Discussion of article "Neural networks made easy (Part 3): Convolutional networks"
A lot of work has been done. A lot of time, I'm sure. What is the purpose of this series of articles?
If it is an exercise in programming on MKL5 - it is clear, you need it. Everyone can try what this beast is. Unfortunately, many may be disappointed by the low results of your implementations and give up further search in this field.
In practical terms, training time of more than 2 hours(!) is unacceptable. A lot of programs have been developed for this purpose in various languages by various large organisations. Just apply it correctly.
Good luck
A lot of work has been done. A lot of time, I'm sure. What is the purpose of this series of articles?
If it is an exercise in programming on MKL5 - it is clear, you need it. Everyone can try what this beast is. Unfortunately, many people may be disappointed by low results of your implementations and give up further search in this field.
In practical terms, training time of more than 2 hours(!) is unacceptable. A lot of programs have been developed for this purpose in various languages by various large organisations. Just apply it correctly.
Good luck
Good day, Vladimir.
The purpose of the articles is to demonstrate the possibilities and to give a tool for everyone to build their own programmes.
Training one epoch for more than 2 hours without using multithreading of video cards is quite an acceptable result. In practice, a neural network is trained on the prepared data, and we work on real data with the processing of one current state for a fraction of seconds.
I agree, the examples show a low result, but the article was not intended to give everyone a ready-made Expert Advisor with a certain profitability. For this purpose, in addition to the tool itself, a long work on neural network architecture is required, accompanied by a number of experiments with different neural network architectures.
it's not 2 hours, it's 2 days.)
what is the sample size? how many features and number of samples?
it's not 2 hours, it's 2 days.)
what is the sample size? how many features and number of samples?
The sample is 12453 candlesticks (H1 for 2 years). The network input is fed with information for 20 candlesticks with 12 signs per candlestick. The training time for one epoch is 2-2,5 hours.
Hi Dimitriy!
Thanks for this series of articles about Neural Networks! I really enjoy reading these articles (by google translation into English) as I have also implemented Neural Networks with back-propagation and different techniques of optimisation (such as Adam, RMSprop, etc.) in MQL5 language. To be honest, the MQL5 language is too slow for training non-toy problems but such implementations are still valuable. Despite the educational aspect of implementing Neural Networks, we notice that we are not limited to MQL5 programming language to train a network. We can use high performance packages to train a network and then transfer the weights (and other hyper-parameters) into our MQL5 programs to test and trade live. Please continue publishing your good work! :)
Best regards, Rasoul
Hi Dimitriy!
Thanks for this series of articles about Neural Networks! I really enjoy reading these articles (by google translation into English) as I have also implemented Neural Networks with back-propagation and different techniques of optimisation (such as Adam, RMSprop, etc.) in MQL5 language. To be honest, the MQL5 language is too slow for training non-toy problems but such implementations are still valuable. Despite the educational aspect of implementing Neural Networks, we notice that we are not limited to MQL5 programming language to train a network. We can use high performance packages to train a network and then transfer the weights (and other hyper-parameters) into our MQL5 programs to test and trade live. Please continue publishing your good work! :)
Best regards, Rasoul
Thanks.
The article is certainly interesting. Especially in terms of the fact that everything is implemented in MQL and no libraries are required. I hope that there will be a sequel. Unfortunately, testing of the Expert Advisor has not yet yielded positive results.
New article Neural networks made easy (Part 3): Convolutional networks has been published:
Author: Dmitriy Gizlyk
Hello. Thanks for this. I have a question for you. How do you get the pixel values for the chart or image? I have looked but can't seem to find an explanation on how to derive the pixel values of the chart to use in the input convolutional layer.
Hello, you right. The convolution layer doesn't take pixels from chart. My idea was else. I take data from differnet indicators and historical price and put it to input of neural network. The conwolution layer looks patterns of this data and returns some value for every candle.
- Free trading apps
- Over 8,000 signals for copying
- Economic news for exploring financial markets
You agree to website policy and terms of use
New article Neural networks made easy (Part 3): Convolutional networks has been published:
As a continuation of the neural network topic, I propose considering convolutional neural networks. This type of neural network are usually applied to analyzing visual imagery. In this article, we will consider the application of these networks in the financial markets.
Testing was performed using the EURUSD pair with the H1 timeframe. Two Expert Advisors, one with a convolutional neural network and the other one with fully connected network, were launched simultaneously on different charts of the same symbol, in the same terminal. The parameters of the fully connected layers of the convolutional neural network match the parameters of the fully connected network of the second Expert Advisor, i. e. we have simply added convolutional and subsampled layers to a previously created network.
Visually, you can see that signals appear less frequently on the convolutional neural network chart, but they are closer to the target.
Author: Dmitriy Gizlyk