zenodo-B

.gitattributes

@@ -1,5 +0,0 @@
-*.wav filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-model/mixer.ckpt filter=lfs diff=lfs merge=lfs -text
-model/mixer_val.ckpt filter=lfs diff=lfs merge=lfs -text

README.md

@@ -11,37 +11,38 @@ The separated stems each have a different <a href="https://ws-choi.github.io/per
 
 ## Installation
 
-Setup [git-lfs](https://git-lfs.github.com/) first. You will need it to download the models inside this repository.
-You'd also need [conda](https://docs.conda.io/en/latest/miniconda.html).
-
-After all those are installed, clone this branch:
+Set up [conda](https://docs.conda.io/en/latest/miniconda.html) first. After it's installed, clone this branch:
 
 ```bash
 git clone -b leaderboard_B https://github.com/kuielab/mdx-net-submission.git
 ```
 
+### Linux
+
 In the cloned repository directory, do
 
 ```bash
 conda env create -f environment.yml -n mdx-submit
 conda activate mdx-submit
 pip install -r requirements.txt
 python download_demucs.py
+wget https://zenodo.org/record/5717356/files/onnx_B.zip
+unzip onnx_B
 ```
 
-Every time when you open a new terminal, conda will default to environment `base`.
-Just do 
+### Windows
+
+In the cloned repository directory, using the conda powershell prompt:
 
 ```bash
+conda env create -f environment.yml -n mdx-submit
 conda activate mdx-submit
+pip install -r requirements.txt
+python download_demucs.py
+Invoke-WebRequest -Uri https://zenodo.org/record/5717356/files/onnx_B.zip -OutFile onnx_B.zip
+Expand-Archive onnx_B.zip -DestinationPath .
 ```
 
-to go back into the environment you have installed MDX's dependencies in.
-
-## Custom models
-
-For custom models (such as the [higher quality vocal model trained by UVR team](https://github.com/Anjok07/ultimatevocalremovergui/releases/tag/MDX-Net-B)), please replace the relevant models in `./onnx/`.
-
 ## Usage
 
 After successful installation, you can put the songs you wish to separate as `./data/test/SONGNAME/mixture.wav`, and run either `run.sh` or
@@ -50,4 +51,13 @@ After successful installation, you can put the songs you wish to separate as `./
 python predict_blend.py
 ```
 
-After the separation completes, the results will be saved in `./data/results/baseline/SONGNAME/`.
+After the separation completes, the results will be saved in `./data/results/kuielab_mdxnet_A/SONGNAME/`.
+
+Also, every time when you open a new terminal / conda prompt, conda will default to environment `base`.
+Just do 
+
+```bash
+conda activate mdx-submit
+```
+
+to go back into the environment you have installed MDX's dependencies in.

aicrowd.json

@@ -4,6 +4,6 @@
   "authors": [
       "kim_min_seok"
   ],
-  "description": "",
+  "description": "KUIELAB-MDX-Net-B",
   "external_dataset_used": true
 }

evaluator/music_demixing.py

@@ -24,16 +24,17 @@ def time_limit(seconds):
     def signal_handler(signum, frame):
         raise TimeoutException("Prediction timed out!")
 
-    signal.signal(signal.SIGALRM, signal_handler)
-    signal.alarm(seconds)
-    try:
-        yield
-    finally:
-        signal.alarm(0)
+    # signal.signal(signal.SIGALRM, signal_handler)
+    # signal.alarm(seconds)
+    # try:
+    #     yield
+    # finally:
+    #     signal.alarm(0)
+    yield
 
 
 class MusicDemixingPredictor:
-    def __init__(self, model_name='baseline'):
+    def __init__(self, model_name='kuielab_mdxnet_B'):
         self.test_data_path = os.getenv("TEST_DATASET_PATH", os.getcwd() + "/data/test/")
         self.results_data_path = os.getenv("RESULTS_DATASET_PATH", os.getcwd() + "/data/results/" + model_name)
         self.inference_setup_timeout = int(os.getenv("INFERENCE_SETUP_TIMEOUT_SECONDS", "900"))

model/.directory

@@ -1 +0,0 @@
-

models.py

@@ -5,7 +5,7 @@
 
 dim_c = 4
 k = 3
-model_path = 'model'
+model_path = 'data/weights'
 n_fft_scale = {'bass': 8, 'drums':2, 'other':4, 'vocals':3, '*':2}
 
 

predict_blend.py

@@ -11,23 +11,23 @@
 device = torch.device('cpu')
 
 class Predictor(MusicDemixingPredictor):
-        
+
     def prediction_setup(self):
         self.models = get_models(model_name, load=False, device=device)
         self.demucs = Demucs(sources=["drums", "bass", "other", "vocals"], channels=48 if '48' in demucs_name else 64)
         self.demucs.load_state_dict(torch.load(f'model/{demucs_name}.ckpt'))
         self.demucs.eval()
 
     def prediction(self, mixture_file_path, bass_file_path, drums_file_path, other_file_path, vocals_file_path):
-        file_paths = [bass_file_path, drums_file_path, other_file_path, vocals_file_path]      
-        mix, rate = sf.read(mixture_file_path)
-        sources = self.demix(mix.T)
+        file_paths = [bass_file_path, drums_file_path, other_file_path, vocals_file_path]
+        sources = self.demix(mixture_file_path)
         for i in range(len(sources)):
-            sf.write(file_paths[i], sources[i].T, rate)
+            sf.write(file_paths[i], sources[i].T, samplerate=44100)
 
-    def demix(self, mix):
+    def demix(self, mix_path):
+        mix = sf.read(mix_path)[0].T
         demucs_out = self.demix_demucs(mix)
-        base_out = self.demix_base(mix)        
+        base_out = self.demix_base(mix)
         sources = base_out * b + demucs_out * (1-b)     
         return sources
 
@@ -36,36 +36,35 @@ def demix_base(self, mix):
         sources = []
         n_sample = mix.shape[1]
         for model in self.models:
-            trim = model.n_fft//2
-            gen_size = model.chunk_size-2*trim
-            pad = gen_size - n_sample%gen_size
-            mix_p = np.concatenate((np.zeros((2,trim)), mix, np.zeros((2,pad)), np.zeros((2,trim))), 1)
+            trim = model.n_fft // 2
+            gen_size = model.chunk_size - 2 * trim
+            pad = gen_size - n_sample % gen_size
+            mix_p = np.concatenate((np.zeros((2, trim)), mix, np.zeros((2, pad)), np.zeros((2, trim))), 1)
 
             mix_waves = []
             i = 0
             while i < n_sample + pad:
-                waves = np.array(mix_p[:, i:i+model.chunk_size])
+                waves = np.array(mix_p[:, i:i + model.chunk_size])
                 mix_waves.append(waves)
-                i += gen_size            
-            mix_waves = torch.tensor(mix_waves, dtype=torch.float32)    
+                i += gen_size
+            mix_waves = torch.tensor(mix_waves, dtype=torch.float32)
 
             with torch.no_grad():
-                _ort = ort.InferenceSession(f'onnx/{model.target_name}.onnx')
+                _ort = ort.InferenceSession(f'{onnx_name}/{model.target_name}.onnx')
                 tar_waves = model.istft(torch.tensor(
                     _ort.run(None, {'input': model.stft(mix_waves).numpy()})[0]
                 ))
-                tar_signal = tar_waves[:,:,trim:-trim].transpose(0,1).reshape(2, -1).numpy()[:, :-pad]
-
+                tar_signal = tar_waves[:, :, trim:-trim].transpose(0, 1).reshape(2, -1).numpy()[:, :-pad]
             sources.append(tar_signal)
         print(time()-start_time)
         return np.array(sources)
 
     def demix_demucs(self, mix):
         start_time = time()
         mix = torch.tensor(mix, dtype=torch.float32)
-        ref = mix.mean(0)        
+        ref = mix.mean(0)
         mix = (mix - ref.mean()) / ref.std()
-        
+
         with torch.no_grad():
             sources = apply_model(self.demucs, mix, split=True, overlap=0.5)
 
@@ -77,6 +76,7 @@ def demix_demucs(self, mix):
 
 model_name = 'tdf_extra'
 demucs_name = 'demucs_extra'
+onnx_name = 'onnx_B'
 
 b = np.array([[[0.5]], [[0.5]], [[0.7]], [[0.9]]])