in a 2D form. You can show a sequence of contour plots side by side, or in a timelapse to add a fourth dimension. There are also creative ways to use colour to add dimensions, while keeping the visualisation comprehensible -- which is the most important thing.
I see you want to write the code yourself. I understand that. Doing so will take a non-trivial effort, and afterwards, you might not have an effective visualisation. My advice is this: use a tool to help you prototype visualisations first! I've used gnuplot with some success, although I'm sure there are other options.
Once you have a good handle on the data, and how to communicate what it means, then you will be well positioned to code a good visualisation.
UPDATE
I'll offer a suggestion for the data you have described. It sounds as though you want/need a point density map. These are popular in geographical information systems, but have other uses. I haven't used one before, but the basic idea is to use a function to enstimate the density in a 3D space. The density becomes the fourth dimension. Something relatively simple, like the equation below, may be good enough.
The point density map might be easier to slice, summarise and render than the raw particle data.
The data I have analysed has been of a different nature, so I have not used this particular method before. Hopefully it proves helpful.
PS. I've just seen your comment below, and I'm not sure that this information will help you with that. However, I am posting my update anyway, just in case it is useful information.
First post here. Using C in Visual Studio 2008. Can work with VS 2005 if necessary.
How do I display numerical data in arrays as in a spreadsheet?
How do I plot numerical data in arrays?
These seem to be simple questions. But I cannot find solutions. So far, I would print the data to a file, import into Excel and view/plot. However, with this code there are too many arrays--so the print/import/plot is tiring.
Some constraints.
I do not want to write 20+ lines of code to do the above. MATFOR or Array Visualizer let you do the plotting with a one line function call.
They cannot display the data in a convenient format. I would like to display the data and the plot in one or two windows so that they are visible simultaneously.
This is a win32 console application---all the code is portable.
Will be using these during debugging.
Free or paid.
While I am looking for something specific, the requirements are substantially the same for any one doing numerical work with arrays and matrices--displaying data and plot simultaneously.
I am hoping that a such a tool has been written and is available.
I am also open to a solution that outputs the array data to an Excel sheet (can keep Excel open) and if it can also plot that can be great but I can live without plotting.
PS: I need this only when debugging the code.
I use ArrayDebugView which is a plug-in you install in Visual studio and draws graphs out of arrays while you are debugging your application. It works as a visual way of variable watch in debug mode. You don't need to write a line of code.
I can't think of any library that would enable what you want in a console app in less than 20 lines of code. My suggestion would be instead to script the plotting-step using MATLAB og GNU Octave to do the actual plotting.
In order to display numerical data in array, you should add the pointer to the first data element you want to observe, into the watch --- if you want to observe the array from the beginning, it would just be the array name, which is the pointer to the first element. In order to view more then one element, you add a "," after the pointer, followed by the number of element you want to observe.
For example, in order to observe the elements of float farray[100];, you should add to the watch farray,100.
In order to plot, you can copy-paste from the watch to your plotting software (i.e. excel), but it is not very convenient as you cannot copy the data column alone, but the columns to the left and right as well, so it involves extra manual editing.
I use GNUPlot (http://www.gnuplot.info/) to display my performance/speedup measurements.
I print my numbers to stdout and wrote a bash script that combines these numbers and calls gnuplot for rendering.
I made a simple plotting program for that purpose. There is only a textbox where I paste the data and a chart where it's drawn.
The data needs to be in either form:
with an automatic X (increment by 1 for each value): seriesName value
for both X and Y specified: seriesName xvalue yvalue
Most of the time I used to plot data from tracepoints.
I copy/paste the whole output window of VS, the plotting program ignores anything that doesn't follow these 2 forms (so I don't have to cleanup the string and put it in excel and all).
It does line, point, colum, area charts and save image, copy to clipboard.
MiniPlot
There are several ways to do this but this will require writing some code. Visualizing data is generally easy and straight forward but visualizing data exactly the way you want them to look will require some additional code and work.
There are several options to visualize data:
A combination of BASH and GNUPLOT
Use MATLAB or OCTAVE for all your calculations and visualization
Use PYTHON and SciPy and matlibplot libraries.
Gnuplot is a great tool to plot data but it is cumbersome to use. It looks fabulous if you invest time to get the plots right and combines excellent with LaTeX and has a good fit implementation for arbitrary functions. Visit http://gnuplot-tricks.blogspot.ch/ great site to learn all about gnuplot.
Numerical programs such as MATLAB and it's open source equivalent OCTAVE are great because they are fast implementation languages for numerical programs and have extensive additional libraries especially MATLAB. For high load numerical computing it is really slow and the plot library is only good for basic plotting needs.
Using PYTHON and its scientific programing libraries (SciPy and matlibplot) are a great combination. This allows excellent plot which are not as cryptic as gnuplot to plrogram and it is more flexible than MATLAB in plotting. Additionally it gives you a environment for numerical programing like MATLAB.