- ago
We're pleased to announce the release of the new WealthLab Deep Learning extension. Watch the demo video for details!
https://www.wealth-lab.com/extension/detail/DeepLearning
https://youtu.be/MAUcOcEsKEU
https://www.wealth-lab.com/extension/detail/DeepLearning
https://youtu.be/MAUcOcEsKEU
Rename
Very interesting and thank you for including it in the premium version and uploading a short video about the extension :-)
What is the difference between NeuroLab and the Deep Learning extension? Both seem to be building models based on neural networks?
- ago
#3
The key differences:
- Deep Learning has pluggable "Learning Engines". Different learning schemes are possible to integrate by us or third parties.
- Deep Learning saves a snapshot of the model at the best point in the out-of-sample training, giving you the ideal model that is the least over-optimized.
- Obviously the UI is cleaned up and hopefully more intuitive.
- Deep Learning has pluggable "Learning Engines". Different learning schemes are possible to integrate by us or third parties.
- Deep Learning saves a snapshot of the model at the best point in the out-of-sample training, giving you the ideal model that is the least over-optimized.
- Obviously the UI is cleaned up and hopefully more intuitive.
- ago
#4
QUOTE:
The key differences:
Yes, but does that mean the Deep Learning extension effectively replaces the NeuroLab extension? What can NeuroLab do that Deep Learning cannot?
- ago
#5
Yes, we'll no longer be offering NeuroLab, in fact it's off the extensions page.
- ago
#6
Does this mean the feature request for supporting hardware accelerators is officially fulfilled? It was denied earlier. https://www.wealth-lab.com/Discussion/Additional-Machine-Learning-Algorithms-and-GPU-support-6014
For the GPU acceleration part (i.e. the Cuda core part), which AI engines are supporting Cuda core? By the way, for those of you with new motherboards, we are now on PCI Express 5.0. But PCIe 4.0 daughter boards should be upward compatible. https://www.amazon.com/PNY-Generation-Express-DisplayPort-Support/dp/B0CJQH8519
And yes, some of the cheaper Cuda core accelerators may still work, but they won't be as fast.
For the GPU acceleration part (i.e. the Cuda core part), which AI engines are supporting Cuda core? By the way, for those of you with new motherboards, we are now on PCI Express 5.0. But PCIe 4.0 daughter boards should be upward compatible. https://www.amazon.com/PNY-Generation-Express-DisplayPort-Support/dp/B0CJQH8519
And yes, some of the cheaper Cuda core accelerators may still work, but they won't be as fast.
Great feature! I will be testing it soon. Thanks Glitch!
Are there any plans to add genetic evolver like mode to this module? So that you can specify a list of indicators, and the module will select several of them, make all sorts of combinations and give the best network. Does this even make sense?
- ago
#9
Cuda: while TorchSharp can support cuda, we don’t have it implemented. For one, we’d need to get appropriate hardware to develop and test with. Two, these kinds of models already train very quickly and I don’t think would get much of a boost from cuda. But it’s possible to add cuda support in a future update.
Evolver style: this is what I alluded to at the end of the demo video so it’s on my mind.
Evolver style: this is what I alluded to at the end of the demo video so it’s on my mind.
QUOTE:
you can specify a list of indicators, and the module will select several of them
This part is what finantic.IndicatorSelection was made for.
- ago
#11
QUOTE:
TorchSharp can support cuda, ... [but] we’d need to get appropriate hardware to develop and test with.
You don't need the fastest Cuda accelerator to test with.
QUOTE:
... these kinds of models ... train very quickly and I don’t think would get much of a boost from cuda.
It depends on how complex the model is. For the simpler stuff, you may be correct. For the more complex stuff, the hardware accelerator may speed you up 100 times if you have many nodes. But I doubt anyone is building complex NN models with WL today. If they are, they should step forward.
I recently purchased a lifetime subscription to NeuroLab and I am actively using it now. Will you provide access to DeepLearning for those who purchased NeuroLab?
- ago
#13
Sure
- ago
#14
You should be good to go.
After updating to version 113 and installing DeepLearning, all my strategies on NNPredictor stopped working for me. This is half a year of my work. I already have many strategies based on NeuroLab with open current positions. I am in a panic. ((
I completely removed WL, installed it again, found NeuroLab in my downloads and installed it too. Now it works. So the problem is not in build 113, but DeepLearning conflicts with NeuroLab.
Can you fix it so that i can gradually switch from NL to DL as i create new strategies on DL?
Can you fix it so that i can gradually switch from NL to DL as i create new strategies on DL?
- ago
#17
Hmmm yes I’ll work on identifying and resolving that conflict.
- ago
#18
I have them working side by side in my production build. Not seeing any issues. What exactly are you seeing? Also email me a problem NeuroLab model so I can trobleshoot.
I use WL on two laptops. I installed DL on one of them, and all strategies with NNPredictor stopped working—simply no trades appeared anywhere. I added one of the NN indicators to the same chart on both laptops, and they looked completely different. Additionally, on the Evaluate tab, all previously saved Networks on the laptop with DL no longer looked the same as they do on the laptop without DL. I have already uninstalled DL and don’t want to install it again for now to avoid breaking everything once more, so I can’t provide screenshots as an example.
- ago
#20
Did you happen to use Swish or LeakyRelu as activations? That might explain it because I applied some corrections to those activations in DL. I should be able to create a NL update with it's own copy of the Encog DLL to avoid this.
Yes, I use LeakyRelu
Is it possible to create an indicator with deep learning, and have more emphasis on recent data points just like EMA moving averages vs SMA? If I use like 10 years of daily input data from SP500, it is clear that price movement is very different now than it was 10 years ago (more rapid fluctuation for example nowdays). It might be useful to be able to give a value of how much more recent price movements affect the result vs. older ones. By this way, created indicator could possibly be more accurate if we had for example a fast market crash, as the model would have more weight on latest similar situations
You can create one indicator trained on 10 years of data and another indicator trained on 5 years of data, and then use a combination of these two indicators in your strategy.
I don't see the benefit of using two or more indicators trained with different length of data at the same time, if there is even a remote possibility to combine them and get only one indicator having more weight on recent events. If indicators with different length of training data show opposite signals, then which one to "obey"? If the one with shorter training data, then why using the longer one at all.
I do understand what you mean, and it is of course possible to use multiple indicators. I just think it would make things too complicated vs. one trained deep learned "EMA indicator"
I do understand what you mean, and it is of course possible to use multiple indicators. I just think it would make things too complicated vs. one trained deep learned "EMA indicator"
Id like to echo ww5's comment about an evolver for the Deep Learning extension. I got some great results from NN predictor but going through different combinations of indicators, models and settings is time consuming and could be automated. Ive upvoted his comment but not sure of we should log this as a feature request.
Also I get an error message when starting WL8 now. I suspect its my local configuration, but I include the messages here anyway in case its important. The Deep Learning extension still seems to work. I did update NVidia Cuda Toolkit but it didnt make any difference. I should add the extension seems to work fine.
Also I get an error message when starting WL8 now. I suspect its my local configuration, but I include the messages here anyway in case its important. The Deep Learning extension still seems to work. I did update NVidia Cuda Toolkit but it didnt make any difference. I should add the extension seems to work fine.
- ago
#26
I've not seen the error, but so far we're not even using cuda in Deep Learning, cpu only. Does this prevent WL8 from continuing to run?
@Glitch,
When I select more than one indicator in DL, I'm not able to exclude all selected at once. It can be done in future, maybe?
Another thing, it would be possible, once a model created, to exchange the Engine also? Then we can clone a model and compare results with different engines.
When I select more than one indicator in DL, I'm not able to exclude all selected at once. It can be done in future, maybe?
Another thing, it would be possible, once a model created, to exchange the Engine also? Then we can clone a model and compare results with different engines.
Request for the same function for DL https://www.wealth-lab.com/Discussion/Copying-and-pasting-layers-in-NeuroLab-12075
Which of the learning models in DeepLearning is closest to what was in NeuroLab with the LeakyRelu activation function?
- ago
#30
NL was using Encog solely.
- ago
#31
Thinking about it, why don't I add a NL-Compatability Learning Engine that could just use pre-existing NL models as is?
The main thing is that after installing the DL, the NN indicators trained with the LeakyRelu continue to work.
- ago
#33
I plan to release DL Build 2 later today which should resolve this.
- ago
#34
Deep Learning Build 2 is out, it should no longer impact NeuroLab.
Note: Before installing Deep Learning Build 2, first re-install NeuroLab from this link:
https://www.wealth-lab.com/extension/detail/NeuroLab
Doing this will overwrite the encog-core-cs.dll to the version NeuroLab required.
Deep Learning now uses a rebuilt wlencog-core-cs.dll so they should not interfere with each other.
Note: Before installing Deep Learning Build 2, first re-install NeuroLab from this link:
https://www.wealth-lab.com/extension/detail/NeuroLab
Doing this will overwrite the encog-core-cs.dll to the version NeuroLab required.
Deep Learning now uses a rebuilt wlencog-core-cs.dll so they should not interfere with each other.
@glitch the link does not seem to work.
- ago
#36
Oh, it should work now.
I have a few questions. Do I understand correctly that it makes sense to submit only ROC(n) at the output, where n looks ahead by n bars and accordingly optimizes the system for the highest profit after n bars after the entry? Or are there other options for various indicators? Is it possible, for example, to submit something like Sharpe at the output, looking ahead, so that the system is optimized according to Sharpe? Or other ideas?
QUOTE:
Or other ideas?
Instead of a blind ROC(5) you could use the actual (profits of) Trades/Positions of a trading startegy. (That is what finantic.IndicatorSelection does)
You could use some position properties to find better Exits. (Comes with a future version of IndicatorSelection)
You could use curated data to reduce noise. Noise disturbs all these Machine Learning algorithms. (comes also with a future version of IndicatorSelection)
- ago
#39
You can set the output to any indicator. In the demo, I set the output to DayOfWeek (1 bar in the future) to train the model to predict tomorrow's day of week.
I have a lifetime subscription to NeuroLab and I am actively using it.
I installed DeepLearning but it says Demo Version.
Do you provide access to Non Demo DeepLearning for those who purchased NeuroLab or I need to buy it ?
I installed DeepLearning but it says Demo Version.
Do you provide access to Non Demo DeepLearning for those who purchased NeuroLab or I need to buy it ?
Is it possible to set up the "out of sample" range the same way as it was in NeuroLab? So that it would be possible to set the exact interval yourself, and not in percentages, as it is now.
- ago
#42
No it’s currently not possible.
@flanda
Anyone who bought Neuro-Lab gets Deep Learning. Probably we have to manually add the new license.
- This is now complete for all customers who purchased Neuro-Lab.
Anyone who bought Neuro-Lab gets Deep Learning. Probably we have to manually add the new license.
- This is now complete for all customers who purchased Neuro-Lab.
Your Response
Post
Edit Post
Login is required