emer / emergent Goto Github PK

much more effective than trying to display all the values, which quickly gets illegible.

These 2 images show input changing from cycle 50 to cycle 75. At this point not sure if it is a problem with history getting out of sync when going back and forth or a problem in my application re-init. I'll take a look and determine.

Hi all, I'm having trouble installing Emergent on my XPS-13 running Ubuntu 16 LTS. I had no issues with GoKi or GoGi, and the widgets built without errors. Go getting leabra (go get -u github.com/emer/leabra) did not complain. The error I get is:

dan@dan-XPS-13:~/go/src/github.com/emer/leabra/examples/ra25$ go get -u ./...
# github.com/emer/leabra/leabra
../../leabra/layer.go:575:37: not enough arguments in call to ext.Floats
	have ()
	want (*[]float64)
../../leabra/layer.go:575:37: ext.Floats() used as value

Additional system info:

dan@dan-XPS-13:~/go/src/github.com/emer/leabra/examples/ra25$ lsb_release -a
LSB Version:	core-9.20160110ubuntu0.2-amd64:core-9.20160110ubuntu0.2-noarch:security-9.20160110ubuntu0.2-amd64:security-9.20160110ubuntu0.2-noarch
Distributor ID:	Ubuntu
Description:	Ubuntu 16.04.6 LTS
Release:	16.04
Codename:	xenial

and go version is go1.13 linux/amd64.

Does anyone have suggestions about how to troubleshoot this? From my perspective (knowing nothing about Go) it just looks like it's just a usage issue in the layer.go code, per the error message, but maybe someone who was involved in writing this code can weigh in.

Still showing old DG layer in hip_bench

Note that in this screenshot the network graph overwrites the controls at the bottom of the window.

not using the proper topzero / bottomzero positioning for the view edit cell.

After running a training session if I click on the various buttons in the EpcPlot window to Save weights, Save SVG, etc. I get erratic responses. Sometimes it gives me the dialog box, but other times it gives me an entirely black panel or it totally freezes the simulator, or it crashes.

@snishiyama found this bug..

Because Go randomly orders a map, there is no deterministic order of parameter application.

@snishiyama is also waiting on this bug fix as well :)

it should stop at end of run but instead it is stopping at start of next run..

Maybe this hasn't been implemented yet, but when I ran the ra25 example it displayed the layers and nodes, but no projections. Is there a switch for this, or is this not implemented yet or a bug?

Infrastructure is in there but code is not!

@snishiyama also waiting on this one :)

Commenting out Blockupdates in this code gets my simulation running again.

func (nv *NetView) GoUpdate() {
if !nv.IsVisible() || !nv.HasLayers() {
return
}
if nv.Viewport.IsUpdatingNode() {
return
}
//nv.Viewport.BlockUpdates()
vs := nv.Scene()
updt := vs.UpdateStart()
nv.UpdateImpl()
//nv.Viewport.UnblockUpdates()
vs.UpdateEnd(updt)
}

cannot open app with macos catalina

dyld: Library not loaded: /usr/local/opt/gsl/lib/libgsl.23.dylib
Referenced from: /usr/local/bin/emergent
Reason: image not found
[1] 24695 abort emergent

The problem is the turning the learn flag off will only stop learning when the layers and prjn are on. Then everything is off, it will still learn.

Following the Wiki Install instructions I successfully installed and built the ra25 example, but when I used the command
./ra25 it told me file not found.
However, if I used the full pathname it worked. The path to the executable is somehow not being set. How to set it should probably be part of the Install instructions.

e.g., in ch7/bg, putting GPiThalLayer for selector doesn't work. presumably MatrixLayer not doing anything either.

If you use either the bottom or right side scroll bars the network disappears. But immediately comes back if you click on one of the right side buttons for the different values to display. Also comes back if you click the movement controls on the bottom (pan, rotate, etc.).

• add emer.Prjn.TypeName() that does string conversion on Prjn.Type() -- must be done in derived packages to get updated list of names

• use that to filter -- just have a string field in Netview view toolbar -- user types in something and use "contains" to filter prjns.

• test once deep is up and running with test project -- first case where needed.

I have noticed that when I leave a project open for a while, such as over night, the window gets very wide, many multiple screen widths wide. to resize it I need to move the window over and then from the right side keep manually shrinking the width.

this is a bug in gonum/plot -- will need to investigate and suggest a fix probably.

I have the net view window open to the full vertical extent of my large scale monitor. It does not show a right hand scroll bar, but it only shows the top half of the controls at the bottom of the window (rotate, pan, change size, etc.). The controls are functional, but they are hard to read.

The fix is to use this logic in the standard AlphaCyc method for updating after QuarterFinal (when ActP is updated) -- this updates for anything <= Quarter:

		ss.Net.QuarterFinal(&ss.Time)
		ss.Time.QuarterInc()
		if ss.ViewOn {
			switch {
			case viewUpdt <= leabra.Quarter:
				ss.UpdateView(train)
			case viewUpdt == leabra.Phase:
				if qtr >= 2 {
					ss.UpdateView(train)
				}
			}
		}

I will update fixes for ra25 and hip examples, and also textbook sims, later.

In developing my model I found that when calling AddCol() the first dimension was not being set. I added dtable_test.go with an AddCol4D test function which also failed.

I'm create a separate issue here because this seems a different problem. Testing with exact the same model can show slightly different results each time, as shown in SSE and other stats. Might be a random stat issue?

This may be a gi bug or an emergent bug. Not sure.

I had the ra25 example open today since this morning on my Mac under Mojave. When I went back to it just now the windows were totally black and stayed black, unless I went full screen. But then when I left full screen mode, it gave me really distorted text.

I am using an LG large screen monitor with new MacBook Pro.

Maybe related, I noticed that sometimes on the large screen monitor the window will rescale to be several times wider than the screen. I have to drag the right edge back to get it scaled to the proper size

I installed the new Go version of Emergent by following the instructions on (https://github.com/emer/emergent/wiki/Install) and while Go and Gogi are installed correctly on my Windows 10 PC, when I build the ra25 example exe in emer/leabra/examlples/ra25 and try to run it via the command line, it doesn't open up the exe window. The window actually remains glitched out and minimized (screenshot attached below) and clicking on the window icon in the toolbar doesn't seem to do anything.

The only way to then force the window open reliably is to go into window's split-screen mode (by, for example, dragging a window to the edge of the screen) and then selecting the 'Leabra Random Associator' window via the split-screen menu which pops up. The exe window then opens up on half the screen and there's no problems after that in rescaling it to fullscreen. I think perhaps the glitchiness has something to go with the initial scaling of the window and how the initial values may not work on all screen resolutions?(mine is 1080p) I wasn't sure if this was just an issue on my computer so I thought to report it. It's possible that if this happens on other people's computers as well, they might think (as I thought) that their emergent/Go/Gogi is not set up correctly when in fact it might be.

I tried the command - and + keys in GoGi preferences, but there are a couple of issues currently.

1. As you said, you can use the command and + or - to zoom in or out, but there are two glitches: a) you can save the change with the Save Zoom button, but if you go back to the program window and hit the Init button it crashes the program. It only works if you quit the program and then restart, b) when you do the command key thing, it is frequently the case that the first command goes in the wrong direction, that is the first command + will zoom out and a command - will zoom in. Any command after that does in the right direction.

2. DPI selection. Although you can change the DPI values in the text box, the Save Zoom button does not save the changes and it just goes back to the previous value

3. Is this documented anywhere?

I'm a bit confused about something. The gui interface would seem to indicate that there are various things that can be changed in the interface, but they don't seem to actually change. And I don't see a way to save the changes.

First, I clicked on RA25 in the Net field and got a dialog box that has Add Layer.. and Connect Layer...I added a Layer and then hit Build, but nothing happened in the gui (although it did create a dialog box with relevant info in it). ra25 crashed while I was trying to edit fields here. I also tried to use the gui to add Parameter sets, but that crashed when I was playing with it.

My assumption was that because everything must be defined in the text file and then the network built, that one could not dynamically change or add things in the gui. But if it is true that these kinds of things can't be changed in the gui, why does the gui have Add buttons that make make it look possible. And if it is true that one can modify these things in the gui, where (or how) do the changes get saved for use later?

This is a question rather than an issue. I had two related questions:

1. where do you set the sse criteria for stopping training?
2. where do you set the number of consecutive epochs at that value to stop training?

Both seem set for projects in chapter 4, but I cannot see where.

would make it easier to then add a param to that by pasting in path.

needs parent context but valueview may provide that.. really a gogi issue but most relevant here.

This may be a GoGi issue, but it would be nice if one could change not just the font, but the font size for the interface.

need to add a VarProps() method to emer.Layer, Prjn which returns go-style tags for range, auto-scale, etc. set for dwt etc to be auto scale.

also auto-scaling not working for prjn vars?

and need to rearrange wt update to happen before trial not after so that dwts are valid in view.

Currently is reading each variable and unit value separately..

• change UnitVals* to take the []float32 slice / tensor so it doesn't alloc mem if not needed
• use unsafe offsets instead of reflect to access unit vars
• double-check about extra unit vals in deep.

see what can be done for prjn-level along same lines.

This isn't really an issue, but I did notice that the active area for hovering over a node to get values is the outline on the 2 D layer surface and not the bar that graphically shows the value. This makes sense and I don't see what would make more sense, but this does have the effect that when the network is tilted, as is the default, this can make it difficult to get values for back rows when there are large values (bars) on the nodes in front. I don't think it effects the readout, but it does make it hard to see where one has to click to get a value.
I also think that the naive thing for a user to do is to think that clicking on the column will always work, when it's the node outline at the base of any column that is actually active.

If I leave a project open for several days, it will gradually get wider and wider to the point where it is several multiples wider. Both the Control panel window on the left and the Neview panel on the right continue to get wider. The height does not change much, if at all during the same time frame.

I was looking through the various ParamSets in the interface for ra25 and it crashed on me.

It not happens the other way around but when you start with a short table but test with more trials than your original table, test trial log only stores the number of rows in the short list.

I was trying out the new functionality for looking at s.wt and r.wt and its not working properly.

1. one initial point is that I had to download the changes and rebuild gi for it to recognize the selections.
2. Once I did that, I was playing around with ra25 and it was not behaving properly. If I set it to show s.wt then when I click on the nodes in the Input layer, I get a single green highlighted node in Hidden1 and then a bunch of red highlighted nodes in Hidden2.(note that not all nodes in the input layer respond to a click) But if I click on a node in Hidden1 nothing happens. Then if I click on nodes in Hidden2, I get a response for some of them, but not all of them. When it works, I get a green highlighted node in the Output layer and the red highlighted nodes in Hidden2 change, but nothing happens in any of the other layers.
3. r.wt has similar problems

I've looked with the debugger and the a2d -> a2d prjn in my model is definitely added and the sender is correct. So, maybe netview isn't showing self connections?

Is there anything in the works to get it so you can view unit specific information in the gui, like before? In the old emergent, as well as viewing activity, input, etc, you could view a unit's receiving weights (r.wt) or sending weights (s.wt). Wondering if that will be possible here. Thanks!

to facilitate more exploratory on-the-fly data analysis.

also don't forget to add IdxView to tableview and the sorting etc

In my copy of emergent (updated just now with go get -u ./...) the mouse location is not registered properly. This screenshot illustrates the problem:

The "All parameters" button is highlighted even though the mouse is upwards and to the left of the button. Maybe this is also related to the poor default resolution on my machine? In case it's relevant information, my display is set to 1920:1080, but the default resolution seems to be poor regardless of my settings. Using ctrl - to change the emergent zoom doesn't resolve the pointer issue.

When you start training with ra25, if you click the TrnEpcPlot button to look at that plot, the other buttons, such as NetView and TstTrlPlot are largely unresponsive. That is you can click on them multiple times and only rarely will they switch you to the relevant plot. For example, the NetView button is largely unresponsive, once you have switched to TrnEpcPlot. Once you stop training, then the buttons respond normally.

When I ran the ra25 example I had two problems. First, the display of the network layers, fragmented when I zoomed in at a high magnification. Second, after moving the layers around and manipulating them, the example crashed.