There are couple of options out there for Deep Learning. Can somebody give me advise with these parameters?
- I would train a sample set on the scale of 200.000-2 million short audio samples, 14 label classes. (Thus the complexity is much lower than training images.)
- What do you estimate about the memory requirements of a built DL model in RAM with the above parameters? It is 1-2MB, 10-20MB or more?
- I use a single PC, Linux environment, but I can install Windows also if needed and I would buy a dedicated GPU to speed up the training process.
- If I understood right, the ideas in Deep Learning are on the training side and the resulting neural networks are still neural networks which are easy to simulate. I don't care too much about the training software, but I must be able to export the built model in csv or other easy-to-parse text format. In the target embedded environment, I have to cross-compile my own or open-source C++ sources to load and simulate the model. Using R, Java, python or other languages is out of question in the simulation. In my ideal world, I can load the result models with the OpenCV's Machine Learning module for simulation, but I can write a loader/simulator in C++ as well.
Dear Csaba,
Some answeres that might help:
- you should definietly make your research on GPU.
- I assume your purpose is classification with labelled data. For this purpose you should apply CNN + FC layer on the top. If you have unlabelled data, you may try to use CNN auto-encoders. The size of the model highly depend on the architecture.
- if application is important you can use Torch7 deep learning framework. It is written in LUA. LUA is a lightweight language over C++ and can be compiled to binary targets.
Best Regards, üdvözlettel,
Balint
Just a few pointers, in case you don't already know.
Audio data is temporal, so it is suggested to use a recurrent neural network (RNN) to do what you want to do.
I would also strongly recommend the Keras deep learning framework to do this in Python, seeing as you have no prior experience. It comes with a few examples of how to train RNNs effectively, so it's quick to pick up!
RNN can be a good soltuion, however 2D CNN applied to the spectrum is very efficient in speech recognition and in music classification as well.
Combining CNNs (feature learning) with RNNs (modeling temporal structure) may lead even to better results.
- Badges
- Science topic