I am working with weight and bias(wandb).
However, it logs by step. And that makes plot disturbing when comparing runs.
For example, I have a run A and run B(assume that they run with same dataset).
run A: 30epochs, 4 batch, 200step/epoch
run B: 30epochs, 8 batch, 100step/epoch
then, the plot of run A gets longer(double, in this case) in axis x when it shows with run B.
How can I scale x axis depend to runs AFTER training?
You can change the x-axis used via the chart settings by clicking on the pencil icon and then selecting a different x-axis. E.g. in your case you could select "epoch" instead of "steps". Just make sure to log "epoch" to your charts, something like:
steps_per_epoch = n_samples / batch_size
epoch = current_step / steps_per_epoch
wandb.log({"epoch":epoch, ...})
Related
I've matrix with values, I've colored the matrix for visual analysis. The green region shows values e.g. 5 and brown shows values 6 and black shows value 0. I want to connect the broken region with value 5. I've tried different structuringElement.g. [110;110;000] and used 2 dilation followed by median filter to get this result.
se_mask = centered(Bool[1 1 0; 1 1 0; 0 0 0])
result = dilate(dilate(gt_mat, se_mask), se_mask)
d_gt_mat = mapwindow(median, result, (5, 5))
I'm not sure whats a better way to connect the broken regions to fill up.
I'm working with JULIA.
The functions in ImageMorphology package might be what you are looking for.
For example, with the first image in the OP, something like this can be done:
using Images
using ImageMorphology
using ImageInTerminal
using IterTools
img = load("/path/to/dir/so-image.png")
gray_img = Gray.(img)
size(gray_img) # (117, 238)
# applying dilate 7 times and then erode 7 times
closed_img =
nth(iterated(erode,nth(iterated(dilate, gray_img),7)),7)
save("/tmp/gray_img.png", gray_img)
save("/tmp/closed_img.png", closed_img)
and the results are:
Original in gray version
After processing
In Vaadin 14.6.1, I tried to create a Vaadin heatmap foollowing the documentation / example from here.
However, I encountered a few problems/questions, listed in descending order of importance below:
The heatmap supported 30 rows by 30 columns; but when I tried 40 rows by 40 columns, the entire heatmap showed a single color (blue in my case).
Is it possible to manually set the minimum numeric value and maximum numeric value for the color scheme. This way, if I plot my data one day and it has values in the range of 0 to 1, but on another dataset from another day, the numeric values range from between 0 and 0.5, the color scheme range won't automatically change (to being between 0 and 0.5) and confuse the user.
In the documentation, it has the following methods listed, but they do not seem to exist in Vaadin 14.6.1
plotOptions.setBorderColor(SolidColor.WHITE);
plotOptions.setBorderWidth(2);
When I hover over the datapoints, is there a way that the tooltip can only show the numeric value (and not the x, y coordinate or the series name etc.?)
Is there any limit to the number of datapoints? I'm hoping that if #1 above is resolved that I can plot a 100 by 100 (ie 10,000) point heatmap.
If I plot 100by100, there will be many labels on the x and y axis. In my case, the x and y axis are actually numerical, so I did what the documentation suggested and just placed category labels in the xaxis and yaxis spots. However, is there any way to only display every "nth" label, so that the x-axis and y-axis is not so crowded? (This would essentially mimic what Vaadin does for normal line charts' xaxis and yaxis).
I'm not that experienced with Vaadin Chart, but these are the questions that I can comment on:
(1) With 40x40 items you go over the threshold of 1000 in which the Chart switches into "turbo" mode for performance reasons. This seems to not be compatible with the heatmap series. You can disable turbo mode by setting plotOptions.setTurboThreshold(0);
(2) Unfortunately the ColorAxis doesn't support this, it only has an API for min and max color. Definitely a valid use-case though, and it seems to be supported by the Highcharts library that the Vaadin Chart uses under the hood. You should consider opening a feature request for this in the Github repo.
(3) This seems to be a documentation issue. The methods are available in later Vaadin platform versions, but not in 14.6.
(5) In theory not, but in practice there will be a huge performance hit in the browser due to the excessive amount of DOM elements (quick test of 100x100 froze the browser for 10s). I'm afraid the component isn't really made for such extreme use-cases. In this case it might be better to utilize a low-level JS drawing library using the canvas, or draw an image on the server-side and display that in the browser. Maybe you can also consider modifying your use-case so that you only display one slice of your data and allow the user to switch between slices.
I tried to plot using histogram_quantile with Prometheus with sum on two labels, i.e.
histogram_quantile(0.9, sum(rate(grpc_request_duration_seconds_bucket{service="$service"}[$interval])) by (le, grpc_method))
The resulting plot was not the same as me manually making a separate plot for each individual value of grpc_method:
histogram_quantile(0.9, sum(rate(grpc_request_duration_seconds_bucket{service="$service",grpc_method="GetGridFile"}[$interval])) by (le, grpc_method))
and
histogram_quantile(0.9, sum(rate(grpc_request_duration_seconds_bucket{service="$service",grpc_method="PostAgvState"}[$interval])) by (le, grpc_method))
I am trying to visualize a time series data set on one plot as a pseudo 3d figure. However, I am having some trouble getting the filledcurves capability working properly. It seems to be adding an unwanted border at the "bottom" of my functions and I do not know how to fix this.
This is my current set up: I have nb_of_frames different files that I want to plot on one figure. Without the filledcurves option, I can do something like this
plot for [i=1:nb_of_frames] filename(i) u ($1):(50.0 * $2 + (11.0 - (i-1)*time_step)) w l linewidth 1.2 lt rgb "black" notitle
which produces a figure like this:
no fill options
Instead of doing this, I want to use the filledcurves option to bring my plots "forward" and highlight the function that is more "forward" which I try to do with:
plot for [i=1:nb_of_frames] filename(i) u ($1):(50. * $2 + (11. - (i-1)*time_step)) w filledcurves fc "white" fs solid 1.0 border lc "black" notitle
This produces a figure as follows:
This is very close to what I want, but it seems that the border option adds a line underneath the function which I do not want. I have tried several variants of with filledcurves y1=0.0 with different values of y1, but nothing seems to work.
Any help would be appreciated. Thank you for your time.
Here is another workaround for gnuplot 5.2.
Apparently, gnuplot closes the filled area from the last point back to the first point. Hence, if you specifiy border, then this line will also have a border which is undesired here (at least until gnuplot 5.4rc2 as #Ethan says).
A straightforward solution would be to plot the data with filledcurves without border and then again with lines. However, since this is a series of shifted data, this has to be plotted alternately. Unfortunately, gnuplot cannot switch plotting styles within a for loop (at least I don't know how). As a workaround for this, you have to build your plot command in a previous loop and use it with a macro # (check help macros) in the plot command. I hope you can adapt the example below to your needs.
Code:
### filledcurves without bottom border
reset session
set colorsequence classic
$Data <<EOD
1 0
2 1
3 2
4 1
5 4
6 5
7 2
8 1
9 0
EOD
myData(i) = sprintf('$Data u ($1-0.1*%d):($2+%d/5.)',i,i)
myFill = ' w filledcurves fc "0xffdddd" fs solid 1 notitle'
myLine = ' w l lc rgb "black" notitle'
myPlotCmd = ''
do for [i=11:1:-1] {
myPlotCmd = myPlotCmd.myData(i).myFill.", ".myData(i).myLine.", "
}
plot #myPlotCmd
### end of code
Result:
I can reproduce this in gnuplot 5.2.8 but not in the output from the release candidate for version 5.4. So I think that some bug-fix or change was applied during the past year or so. I realize that doesn't help while you are using verion 5.2, but if you can download and build from source for the 5.4 release candidate that would take care of it.
Update
I thought of a work-around, although it may be too complicated to be worth it.
You can treat this as a 2D projection of a 3D fence plot constructed using plot style with zerrorfill. In this projection the y coordinate is the visual depth. X is X. Three quantities are needed on z: the bounding line, the bottom, and the top. I.e. 5 fields in the using clause: x depth zline zbase ztop.
unset key
set view 90, 180
set xyplane at 0
unset ytics
set title "3D projection into the xz plane\nplot with zerrorfill" offset 0,-2
set xlabel "X axis" offset 0,-1
set zlabel "Z"
splot for [i=1:25] 'foo.dat' using ($1+i):(i/100.):($2-i):(-i):($2-i) \
with zerrorfill fc "light-cyan" lc "black" lw 2
Scenario: I need to draw a plot that has a background image. Based on the information on that image there have to be multiple origins (let's call them 'targets') that can move over time. The movements of these targets will have to be indicated by arrows/vectors where the first vector originates at the location of the target, the second vector originates where the previous vector ended and so on.
The result should look similar to this:
Plot with targets and movement vectors
While trying to implement this, i stumbled upon different questions:
I would use a chart with combined series: a Scatter plot to add the targets at exact x/y locations and a vector plot to insert the vectors. Would this be a correct way?
Since i want to set each vectors starting point to exact x/y coordinates i use rotationOrigin: 'start'. When i now change vectorLength to something other than 20 the vector is still shifted by 10 pixels (http://jsfiddle.net/Chop_Suey/cx35ptrh/) this looks like a bug to me. Can it be fixed or is there a workaround?
When i define a vector it looks like [x, y, length, direction]. But length is a relative unit that is calculated with some magic relative to the longest vector which is 20 (pixels) by default or whatever i set vectorLength to. Thus, the vectors are not connected and the space between them changes depending on plot size axes min/max). I actually want to corellate the length with the plot axis (which might be tricky since the x-axis and y-axis might have different scales). A workaround could be to add a redraw event and recalculate the vectors on every resize and set the vectorLength to the currently longest vector (which again can be calculated to correlate to the axes). This is very cumbersome and i would prefer to be able to set the vectors somehow like [x1, y1, x2, y2] where (x1/y2) denotes the starting- and (x2/y2) the ending-point of the vector. Is this possible somehow? any recommendations?
Since the background image is not just a decoration but relevant for the displayed data to make sense it should change when i zoom in. Is it possible to 'lock' the background image to the original plot min/max so that when i zoom in, the background image is also zoomed (image quality does not matter)?
Combining these two series shoudn't be problematic at all, and that will be the correct way, but it is necessary to change the prototype functions a bit for that the vectors will draw in a different way. Here is the example: https://jsfiddle.net/6vkjspoc/
There is probably the bug in this module and we will report it as new issue as soon as it is possible. However, we made the workaround (or fix) for that and now it's working well, what you can notice in example above.
Vector length is currently calculated using scale, namely - if vectorLength value is equal to 100 (for example), and vector series has two points which looks like that:
{
type: 'vector',
vectorLength: 100,
rotationOrigin: 'start',
data: [
[1, 50000, 1, 120],
[1, 50000, 2, -120]
]
}
Then the highest length of all points is taken and basing on it the scale is calculated for each point, so first one length is equal to 50, because the algorithm is point.length / lengthMax, what you can deduce from the code below:
H.seriesTypes.vector.prototype.arrow = function(point) {
var path,
fraction = point.length / this.lengthMax,
u = fraction * this.options.vectorLength / 20,
o = {
start: 10 * u,
center: 0,
end: -10 * u
}[this.options.rotationOrigin] || 0;
// The stem and the arrow head. Draw the arrow first with rotation 0,
// which is the arrow pointing down (vector from north to south).
path = [
'M', 0, 7 * u + o, // base of arrow
'L', -1.5 * u, 7 * u + o,
0, 10 * u + o,
1.5 * u, 7 * u + o,
0, 7 * u + o,
0, -10 * u + o // top
];
return path;
}
Regarding your question about defining start and end of a vector by two x, y values, you need to refactor entire series code, so that it won't use the vectorLength at all as like as scale, because you will define the points length. I suspect that will be very complex solution, so you can try to do it by yourself, and let me know about the results.
In order to make it works, you need to recalculate and update vectorLength of your vector series inside of chart.events.selection handler. Here is the example: https://jsfiddle.net/nh7b6qx9/