Begin adding evaluation scores for all models, initially MUSEDB18 + model fix (#169)

* Added initial attempt at script to evaluate model performance using MUSDB18 dataset and evaluation tooling

* Added loading of existing raw scores

* WIP combined scores

* Updated script to process all tracks for all models, and evaluate SDR for all stems supported by MUSDB18. Added combined scores calculated so far.

* Added median score calculation, changed inner/outer loop for tracks vs. model processing, renamed accompaniment

* Fixed UVR-De-Reverb-aufr33-jarredou.pth model by adding new model-data dict appended to the UVR one

Author: beveradb

.gitignore

@@ -11,6 +11,8 @@
 *.wav
 *.flac
 *.mp3
+tests/model-metrics/results
+tests/model-metrics/datasets
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

audio_separator/model-data.json

@@ -0,0 +1,14 @@
+{
+    "vr_model_data": {
+        "97dc361a7a88b2c4542f68364b32c7f6": {
+            "vr_model_param": "4band_v4_ms_fullband",
+            "primary_stem": "Dry",
+            "nout": 32,
+            "nout_lstm": 128,
+            "is_karaoke": false,
+            "is_bv_model": false,
+            "is_bv_model_rebalanced": 0.0
+        }
+    },
+    "mdx_model_data": {}
+}