Missing parameter causing errors in NT3 (P1B4)

[vgutta]

While running Pilot 1 NT3 using the command python nt3_baseline_keras2.py --conf nt3_perf_bench_model.txt I ran into this error caused by a missing parameter

Traceback (most recent call last): File "nt3_baseline_keras2.py", line 290, in <module> main() File "nt3_baseline_keras2.py", line 286, in main run(gParameters) File "nt3_baseline_keras2.py", line 101, in run X_train, Y_train, X_test, Y_test = load_data(train_file, test_file, gParameters) File "nt3_baseline_keras2.py", line 70, in load_data if gParameters['add_noise']: KeyError: 'add_noise'

Issue is caused by these lines

Benchmarks/Pilot1/NT3/nt3_baseline_keras2.py

Lines 68 to 82 in a48c85a

	# TODO: Add better names for noise boolean, make a featue for both RNA seq and label noise together
	# check if noise is on (this is for label)
	if gParameters['add_noise']:
	# check if we want noise correlated with a feature
	if gParameters['noise_correlated']:
	Y_train, y_train_noise_gen = candle.label_flip_correlated(Y_train,
	gParameters['label_noise'], X_train,
	gParameters['feature_col'],
	gParameters['feature_threshold'])
	# else add uncorrelated noise
	else:
	Y_train, y_train_noise_gen = candle.label_flip(Y_train, gParameters['label_noise'])
	# check if noise is on for RNA-seq data
	elif gParameters['noise_gaussian']:
	X_train = candle.add_gaussian_noise(X_train, 0, gParameters['std_dev'])

It seems like the candle parser being used never includes the parameters being checked here

Benchmarks/Pilot1/NT3/nt3_baseline_keras2.py

Lines 30 to 31 in a48c85a

	# Initialize parameters
	gParameters = candle.finalize_parameters(nt3Bmk)

So is there a different way to run this (maybe different flags) to avoid this issue. Obviously commented out lines 68-82 in nt3_baseline_keras2.py works but was not sure if parameters such as 'add_noise' will ever make it through to NT3. If not then maybe commenting out these lines permanently will save others some trouble?

[jmohdyusof]

OK, you need to add the lines:
add_noise = False
noise_gaussian = False
to the nt3_perf_bench_model.txt.

All the keywords are parsed from the config file, so if they are not there they will not appear in gParameters and will cause an error. Adding those 2 lines is sufficient.

Alternatively, just run the benchmark with 'python nt3_baseline_keras2.py' and it will use the nt3_default_model.txt which contains the required keywords.

[vgutta]

Actually I tried without any changes to the repo on a different system and I did not run into any of those issues. I'll close this for now but I'll reply if I find the reason why it did that on the first system. Thanks for your help @jmohdyusof

[jmohdyusof]

I can't see how it would run correctly without those keywords, unless you are using the default model file. We will push changes to master to fix the perf_bench_model

[vgutta]

You are right, I ran the default model file this time. Sorry for the confusion

[jmohdyusof]

OK, the original answer still stands. The perf_bench_model file was not updated to include the keywords needed by the noise methods, so we will push an update to master to fix that.