eldr.ai | Train an ELDR AI Model
If you haven't already, please have a look at the
Create, View & Edit Models
guide before going through this tutorial as this guide is an extension of that.
Training is the process where ELDR AI takes your AI Data, passes it through your AI Model
and learns all the links, relationships, pathways, trends and patterns within and between it, so in the near future when you present ELDR AI with information
it hasn't seen before it is able make accurate predictions & recommendations as well as give insights.
Let's quickly recap what we've done so far:
(1) Created Data - in your Data, you provided ELDR AI with the information it needs to learn from -
all your inputs (ip - values, ipc - categories/words) and an output(s). Here you are telling ELDR AI that you want it to learn how a set of inputs leads
to an output(s).
(2) Created a Model - in your Model, you have created an Artificial Neural Network - the core of AI - that's designed to learn from
your data by receiving it at one end (input layer), passing it through some interconnected artificial neurons (hidden layers) and finally churn it out
at the other end (output layer). Every time the data passes through the Model, the neurons continually adjust to make the Model Output match as close as possible
to the Output(s) we specified in our data.
Training is simply the process that governs Data moving through the Model until the Model has learnt what we want it to.
The outcome is a Trained Model capable of making predictions, recommendations and giving insights.
After Training, ELDR AI no longer needs your Data - much like a human brain, it stores everything it has
learnt in the Artificial Neurons of the Model - this is one of the main reasons why AI is so powerful - no databases are required.
You can get to the Training screen by clicking the "Train Model" link on the Dashboard, side menu or nav menu, or as below by clicking on the Training icon (
graduation cap, green box) from the View Models page:
On the Training page before you start training you will see a truncated version of your Data.
This is so you can do a final once-over to make sure the structure, inputs and outputs are what you're expecting.
You will also see the Model parameters displayed again so you can do a final confirmation
before starting training. If you need to make any change click "Edit Model".
When you're happy, click "Train Model" near the top of the page to begin Training:
When ELDR AI starts training your Model, several displays, dials and charts will appear, allowing you to monitor Training if you wish to.
It's a good idea to monitor training for at least the first five minutes, to confirm training is moving in the right direction e.g. Loss is decreasing.
After that you can leave the page, Training will continue in the background and you can come back to the Training page any time to see progress.
As you get to know ELDR AI, you will soon learn what to expect during Training, and what works best for your data/models.
Let's have a look at the displays in more detail:
(1) Server Output (red box) - This is to show raw server output and is designed to look nostalgic at the same time as giving
useful information. The key output to look for is the Loss (yellow box). This should be decreasing each time the output refreshes.
Initially the change should be quite large, but then changes will be small. As long as Loss is clearly decreasing and not stalling then Training is working.
If Loss stalls above 0.1, generally speaking Training will have probably failed. If Training is allowed to carry on unimpeded Loss will get down to
very small numbers e.g. 0.000000000001 and beyond. Seconds and Epochs are also shown.
(2) Loss, Epochs and Time % Dials (light blue box) - These are graphical dial representations of how much of a percentage of your
specified maximum Loss, Epochs and Time has been used. E.g we have used 4.2 seconds of our maximum of 300 seconds/5 minutes = 1%, 700 Epochs of our maximum 10000 = 7%,
and our Loss is 3985% over our target Loss (0.0000398 of 0.0000001). These are there to give you an idea of when Training will finish if left unimpeded.
(3) Loss/Time graph. - As you can see, when Training starts, Loss change is quite dramatic and then tails off. In this example we
can see our Loss is very close to 0 which is ideal.
(4) Layer Weight Contribution (purple box) - To show which layers are contributing most to learning. This can help with determining
if you have an efficient number of Layers. If one Layer is dominating while others are hardly contributing,
consider changing the number of neurons in each layer, or changing the number of layers.
(5) Epochs/Time graph - This tells you how fast Epochs are taking.
The larger your Data and/or Model, the longer it will take for each Epoch. Again, as you get to know ELDR AI you will get to know what's working for you.
(6) Neuron Chart - This shows you how much each individual Artificial Neuron is changing during Training. If the number of Neurons
in your Model is quite large, only a percentage of Neurons will be shown here so it should be used for indicative purposes only. The main thing to look for is
that most of the Neurons are changing - this is an indication of Training.
If you want to prematurely stop training for any reason at all e.g. Loss decrease has stalled
or you can see Training has got to a state that you're happy with, press the black "FORCE STOP" button. This will stop training within about
1000 Epochs:
After training has finished e.g. it has met either your target Loss, maximum time, or maximum epochs, you will receive a status alert, or which
there are several flavours:
Success
If your Model gets to the desired Loss within the specified Time and Epochs, you will get a success message.
Loss not met
If your Model completes Training but doesn't meet the desired Loss after the Time and/or Epochs have reached the limits, you will get this message.
This doesn't mean it's a total fail - you should check the predictions, recommendations and insights from this Model to see
if it has trained sufficiently well before dismissing the Model. It could just be that you requested a Loss that was practically impossible to reach with
the particular data you provided.
Timed Out
If your Model completes Training but doesn't meet the desired Loss after the maximum time, you will get this message.
This doesn't mean it's a total fail - you should check the predictions, recommendations and insights from this Model to see
if it has trained sufficiently well before dismissing the Model. It could just be that you requested a Loss that was practically impossible to reach with
the particular data you provided.
Failed Data Validation
If you get this message it means there's a fatal/total fail with your data so the Model will never Train and the data needs fixing. Data is validated
all over ELDR AI during most actions, so it's unusual to get this message without knowing about a problem before Training. It could happen if your data is
provided by API and the data coming back from the API has changed shortly after you started Training.
Total Fail
Again it is unlikely you will get this error. If this does occur it is likely to be a problem you have no control over such as a server or software error.
Please let us know about these.
During and after Training you can view current and historical Training status' of all your Models
To see all current and past Training, click on View Training from the Dashboard, sidebar or nav menu:
You can see individual Model Training History by clicking the "clock" history icon (red box) from the View Models page:
In both all and single Model Training Histories you will see the current and past training status' for each Training run you carried out; the
status messages corresponding to the alert messages you got after Training. You will also see the full Data and Model Parameters associated with that particular
Training.Each one of these training runs/rows is a Model Variant
We will discuss Model Variants in the next section - Optimisation.
That concludes the ELDR AI guide to Training Models