Machine learning in trading: theory, models, practice and algo-trading - page 2253
You are missing trading opportunities:
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
Registration
Log in
You agree to website policy and terms of use
If you do not have an account, please register
Well, let's say (not everything is clear to me in the Python code). And where is the training of this BAE?
Is it in pyTorch ?all on the torch
the coder works, just the results are worse with it
in short, the Kullback-Leibner distance model is very hard to learn, it's the described problem of variational coders
in fact, algorithm "converges" after 2-3 epochs and nothing happens further
One of the main drawbacks of variational autoencoders is that the integral of the KL divergence term does not have a closed form analytical solution except for a handful of distributions. Furthermore, it is not straightforward to use discrete distributions for the latent codez z . This is because backpropagation through discrete variables is generally not possible, making the model difficult to train efficiently. One approach to do this in the VAE setting was introducedhere.
you have access to the weights of the network and the ability to change them?
sure, but i don't need to.
everything on the torch
the coder works, the results are just worse
in short, the Kullback-Leibner distance models are very hard to learn, this is the described problem of variation coders
In fact, the algorithm "converges" after 2-3 epochs, and nothing happens further
One of the main drawbacks of variational autoencoders is that the integral of the KL divergence term does not have a closed form analytical solution except for a handful of distributions. Furthermore, it is not straightforward to use discrete distributions for the latent codez z . This is because backpropagation through discrete variables is generally not possible, making the model difficult to train efficiently. One approach to do this in the VAE setting was introducedhere.
The article talks about something else. It deals with the case where all predictors are discrete [0, 1]. Then there is a problem. The neural network does not understand predictors with zero variation.
You, as I understand it, have a slightly different case. You have combined in input predictors (continuous) and target (discrete matrix ncol=3). You are trying to get a qualitative distribution of latent from which you generate (restore) the input including the target, practically without training. Did I understand you correctly? It will not work qualitatively. The article shows the way of solution. Translate discrete target to continuous using RBM, connect with other predictors and use BAE (training!). And then retrieve examples from the trained VAE and use RBM to restore the target again. It's a bit complicated. But it might work.
I'll try it with regular AE.
Good luck
The article deals with something else. It considers the case when all predictors are discrete [0, 1]. Then there is a problem. The neural network does not understand predictors with zero variation.
You, as I understand it, have a slightly different case. You have combined in input predictors (continuous) and target (discrete matrix ncol=3). You try to get a qualitative distribution of latents from which you generate (restore) the input including the target, practically without training. Did I understand you correctly? It will not work qualitatively. The article shows the way of solution. Translate discrete target to continuous using RBM, connect with other predictors and use BAE (training!). And then retrieve examples from the trained VAE and use RBM to restore the target again. It's a bit complicated. But it might work.
I'll try it with regular AE.
Good luck
I have CVAE. On the encoder and decoder class labels (1,0) among other things. But I compared VAE and CVAE, the results are not much different (equally mediocre)
The targets are not reconstructed, but set when generating the traits. I.e. for which target to generate features. This is what is done in the picture generation examples, only there are more targets, so they are done van-hot
looked at the comparison tables of VAE and GAN. The latter is much better at generating, I think to spit on the coders. Don't have as much desire to reinvent the wheel.
Might make sense to convert to continuous, yes...but not for sure.
there is a small chance that I'm doing something wrong... but have practiced on cats before)
Z.U. even in the example with the numbers, it learns only 10 epochs, and then the error stops dropping
Autoencoders are more suitable for compressing data to lower dimensions or generating semantic vectors from it. Where GANs are more suitable for generating data
https://medium.com/lis-computer-vision-blogs/gans-cgans-ae-aae-ave-caae-cave-2e7d23255b52
I think the results are blurred because of data compression. Even if you increase the number of neurons, there is still some compression, a shift to a different distribution. The new data will always be smeared, smoothed, etc. And I don't want a smudge, I want plausible samples.
You can get exactly the same frosty look by compressing samples into PCA, then converting them into GMM, and then uncompressing them out of PCA. I did that, and it also turned out to be a fudge.
Coder is PCA + GMM in one bottle, if you draw the analogy
Has anyone tried the Relevant Vectors method?
You'll be the first
You can get exactly the same fuzzy by compressing PCA attributes, then converting them to GMM, and then decompressing from PCA. I've done it and got the same fuzzy look.
Coder is PCA + GMM in one package, if you draw the analogy.
So, leave all PCA components and you won't get smudges, you'll get a crisp picture
You'll be the first.
So leave all the PCA components and there will be no smear, there will be a clear picture
No it won't, the reverse conversion of the nageneric adds a lot of noise
or maybe not... but it looks like itit won't, it adds a lot of noise when converting back the nagenerated
I don't understand, gmm already adds noise? or what?
I do not understand, gmm already adds noise ? or what ?
I didn't think much of it, just a guess based on poking around.
gmm will give you features you haven't seen before. But similar to the ones you've seen. In a reverse transformation it might have some effect, I guess. Add some noise.
This is an assumption.