[Discuss-gnuradio] gr-inspector & RadioML model issue

[Christopher Richardson]

Hi,

I'm just wondering if anyone might have any ideas, why I'm getting 2PSK recognised

as QAM64 when using the RadioML model & data (I'm assuming I'm doing something

silly, but can't for the life of me work out what).

(You will need TensorFlow 1.0 and Keras)

If you download the TF model from - https://www.dropbox.com/s/7lbm8q6w4vnxbks/cnn.zip?dl=1

Which was trained for around 25 epochs.  And place this model in /tmp.

You can then test the accuracy of the model using cnn_test.py in:
https://github.com/chrisruk/models

(You will first need to download https://radioml.com/ds/2016.10/RML2016.10a_dict.dat.bz2 and place RML2016.10a_dict.dat in the models directory)

It should display something like the following:
Overall Accuracy:  0.0980786825252
Overall Accuracy:  0.0959128065395
Overall Accuracy:  0.0987698986975
Overall Accuracy:  0.11241429087
Overall Accuracy:  0.139106654512
Overall Accuracy:  0.210414074996
Overall Accuracy:  0.304243983098
Overall Accuracy:  0.467925219941
Overall Accuracy:  0.598706099815
Overall Accuracy:  0.660195936139
Overall Accuracy:  0.701094890511
Overall Accuracy:  0.708106143221
Overall Accuracy:  0.725403817915
Overall Accuracy:  0.724187725632
Overall Accuracy:  0.724977457169
Overall Accuracy:  0.725683060109
Overall Accuracy:  0.715372080391
Overall Accuracy:  0.719689979701
Overall Accuracy:  0.713665158371
Overall Accuracy:  0.717371737174

Then if you install my version of gr-inspector (which has an extra python file to test and will require TensorFlow 1.0):
https://github.com/chrisruk/gr-inspector

If you run:
examples/amc_cnn.grc

In GRC, you will note PSK seems to be recognised as QAM64.

In the root directory of my gr-inspector there is also 'amc_cnn_channel.py'

which uses the same channel model as the RadioML dataset, but also still doesn't seem to recognise PSK.

Any ideas would be most appreciated!

Cheers

Chris