add inst.py

args.py

@@ -0,0 +1,419 @@
+import os
+
+import yaml
+from addict import Dict
+
+from ml_collections import ConfigDict
+
+from UVR import DENOISER_MODEL_PATH, DEVERBER_MODEL_PATH, MDX_C_CONFIG_PATH
+from gui_data.constants import DEMUCS_2_SOURCE, DEMUCS_2_SOURCE_MAPPER, VR_ARCH_TYPE, MDX_ARCH_TYPE, DEMUCS_ARCH_TYPE
+
+vr_cache_source_mapper = {}
+mdx_cache_source_mapper = {}
+demucs_cache_source_mapper = {}
+
+
+def cached_source_callback(process_method, model_name=None):
+    model, sources = None, None
+
+    if process_method == VR_ARCH_TYPE:
+        mapper = vr_cache_source_mapper
+    if process_method == MDX_ARCH_TYPE:
+        mapper = mdx_cache_source_mapper
+    if process_method == DEMUCS_ARCH_TYPE:
+        mapper = demucs_cache_source_mapper
+
+    for key, value in mapper.items():
+        if model_name in key:
+            model = key
+            sources = value
+
+    return model, sources
+
+
+def cached_model_source_holder(self, process_method, sources, model_name=None):
+    if process_method == VR_ARCH_TYPE:
+        self.vr_cache_source_mapper = {**vr_cache_source_mapper, **{model_name: sources}}
+    if process_method == MDX_ARCH_TYPE:
+        self.mdx_cache_source_mapper = {**mdx_cache_source_mapper, **{model_name: sources}}
+    if process_method == DEMUCS_ARCH_TYPE:
+        self.demucs_cache_source_mapper = {**demucs_cache_source_mapper, **{model_name: sources}}
+
+
+config_path = os.path.join(MDX_C_CONFIG_PATH, 'model_2_stem_full_band_8k.yaml')
+with open(config_path) as f:
+    config = ConfigDict(yaml.load(f, Loader=yaml.FullLoader))
+
+mdx23c_8kfft_instvoc_hq_model_data = {
+    'DENOISER_MODEL': DENOISER_MODEL_PATH,
+    'DEVERBER_MODEL': DEVERBER_MODEL_PATH,
+    'all_models': [],
+    'bv_model_rebalance': 0,
+    'chunks': 0,
+    'compensate': None,
+    'demucs_4_stem_added_count': 0,
+    'demucs_source_list': [],
+    'demucs_stem_count': 0,
+    'demucs_stems': 'All Stems',
+    'deverb_vocal_opt': 'Vocals',
+    'device_set': 'Default',
+    'ensemble_primary_stem': 'Vocals',
+    'ensemble_secondary_stem': 'Instrumental',
+    'is_4_stem_ensemble': False,
+    'is_bv_model': False,
+    'is_change_def': False,
+    'is_demucs_4_stem_secondaries': False,
+    'is_demucs_combine_stems': True,
+    'is_demucs_pre_proc_model_inst_mix': False,
+    'is_denoise': False,
+    'is_denoise_model': False,
+    'is_deverb_vocals': False,
+    'is_dry_check': False,
+    'is_ensemble_mode': True,
+    'is_get_hash_dir_only': False,
+    'is_gpu_conversion': 0,
+    'is_inst_only_voc_splitter': False,
+    'is_invert_spec': False,
+    'is_karaoke': False,
+    'is_match_frequency_pitch': True,
+    'is_mdx_c': True,
+    'is_mdx_c_seg_def': False,
+    'is_mdx_ckpt': False,
+    'is_mdx_combine_stems': True,
+    'is_mixer_mode': False,
+    'is_multi_stem_ensemble': False,
+    'is_normalization': False,
+    'is_pitch_change': False,
+    'is_pre_proc_model': False,
+    'is_primary_model_primary_stem_only': False,
+    'is_primary_model_secondary_stem_only': False,
+    'is_primary_stem_only': False,
+    'is_save_inst_vocal_splitter': False,
+    'is_save_vocal_only': False,
+    'is_sec_bv_rebalance': False,
+    'is_secondary_model': False,
+    'is_secondary_model_activated': False,
+    'is_secondary_stem_only': False,
+    'is_use_opencl': False,
+    'is_vocal_split_model': False,
+    'is_vocal_split_model_activated': False,
+    'is_vr_51_model': False,
+    'manual_download_Button': None,
+    'margin': 44100,
+    'mdx_batch_size': 1,
+    'mdx_c_configs': config,
+    'mdx_dim_f_set': None,
+    'mdx_dim_t_set': None,
+    'mdx_model_stems': ['Vocals', 'Instrumental'],
+    'mdx_n_fft_scale_set': None,
+    'mdx_segment_size': 256,
+    'mdx_stem_count': 2,
+    'mdxnet_stem_select': 'Vocals',
+    'mixer_path': '/Users/taoluo/Workspace/github/ultimatevocalremovergui/models/MDX_Net_Models/mixer_val.ckpt',
+    'model_and_process_tag': 'MDX-Net: MDX23C-InstVoc HQ',
+    'model_basename': 'MDX23C-8KFFT-InstVoc_HQ',
+    'model_capacity': (32, 128),
+    'model_data': {'config_yaml': 'model_2_stem_full_band_8k.yaml'},
+    'model_hash': '99b6ceaae542265a3b6d657bf9fde79f',
+    'model_hash_dir': '/Users/taoluo/Workspace/github/ultimatevocalremovergui/models/MDX_Net_Models/model_data/99b6ceaae542265a3b6d657bf9fde79f.json',
+    'model_name': 'MDX23C-InstVoc HQ',
+    'model_path': '/Users/taoluo/Workspace/github/ultimatevocalremovergui/models/MDX_Net_Models/MDX23C-8KFFT-InstVoc_HQ.ckpt',
+    'model_samplerate': 44100,
+    'model_status': True,
+    'mp3_bit_set': '320k',
+    'overlap': 0.25,
+    'overlap_mdx': 'Default',
+    'overlap_mdx23': 8,
+    'pre_proc_model': None,
+    'pre_proc_model_activated': False,
+    'primary_model_primary_stem': None,
+    'primary_stem': 'Vocals',
+    'primary_stem_native': None,
+    'process_method': 'MDX-Net',
+    'save_format': 'WAV',
+    'secondary_model': None,
+    'secondary_model_4_stem': [],
+    'secondary_model_4_stem_model_names_list': [],
+    'secondary_model_4_stem_names': [],
+    'secondary_model_4_stem_scale': [],
+    'secondary_model_bass': None,
+    'secondary_model_drums': None,
+    'secondary_model_other': None,
+    'secondary_model_scale': None,
+    'secondary_model_scale_bass': None,
+    'secondary_model_scale_drums': None,
+    'secondary_model_scale_other': None,
+    'secondary_stem': 'Instrumental',
+    'semitone_shift': 0.0,
+    'vocal_split_model': None,
+    'wav_type_set': 'PCM_16'}
+
+mdx23c_8kfft_instvoc_hq_process_data = {
+    'model_name': 'MDX23C Model: MDX23C-InstVoc HQ',
+    'model_data': Dict(mdx23c_8kfft_instvoc_hq_model_data),
+    'export_path': '/Users/taoluo/Downloads/Ensembled_Outputs_1702201464',
+    'audio_file_base': '1_test_audio_MDX23C-8KFFT-InstVoc_HQ',
+    'audio_file': '/Users/taoluo/Downloads/test_audio.mp3',
+
+    'set_progress_bar': lambda step, inference_iterations=0: print(
+        f"iteration {inference_iterations} of step #{step}"),
+    'write_to_console': lambda progress_text, base_text='': print(
+        f"{progress_text} {base_text}"),
+    'process_iteration': lambda iteration: iteration + 1,
+    'cached_source_callback': cached_source_callback,
+    'cached_model_source_holder': cached_model_source_holder,
+
+    'list_all_models': ['MDX23C-8KFFT-InstVoc_HQ', 'UVR-MDX-NET-Voc_FT',
+                        'htdemucs_ft'],
+    'is_ensemble_master': True,
+    'is_4_stem_ensemble': False}
+
+uvr_mdx_net_voc_ft_model_data = {
+    'DENOISER_MODEL': DENOISER_MODEL_PATH,
+    'DEVERBER_MODEL': DEVERBER_MODEL_PATH,
+    'all_models': [],
+    'bv_model_rebalance': 0,
+    'chunks': 0,
+    'compensate': 1.021,
+    'demucs_4_stem_added_count': 0,
+    'demucs_source_list': [],
+    'demucs_stem_count': 0,
+    'demucs_stems': 'All Stems',
+    'deverb_vocal_opt': 'Vocals',
+    'device_set': 'Default',
+    'ensemble_primary_stem': 'Vocals',
+    'ensemble_secondary_stem': 'Instrumental',
+    'is_4_stem_ensemble': False,
+    'is_bv_model': False,
+    'is_change_def': False,
+    'is_demucs_4_stem_secondaries': False,
+    'is_demucs_combine_stems': True,
+    'is_demucs_pre_proc_model_inst_mix': False,
+    'is_denoise': False,
+    'is_denoise_model': False,
+    'is_deverb_vocals': False,
+    'is_dry_check': False,
+    'is_ensemble_mode': True,
+    'is_get_hash_dir_only': False,
+    'is_gpu_conversion': 0,
+    'is_inst_only_voc_splitter': False,
+    'is_invert_spec': False,
+    'is_karaoke': False,
+    'is_match_frequency_pitch': True,
+    'is_mdx_c': False,
+    'is_mdx_c_seg_def': False,
+    'is_mdx_ckpt': False,
+    'is_mdx_combine_stems': True,
+    'is_mixer_mode': False,
+    'is_multi_stem_ensemble': False,
+    'is_normalization': False,
+    'is_pitch_change': False,
+    'is_pre_proc_model': False,
+    'is_primary_model_primary_stem_only': False,
+    'is_primary_model_secondary_stem_only': False,
+    'is_primary_stem_only': False,
+    'is_save_inst_vocal_splitter': False,
+    'is_save_vocal_only': False,
+    'is_sec_bv_rebalance': False,
+    'is_secondary_model': False,
+    'is_secondary_model_activated': False,
+    'is_secondary_stem_only': False,
+    'is_use_opencl': False,
+    'is_vocal_split_model': False,
+    'is_vocal_split_model_activated': False,
+    'is_vr_51_model': False,
+    'manual_download_Button': None,
+    'margin': 44100,
+    'mdx_batch_size': 1,
+    'window_size': 512,
+    'mdx_c_configs': None,
+    'mdx_dim_f_set': 3072,
+    'mdx_dim_t_set': 8,
+    'mdx_model_stems': [],
+    'mdx_n_fft_scale_set': 7680,
+    'mdx_segment_size': 256,
+    'mdx_stem_count': 1,
+    'mdxnet_stem_select': 'All Stems',
+    'mixer_path': '/Users/taoluo/Workspace/github/ultimatevocalremovergui/models/MDX_Net_Models/mixer_val.ckpt',
+    'model_and_process_tag': 'MDX-Net: UVR-MDX-NET-Voc_FT',
+    'model_basename': 'UVR-MDX-NET-Voc_FT',
+    'model_capacity': (32, 128),
+    'model_data': {'compensate': 1.021,
+                   'mdx_dim_f_set': 3072,
+                   'mdx_dim_t_set': 8,
+                   'mdx_n_fft_scale_set': 7680,
+                   'primary_stem': 'Vocals'},
+    'model_hash': '77d07b2667ddf05b9e3175941b4454a0',
+    'model_hash_dir': '/Users/taoluo/Workspace/github/ultimatevocalremovergui/models/MDX_Net_Models/model_data/77d07b2667ddf05b9e3175941b4454a0.json',
+    'model_name': 'UVR-MDX-NET-Voc_FT',
+    'model_path': '/Users/taoluo/Workspace/github/ultimatevocalremovergui/models/MDX_Net_Models/UVR-MDX-NET-Voc_FT.onnx',
+    'model_samplerate': 44100,
+    'model_status': True,
+    'mp3_bit_set': '320k',
+    'overlap': 0.25,
+    'overlap_mdx': 'Default',
+    'overlap_mdx23': 8,
+    'pre_proc_model': None,
+    'pre_proc_model_activated': False,
+    'primary_model_primary_stem': None,
+    'primary_stem': 'Vocals',
+    'primary_stem_native': 'Vocals',
+    'process_method': 'MDX-Net',
+    'save_format': 'WAV',
+    'secondary_model': None,
+    'secondary_model_4_stem': [],
+    'secondary_model_4_stem_model_names_list': [],
+    'secondary_model_4_stem_names': [],
+    'secondary_model_4_stem_scale': [],
+    'secondary_model_bass': None,
+    'secondary_model_drums': None,
+    'secondary_model_other': None,
+    'secondary_model_scale': None,
+    'secondary_model_scale_bass': None,
+    'secondary_model_scale_drums': None,
+    'secondary_model_scale_other': None,
+    'secondary_stem': 'Instrumental',
+    'semitone_shift': 0.0,
+    'vocal_split_model': None,
+    'wav_type_set': 'PCM_16'}
+
+uvr_mdx_net_voc_ft_process_data = {
+    'model_name': 'MDX-Net Model: UVR-MDX-NET Voc FT',
+    'model_data': Dict(uvr_mdx_net_voc_ft_model_data),
+    'export_path': '/Users/taoluo/Downloads/Ensembled_Outputs_1702201464',
+    'audio_file_base': '1_test_audio_UVR-MDX-NET-Voc_FT',
+    'audio_file': '/Users/taoluo/Downloads/test_audio.mp3',
+    'set_progress_bar': lambda step, inference_iterations=0: print(
+        f"iteration {inference_iterations} of step #{step}"),
+    'write_to_console': lambda progress_text, base_text='base_text': print(
+        f"{progress_text} {base_text}"),
+    'process_iteration': lambda iteration: iteration + 1,
+    'cached_source_callback': cached_source_callback,
+    'cached_model_source_holder': cached_model_source_holder,
+    'list_all_models': ['MDX23C-8KFFT-InstVoc_HQ', 'UVR-MDX-NET-Voc_FT', 'htdemucs_ft'],
+    'is_ensemble_master': True,
+    'is_4_stem_ensemble': False}
+
+htdemucs_ft_model_data = {
+    'DENOISER_MODEL': DENOISER_MODEL_PATH,
+    'DEVERBER_MODEL': DEVERBER_MODEL_PATH,
+    'all_models': [],
+    'bv_model_rebalance': 0,
+    'chunks_demucs': 0,
+    'compensate': None,
+    'demucs_4_stem_added_count': 0,
+    'demucs_source_list': DEMUCS_2_SOURCE,
+    'demucs_source_map': DEMUCS_2_SOURCE_MAPPER,
+    'demucs_stem_count': 2,
+    'demucs_stems': 'All Stems',
+    'demucs_version': 'v4',
+    'deverb_vocal_opt': 'Vocals',
+    'device_set': 'Default',
+    'ensemble_primary_stem': 'Vocals',
+    'ensemble_secondary_stem': 'Instrumental',
+    'is_4_stem_ensemble': False,
+    'is_bv_model': False,
+    'is_change_def': False,
+    'is_chunk_demucs': False,
+    'is_demucs_4_stem_secondaries': False,
+    'is_demucs_combine_stems': True,
+    'is_demucs_pre_proc_model_inst_mix': False,
+    'is_denoise': False,
+    'is_denoise_model': False,
+    'is_deverb_vocals': False,
+    'is_dry_check': False,
+    'is_ensemble_mode': True,
+    'is_get_hash_dir_only': False,
+    'is_gpu_conversion': 0,
+    'is_inst_only_voc_splitter': False,
+    'is_invert_spec': False,
+    'is_karaoke': False,
+    'is_match_frequency_pitch': True,
+    'is_mdx_c': False,
+    'is_mdx_c_seg_def': False,
+    'is_mdx_ckpt': False,
+    'is_mdx_combine_stems': True,
+    'is_mixer_mode': False,
+    'is_multi_stem_ensemble': False,
+    'is_normalization': False,
+    'is_pitch_change': False,
+    'is_pre_proc_model': False,
+    'is_primary_model_primary_stem_only': False,
+    'is_primary_model_secondary_stem_only': False,
+    'is_primary_stem_only': False,
+    'is_save_inst_vocal_splitter': False,
+    'is_save_vocal_only': False,
+    'is_sec_bv_rebalance': False,
+    'is_secondary_model': False,
+    'is_secondary_model_activated': False,
+    'is_secondary_stem_only': False,
+    'is_split_mode': True,
+    'is_use_opencl': False,
+    'is_vocal_split_model': False,
+    'is_vocal_split_model_activated': False,
+    'is_vr_51_model': False,
+    'manual_download_Button': None,
+    'margin_demucs': 44100,
+    'mdx_batch_size': 1,
+    'mdx_c_configs': None,
+    'mdx_dim_f_set': None,
+    'mdx_dim_t_set': None,
+    'mdx_model_stems': [],
+    'mdx_n_fft_scale_set': None,
+    'mdx_stem_count': 1,
+    'mdxnet_stem_select': 'All Stems',
+    'mixer_path': '/Users/taoluo/Workspace/github/ultimatevocalremovergui/lib_v5/mixer.ckpt',
+    'model_and_process_tag': 'Demucs: v4 | htdemucs_ft',
+    'model_basename': 'htdemucs_ft',
+    'model_capacity': (32, 128),
+    'model_hash_dir': None,
+    'model_name': 'v4 | htdemucs_ft',
+    'model_path': '/Users/taoluo/Workspace/github/ultimatevocalremovergui/models/Demucs_Models/v3_v4_repo/htdemucs_ft.yaml',
+    'model_samplerate': 44100,
+    'model_status': True,
+    'mp3_bit_set': '320k',
+    'overlap': 0.25,
+    'overlap_mdx': 'Default',
+    'overlap_mdx23': 8,
+    'pre_proc_model': None,
+    'pre_proc_model_activated': False,
+    'primary_model_primary_stem': None,
+    'primary_stem': None,
+    'primary_stem_native': None,
+    'process_method': 'Demucs',
+    'save_format': 'WAV',
+    'secondary_model': None,
+    'secondary_model_4_stem': [],
+    'secondary_model_4_stem_model_names_list': [],
+    'secondary_model_4_stem_names': [],
+    'secondary_model_4_stem_scale': [],
+    'secondary_model_bass': None,
+    'secondary_model_drums': None,
+    'secondary_model_other': None,
+    'secondary_model_scale': None,
+    'secondary_model_scale_bass': None,
+    'secondary_model_scale_drums': None,
+    'secondary_model_scale_other': None,
+    'secondary_stem': None,
+    'segment': 'Default',
+    'semitone_shift': 0.0,
+    'shifts': 2,
+    'vocal_split_model': None,
+    'wav_type_set': 'PCM_16'}
+
+htdemucs_ft_process_data = {
+    'model_name': 'Demucs v4: htdemucs_ft',
+    'model_data': Dict(htdemucs_ft_model_data),
+    'export_path': '/Users/taoluo/Downloads/Ensembled_Outputs_1702201464',
+    'audio_file_base': '1_test_audio_htdemucs_ft',
+    'audio_file': '/Users/taoluo/Downloads/test_audio.mp3',
+    'set_progress_bar': lambda step, inference_iterations=0: print(
+        f"iteration {inference_iterations} of step #{step}"),
+    'write_to_console': lambda progress_text, base_text='base_text': print(
+        f"{progress_text} {base_text}"),
+    'process_iteration': lambda iteration: iteration + 1,
+    'cached_source_callback': cached_source_callback,
+    'cached_model_source_holder': cached_model_source_holder,
+    'list_all_models': ['MDX23C-8KFFT-InstVoc_HQ', 'UVR-MDX-NET-Voc_FT', 'htdemucs_ft'],
+    'is_ensemble_master': True,
+    'is_4_stem_ensemble': False}

inst.py

@@ -0,0 +1,100 @@
+import hashlib
+import os
+import shutil
+import sys
+from datetime import datetime
+from pathlib import Path
+
+import click
+import yaml
+from ml_collections import ConfigDict
+
+from UVR import ModelData, AudioTools, Ensembler, DENOISER_MODEL_PATH, DEVERBER_MODEL_PATH, MDX_MODELS_DIR, \
+    MDX_MIXER_PATH, load_model_hash_data, MDX_MODEL_NAME_SELECT, model_hash_table, MDX_HASH_DIR, MDX_C_CONFIG_PATH
+from args import mdx23c_8kfft_instvoc_hq_process_data, htdemucs_ft_process_data, uvr_mdx_net_voc_ft_process_data
+from download import download_model, get_model_file
+from gui_data.constants import VR_ARCH_TYPE, MDX_ARCH_TYPE, DEMUCS_ARCH_TYPE, ENSEMBLE_MODE, TIME_STRETCH, \
+    MANUAL_ENSEMBLE, MATCH_INPUTS, ALIGN_INPUTS, ALL_STEMS, DEFAULT, VOCAL_STEM, MP3_BIT_RATES, WAV, DEMUCS_2_SOURCE, \
+    DEMUCS_2_SOURCE_MAPPER, INST_STEM, CKPT, ONNX, MDX_POP_NFFT, secondary_stem, PRIMARY_STEM, SECONDARY_STEM
+from lib_v5 import spec_utils
+from separate import (
+    SeperateDemucs, SeperateMDX, SeperateMDXC, SeperateVR,  # Model-related
+    save_format, clear_gpu_cache,  # Utility functions
+    cuda_available, mps_available,  # directml_available,
+)
+
+
+def run_ensemble_models(audio_path, export_path, format=WAV, clean=True):
+    start = datetime.now()
+    process_datas = [mdx23c_8kfft_instvoc_hq_process_data, uvr_mdx_net_voc_ft_process_data,
+                     htdemucs_ft_process_data]
+    # download models
+    for process_data in process_datas:
+        download_model(process_data['model_name'])
+
+    # create folder
+    os.makedirs(export_path, exist_ok=True)
+    temp_export_path = os.path.join(export_path, 'uvr5_' + datetime.now().strftime("%Y-%m-%d %H:%M:%S"))
+    os.makedirs(temp_export_path, exist_ok=True)
+    print(f'temp_export_path', temp_export_path)
+
+    instrumental_export_paths = []
+    vocals_export_paths = []
+
+    for process_data in process_datas:
+        current_model = process_data['model_data']
+        audio_file_base = Path(audio_path).stem + '_' + current_model.model_basename
+        process_data['export_path'] = temp_export_path
+        process_data['audio_file_base'] = audio_file_base
+        process_data['audio_file'] = audio_path
+
+        if current_model.process_method == VR_ARCH_TYPE:
+            seperator = SeperateVR(current_model, process_data)
+        elif current_model.process_method == MDX_ARCH_TYPE:
+            seperator = SeperateMDXC(current_model, process_data) if current_model.is_mdx_c else SeperateMDX(
+                current_model, process_data)
+        elif current_model.process_method == DEMUCS_ARCH_TYPE:
+            seperator = SeperateDemucs(current_model, process_data, vocal_stem_path=(audio_path, audio_file_base))
+        else:
+            raise Exception(f'model not found')
+
+        seperator.seperate()
+
+        instrumental_path = Path(temp_export_path) / f"{audio_file_base}_(Instrumental).{format.lower()}"
+        vocals_path = Path(temp_export_path) / f"{audio_file_base}_(Vocals).{format.lower()}"
+        instrumental_export_paths.append(str(instrumental_path))
+        vocals_export_paths.append(str(vocals_path))
+
+    # merge each model outputs
+    vocals_final_path = Path(export_path) / f"{Path(audio_path).stem}.vocal.{format.lower()}"
+    instrumental_final_path = Path(export_path) / f"{Path(audio_path).stem}.instrumental.{format.lower()}"
+
+    ensemble(vocals_export_paths, vocals_final_path)
+    ensemble(instrumental_export_paths, instrumental_final_path)
+
+    if clean:
+        shutil.rmtree(temp_export_path)
+    print(f'instrumental_final_path', instrumental_final_path)
+    print(f'vocals_final_path', vocals_final_path)
+    print(f'Finished in {datetime.now() - start}')
+    return instrumental_final_path, vocals_final_path
+
+
+def ensemble(stem_outputs, stem_save_path, format=WAV):
+    algorithm = 'Average'
+    is_normalization = True
+    spec_utils.ensemble_inputs(stem_outputs, algorithm, is_normalization, 'PCM_16', stem_save_path, is_wave=True)
+    save_format(stem_save_path, format, '320k')
+
+
+# /Users/taoluo/Downloads/test/kimk_audio_MDX23C-8KFFT-InstVoc_HQ_(Instrumental).WAV
+#
+if __name__ == '__main__':
+    audio_file = '/Users/taoluo/Downloads/assets/audio/kimk_audio.mp3'
+    audio_file = sys.argv[1]
+    if not os.path.isfile(audio_file):
+        raise FileNotFoundError(f"File {audio_file} not exist.")
+
+    output_dir = os.path.dirname(audio_file)
+    print(output_dir)
+    run_ensemble_models(audio_file, output_dir)