Possibly inaccurate information on dataset.

[el-dAna]

We tried to replicate your results and use your code for our own implementation but there is a little misunderstanding on your the dataset.

1. On the git report readme there are 23 subdasets but only 19 in the dataset download link. The extracted zip file yield 19. Have we missed anything?

2. The training_aug.h5 has 1000 samples with labels as 0. Each h5 file in the readme should have N labels as specified in the data structure of the readme. I guess we have misunderstood you, could you clarify a little bit?

3. For the training in the code you passed the path to the training_aug.h5 file which has only one class. Except that we missed something it will be meaningless to train with just one label. Did you therefore generate your training dataset by merging the three sub datasets? (training_aug.h5, training_0hz.h5 and training_no_aug.h5)

[gxhen]

Sorry for the late reply.

1. There are 19 subdatasets in the zip file. Please double-check the readme.
2. The 'label' in 'training_aug.h5' file should be a vector like [0 0 0 0 ... 1 1 1 1 ... 2 2 2 2 until 9 9 9 9].
3. There should be multiple labels in training_aug.h5. Please debug by setting a breakpoint or try to print it.

[el-dAna]

According to the insert the 'training_aug.h5' 'label' has 29 unique classes corresponding to the 30 devices used. If 'label' stored a vector with only 9 unique values that would mean 9 devices.

[el-dAna]

the pkt_range was also set to 30000 to load all the labels. Thank you for the support. Standing by.