V0.12.3 Various + Weights graphic

Quick summary - good fix on CSV BOM, & yes - set it up and it runs. Nice!!

The weights graphic

It’s nice too, but I’m not sure about the scale…
Convolution 1 Weights mnist 0_12_3

I was messing around with some hyperreal arithmetic the other day (infinitesimals & infinities) but I’m not sure there’s much between +0 and -0 even in the hyperreals :wink:

(Also, it might be nice to put the actual values in the centres of the boxes - the value on mouseover is nice but I can only see one at a time and I can’t remember them all; also - what’s the leading "#: "? Looks like it should be a coordinate but there’s only the x value)

Spinning Circles

I kept waiting for the model previews to appear, but they didn’t - then I realised it had started training automatically and I was in the Statistics view. Previews do appear in model view, but even after training they don’t appear in statistics.

I do think it should be up to the user to start training… I expect to modify the recommended model.

:man_facepalming: dumb comment of the year above

Of course, if you actually click the button that says “Run model” you should not be surprised - or complain - if it actually does that. Sometimes the train of thought runs right over the answer. My excuse is that I was trying to run PL while wearing a blindfold and straitjacket - honestly.

Performance

Nice :slight_smile: 511s to do a single Epoch training run on ~120k mnist pngs (5% held back for validation, 10% for testing, by order of values in CSV with randomise off)

Test

Thank you for the warning about “large” dataset - let’s see what happens with the 14k of the partitioned dataset…

Alas, it’s not clear that it “fails” because nothing at all seems to happen: selected the model then Confusion, Metrics, Metris & Confusion and each time on Run Test back to the Test page and “There are no test [sic] running at the moment”.

But now it seems the kernel is offline… was that the error you were expecting? (probably happened 1st time and I didn’t notice before I tried the other test types)

Keep up the great work!

@JulianSMoore
Thanks as always for the feedback! :slight_smile:

The weights graphic
I agree that there is not much room between +0 and -0 :sweat_smile: I’ve added it as a bug, should be a simple fix of just sending the correct min and max values (maybe the decimals are the isse :thinking: )
Great suggestion with showing the numbers centered on each box as well! Will check if that’s a possibility with this framework.

Spinning Circles

My excuse is that I was trying to run PL while wearing a blindfold and straitjacket - honestly.

Hahaha XD

For the spinning circles, it’s a known bug that if you start training straight from the Data Wizard you will have those spinning circles, should be fixed and released in a week or two :slight_smile:

Performance
Great to hear it ran well! We put quite a bit of time into optimizing it, and it will soon be optimized more still. :slight_smile:

Test
The test is very MVP right now unfortunately. One feature we are working on right now is to introduce loading bars to show how long the test will take to generate (as well as fix the issue where it can crash for large datasets/models).
What happens is that it takes so long to generate certain tests that:

  1. The kernel gets locked up (which is why you are seeing Kernel offline while trying to do something else)
  2. At some point (I think after a minute or so) the websocket connection will time out. You will now be able to use the rest of the tool as normal but the tests will have been lost.