Added balanced dataset, fixed generation bug

degradation/balanced_dataset.json

@@ -0,0 +1,57 @@
+{
+	"name":		"BalancedDataset",
+	"train": 	0.75,
+	"test":		0.25,
+	"weights": {
+		"add-white":		0.5,
+		"add-aircon":		0.5,
+		"add-hum":			0.5,
+		"add-infra":		0.5,
+		"clipping-hard":	0.5,
+		"clipping-soft":	0.5,
+		"mute":				0.5
+	},
+	"pipeline": [
+		"pad-1",
+		"add-white",
+		"add-aircon",
+		"add-hum",
+		"add-infra",
+		"mute",
+		"clipping-soft",
+		"clipping-hard",
+		"normalize-rand"
+	],
+	"incompatible": [
+		["clipping-soft", "clipping-hard", "normalize-rand"]
+	],
+	"operations": [
+		{
+			"name": "pad-1",
+			"degradations": [
+				{
+					"name": "pad",
+					"before": 1,
+					"after": 1
+				}
+			]
+		},
+		{
+			"name": "normalize",
+			"degradations": [
+				{
+					"name": "normalize"
+				}
+			]
+		},
+		{
+			"name": "normalize-rand",
+			"degradations": [
+				{
+					"name": "normalize",
+					"maxAmplitude": {"randomRange": [0.5, 0.999]}
+				}
+			]
+		}
+	]
+}

degradation/create_dataset.py

@@ -151,6 +151,12 @@ def create(args):
 			print(f"Noise class {nc_id} not specified in {args.noise_classes}", file=sys.stderr)
 			m_eng.cancel()
 			sys.exit(1)
+	if "noise_class_degradations" in spec:
+		print("Warning: when using an existing dataset as the specification,", file=sys.stderr)
+		print("the degradations are still taken from the current noise class file:", file=sys.stderr)
+		print(args.noise_classes, file=sys.stderr)
+		print("The 'noise_class_degradations' field in the specifications of the", file=sys.stderr)
+		print("two datasets may differ as a result.", file=sys.stderr, flush=True)
 
 	# Check that the pipeline is correct
 	for _id in spec["pipeline"]:
@@ -216,8 +222,8 @@ def create(args):
 		for labels in (labels_train, labels_test):
 			idxs, = ((labels & combo) == combo).all(axis=1).nonzero()
 			for i in range(len(combo_weights)):
-				s = slice(None if i == 0 else int(combo_weights[i-1] * len(labels)),
-				          None if i == len(combo_weights) - 1 else int(combo_weights[i] * len(labels)))
+				s = slice(None if i == 0 else int(combo_weights[i-1] * len(idxs)),
+				          None if i == len(combo_weights) - 1 else int(combo_weights[i] * len(idxs)))
 				# Replace with only one of the incompatible labels, proportional to how common each label should be
 				labels[idxs[s]] &= ~combo | (combo.cumsum() == i + 1)
 
@@ -325,6 +331,11 @@ def create(args):
 
 		print("Done")
 
+	# Append used noise class degradations to specification and save it
+	spec["noise_class_degradations"] = dict()
+	for nc_id in spec["weights"].keys():
+		spec["noise_class_degradations"][nc_id] = noise_classes[nc_id].degradations
+
 	with open(os.path.join(dataset_folder, "source.json"), "w") as f:
 		json.dump(spec, f, indent='\t')
 	print("Wrote source.json")