Extracting input of sim_ber function

[BramvBk]

Hello,
I am using Sionna and started looking into the Toward Learned Receivers example. I would like to know if it is possible to extract and save the output data of the sim_ber function used in the evolution step of the neural receiver. so I can use the exact input for a hardware implementation to use as a comparison.

Best regards,

[SebastianCa]

Hi,

in the Neural Receiver example, the sim_ber function simply runs the model multiple times and internally evaluates the BER/BLER for a given SNR (Monte-Carlo simulation).

If you want to store the exact neural receiver input/output, you need to modify the E2ESystem Model by simply returning y in the last line, i.e., return b, b_hat, y.
You can then call the new model via
b, b_hat, y = model(batch_size, ebno_db)
where b containts the ground truth bits, b_hat is the neural receiver's estimate and y is the channel output/input to the receiver. batch_size and ebno_db depend on your desired system configuration.

[BramvBk]

Thanks for your reply i am also trying to extract the exact input data going into the neural receiver so that is z = tf.concat([tf.math.real(y), tf.math.imag(y),no], axis=-1) . I thought indeed sim ber just ran the model multiple times but when i try to exports using either tf.io.write.file or write it to a diffrent variable (as z is reused for the diffrent resudial blocks) so i can return it in the last line it only seems to be called during the following part:

model(1, tf.constant(10.0, tf.float32))
with open(model_weights_path, 'rb') as f:
weights = pickle.load(f)

model.set_weights(weights)

But not when running the line

_,bler = sim_ber(model, ebno_dbs, batch_size=64, num_target_block_errors=50, max_mc_iter=50, graph_mode=None )

[SebastianCa]

Keep in mind that the model runs in Graph mode due to the @tf.function decorator. This means that any Python code in the call() function of the model will be traced only once during the first call (or if the input signature changes).

Without knowing your exact application, I would suggest to replace the sim_ber function with a customized for-loop which saves the returned values & calculates the BLER manually or you could run the model in Eager mode.

[BramvBk]

thank you for your reply, I think this is indeed the issue I am having I also noticed that sim_ber has the argument graph mode that I am able to set. (graph_mode (One of [“graph”, “xla”], str) – A string describing the execution mode of mc_fun. Defaults to None. In this case, mc_fun is executed as is.) would this allow it to be traced multiple times or is this something else. can I run in eager mode by removing the decorator? (I will read into this as I am still relatively new in training nn and using tensorflow) but thanks for pointing me toward this. how would changing the sim_ber function change the execution as you explained sim_ber just runs the model multiple times so if I do the same whit a for loop how would that change . as I imagine the python code in the call is only executed the first time. or does running the customized loop cause the input signature to change?

[SebastianCa]

You need to remove the function decorator in the model. The graph_mode parameter in the sim_ber function only allows to run an Eager function in graph (or XLA) mode. Unfortunately, the other direction does NOT work.

If the model is properly fed with input tensors for the batch_size and the snr, the signature should not change.
You can find a detailed explanation in Eager vs. graph mode for details.

[BramvBk]

thanks for your responses i will try this out