search

alexivaner / Deep-Learning-Based-Radio-Signal-Classification

Final Project for AI Wireless
MIT License
89 stars 25 forks source link

Code error #1

Open autoasd opened 3 years ago

autoasd commented 3 years ago

I think your code does not solve the problem of signal modulation classification under low signal-to-noise ratio。 The reason why your code can achieve such high accuracy under low signal-to-noise ratio is because there are duplicate data in the data in your data set. code： idx = np.random.choice(range(0,sample_num),size=60000) The number of 60000 leads to duplication in your data set When you change 60000 to sample_num and do not load pre-trained weights, your results will be worse.

alexivaner commented 3 years ago

I think your code does not solve the problem of signal modulation classification under low signal-to-noise ratio。 The reason why your code can achieve such high accuracy under low signal-to-noise ratio is because there are duplicate data in the data in your data set. code： idx = np.random.choice(range(0,sample_num),size=60000) The number of 60000 leads to duplication in your data set When you change 60000 to sample_num and do not load pre-trained weights, your results will be worse.

Hi thank you for your interest in our project, actually, we should admit that in this repository, we only focused on the model and we are not experts in signals. I think that we already use randomized data for our deep learning model. Regarding the pre-trained model, of course, we need to load pre-trained weights, because that one of the purposes we create a neural network, we want to have trained weight to produce a better result.

alexivaner commented 3 years ago

If you want to use another dataset composition, I suggest you train again, but we believe that transformer at least achieves a better result than a conventional way for Low SNR signal.

wxw-cyber commented 2 years ago

Why I can't train your model with the full dataset (RML2018.01A), the code always has an ''OOM'' error.