Files
RLD-UVR-Opt/separate.py
RevLayDust 90050a048e feat(engine): add TensorRT-RTX execution provider backend with in-memory FP16 downcast
- Implement TensorRT-RTX backend via onnxruntime-ep-nv-tensorrt-rtx plugin in engine/tensorrt_rtx/

- Support in-memory FP16 downcast for MDX-Net ONNX models with zero disk writes

- Implement automatic fallback chain: TensorRT-RTX -> CUDA -> CPU

- Add GUI execution provider selection with dynamic hardware gating

- Add CLI provider flags and model resolver support for trt_rtx

- Add install/requirements_trt_rtx.txt and stack profile tracking

- Add test suite in tests/test_tensorrt_rtx.py and test_uvr_cli.py

- Update run_uvr.bat to execute UVR.py via native CMD
2026-10-09 16:05:22 +07:00

2318 lines
111 KiB
Python

from __future__ import annotations
from typing import TYPE_CHECKING
from demucs.apply import apply_model, demucs_segments
from demucs.hdemucs import HDemucs
from demucs.model_v2 import auto_load_demucs_model_v2
from demucs.pretrained import get_model as _gm
from demucs.utils import apply_model_v1
from demucs.utils import apply_model_v2
from lib_v5.tfc_tdf_v3 import TFC_TDF_net, STFT
from lib_v5 import spec_utils
from lib_v5.vr_network import nets
from lib_v5.vr_network import nets_new
from lib_v5.vr_network.model_param_init import ModelParameters
from pathlib import Path
from gui_data.constants import *
from gui_data.error_handling import *
from scipy import signal
import audioread
import gzip
import librosa
import math
import numpy as np
import torch
import onnxruntime as ort
import os
import warnings
import pydub
import soundfile as sf
import lib_v5.mdxnet as MdxnetSet
import math
#import random
from onnx import load
from onnx2pytorch import ConvertModel
import gc
import queue
import threading
import time
class ONNXExecutionWorker:
def __init__(self):
self.queue = queue.Queue()
self.thread = threading.Thread(target=self._run, daemon=True, name="ONNX_Execution_Worker")
self.thread.start()
def _run(self):
while True:
func, args, kwargs, result_queue = self.queue.get()
result = None
try:
result = func(*args, **kwargs)
result_queue.put((result, None))
except Exception as error:
result_queue.put((None, error))
finally:
self.queue.task_done()
func = None
args = None
kwargs = None
result_queue = None
result = None
def execute(self, func, *args, **kwargs):
res_queue = queue.Queue()
self.queue.put((func, args, kwargs, res_queue))
res, err = res_queue.get()
if err is not None:
raise err
return res
ONNX_WORKER = ONNXExecutionWorker()
class ModelCacheManager:
def __init__(self, max_cached_models=1, idle_timeout_seconds=300):
self.cache = {}
self.max_cached_models = 1
self.idle_timeout_seconds = idle_timeout_seconds
self._lock = threading.RLock()
if max_cached_models != 1:
print("[MODEL CACHE] Multi-model caching is disabled; only one model/precision session may remain active.")
@staticmethod
def _normalize_path(model_path):
return os.path.normpath(os.path.abspath(model_path))
@staticmethod
def _same_precision(cached_precision, requested_precision):
return requested_precision is None or cached_precision == requested_precision
def get_model(self, model_path, precision_mode=None):
normalized_path = self._normalize_path(model_path)
with self._lock:
self._evict_inactive(normalized_path, precision_mode)
print(f"[DIAGNOSTIC] get_model lookup: {model_path} -> normalized: {normalized_path}")
entry = self.cache.get(normalized_path)
if entry is not None and self._same_precision(entry.get('precision_mode'), precision_mode):
entry['last_used'] = time.time()
print(f"Model Cache Hit: Reusing resident model {os.path.basename(model_path)}")
return entry['model_obj']
print(f"[DIAGNOSTIC] get_model Cache Miss for: {normalized_path}")
return None
def set_model(self, model_path, model_obj, size_bytes=0, precision_mode=None):
normalized_path = self._normalize_path(model_path)
with self._lock:
self._evict_inactive(normalized_path, precision_mode)
if normalized_path in self.cache:
self.unload_model(normalized_path, reason="Session Replacement")
self.cache[normalized_path] = {
'model_obj': model_obj,
'last_used': time.time(),
'size_bytes': size_bytes,
'precision_mode': precision_mode
}
print(f"Model Cache Store: Resident model cached {os.path.basename(model_path)} (precision: {precision_mode})")
def _evict_inactive(self, requested_path, requested_precision):
for cached_path, entry in list(self.cache.items()):
same_path = cached_path == requested_path
same_precision = self._same_precision(entry.get('precision_mode'), requested_precision)
if same_path and same_precision:
continue
reason = "Precision Change" if same_path else "Model Change"
self.unload_model(cached_path, reason=reason)
def evict_if_necessary(self, incoming_size_bytes=0):
with self._lock:
now = time.time()
for cached_path, entry in list(self.cache.items()):
if now - entry['last_used'] > self.idle_timeout_seconds:
self.unload_model(cached_path, reason="Idle Timeout")
def unload_model(self, model_path, reason="Explicit Release"):
normalized_path = self._normalize_path(model_path)
with self._lock:
entry = self.cache.pop(normalized_path, None)
if entry is None:
return False
print(f"Model Cache Eviction ({reason}): Unloading {os.path.basename(normalized_path)}")
model_obj = entry.pop('model_obj', None)
entry.clear()
del model_obj
del entry
gc.collect()
if torch.cuda.is_available():
try:
torch.cuda.synchronize()
except Exception:
pass
torch.cuda.empty_cache()
return True
def clear_all(self):
with self._lock:
for cached_path in list(self.cache):
self.unload_model(cached_path, reason="Clear All")
MODEL_CACHE = ModelCacheManager()
class SmartPrecisionPolicy:
"""
Smart Precision Policy abstraction that controls:
- speed
- VRAM
- numerical stability
- output quality
Precision Preset Modes:
1. Ultra Quality: FP32
2. Balanced: BF16
3. Performance: FP16
"""
def __init__(self, mode_setting: str):
self.raw = normalize_model_precision(mode_setting)
self.mode = self._parse_mode(self.raw)
@staticmethod
def _parse_mode(raw: str) -> str:
if raw == MODEL_PRECISION_FP16:
return 'FP16'
elif raw == MODEL_PRECISION_BF16:
return 'BF16'
return 'FP32'
@property
def is_fp32(self) -> bool:
return self.mode == 'FP32'
@property
def is_bf16(self) -> bool:
return self.mode == 'BF16'
@property
def is_fp16(self) -> bool:
return self.mode == 'FP16'
@property
def is_half_or_lower(self) -> bool:
return self.mode != 'FP32'
@property
def enabled_amp(self) -> bool:
return self.mode != 'FP32'
@property
def amp_dtype(self):
if self.mode == 'FP32':
return torch.float32
elif self.mode == 'BF16':
return torch.bfloat16
elif self.mode == 'FP16':
return torch.float16
return torch.float32
def cast_model(self, model: torch.nn.Module) -> torch.nn.Module:
"""Casts PyTorch neural net weights based on policy."""
if self.mode == 'FP32':
return model
elif self.mode == 'BF16':
return model.bfloat16()
elif self.mode == 'FP16':
return model.half()
return model
def prepare_spectrogram(self, spek: torch.Tensor) -> torch.Tensor:
"""Applies input precision transforms based on policy."""
if self.mode == 'FP32':
return spek.to(torch.float32)
elif self.mode == 'BF16':
return spek.bfloat16()
elif self.mode == 'FP16':
return spek.half()
return spek
def effective_mode_for(self, model_type: str) -> str:
"""Returns what precision will actually be used for a given model type.
Args:
model_type: 'ONNX', 'PyTorch-CKPT', or 'PyTorch-ConvertModel'
Returns:
The actual precision mode that will be applied at runtime.
"""
if model_type == 'ONNX':
if self.mode == 'FP32':
return 'FP32'
elif self.mode == 'FP16':
# ONNX only supports FP16 quantization via onnxconverter-common
return 'FP16'
elif self.mode == 'BF16':
# ONNX Runtime has no native BF16 support — must stay FP32
return 'FP32'
return 'FP32'
else:
# PyTorch models support the available precision modes natively
return self.mode
def diagnose(self, model_run, model_type: str, execution_provider: str = "") -> dict:
"""Detect and report actual runtime precision state for full transparency.
Args:
model_run: The model object (torch.nn.Module or ort.InferenceSession)
model_type: 'ONNX', 'PyTorch-CKPT', or 'PyTorch-ConvertModel'
execution_provider: e.g. 'CUDAExecutionProvider'
Returns:
Dict with precision diagnostics.
"""
effective = self.effective_mode_for(model_type)
report = {
'requested_mode': self.mode,
'effective_mode': effective,
'model_type': model_type,
'execution_provider': execution_provider or 'N/A',
}
if isinstance(model_run, torch.nn.Module):
first_param = next(model_run.parameters(), None)
report['model_weight_dtype'] = str(first_param.dtype) if first_param is not None else 'N/A'
report['autocast_enabled'] = self.enabled_amp
report['autocast_dtype'] = str(self.amp_dtype)
# Verify consistency
if first_param is not None:
actual_dtype_str = str(first_param.dtype)
if self.mode == 'FP16' and 'float16' not in actual_dtype_str:
report['warning'] = f"Requested FP16 but model weights are {actual_dtype_str}"
elif self.mode == 'BF16' and 'bfloat16' not in actual_dtype_str:
report['warning'] = f"Requested BF16 but model weights are {actual_dtype_str}"
elif hasattr(model_run, 'get_inputs'):
# ONNX InferenceSession
input_type = model_run.get_inputs()[0].type
report['onnx_input_type'] = input_type
report['onnx_output_type'] = model_run.get_outputs()[0].type if model_run.get_outputs() else 'N/A'
active_providers = model_run.get_providers() if hasattr(model_run, 'get_providers') else []
report['active_providers'] = active_providers
# Detect actual ONNX runtime precision
if 'float16' in input_type:
report['effective_mode'] = 'FP16'
else:
report['effective_mode'] = 'FP32'
# Report limitations honestly
if self.mode != report['effective_mode']:
if self.mode == 'BF16':
report['fallback_reason'] = (
f"ONNX Runtime does not support {self.mode}. "
f"Model executes at {report['effective_mode']}. "
f"BF16 requires PyTorch (.ckpt) models."
)
elif self.mode == 'FP16' and 'float16' not in input_type:
report['fallback_reason'] = (
f"FP16 quantization of ONNX model failed or was not applied. "
f"Model executes at FP32."
)
else:
# Lambda wrapper (ONNX via ONNX_WORKER)
report['model_weight_dtype'] = 'opaque (lambda wrapper)'
report['note'] = 'Model is accessed via lambda; precision is determined by its ONNX Runtime session.'
# Print the report
print(f"[PRECISION] === Runtime Precision Report ===")
for k, v in report.items():
print(f"[PRECISION] {k}: {v}")
if report.get('fallback_reason'):
print(f"[PRECISION] [WARN] LIMITATION: {report['fallback_reason']}")
if self.mode != report['effective_mode']:
print(f"[PRECISION] [WARN] MISMATCH: Requested={self.mode}, Actual={report['effective_mode']}")
else:
print(f"[PRECISION] [OK] Precision applied correctly: {report['effective_mode']}")
print(f"[PRECISION] ==============================")
return report
class AdaptiveChunkScheduler:
def __init__(self, device, precision_mode, model_size_bytes, batch_size):
self.device = device
self.precision_mode = precision_mode
self.policy = SmartPrecisionPolicy(precision_mode)
self.model_size_bytes = model_size_bytes
self.batch_size = batch_size
def get_free_vram(self):
if torch.cuda.is_available():
try:
free_mem, total_mem = torch.cuda.mem_get_info(self.device)
return free_mem
except Exception:
pass
return 8 * 1024 * 1024 * 1024 # Default fallback 8GB
def determine_chunk_size(self):
free_vram = self.get_free_vram()
free_vram_gb = free_vram / (1024 ** 3)
# Specification workflow
if free_vram_gb > 14.0:
chunk_size = 8192
elif free_vram_gb > 10.0:
chunk_size = 4096
elif free_vram_gb > 6.0:
chunk_size = 2048
else:
chunk_size = 1024
# Refine based on precision, model size, and batch size
if not self.policy.is_half_or_lower:
# FP32 uses twice the memory of lower precision
chunk_size = max(1024, chunk_size // 2)
if self.batch_size > 4:
# Larger batch size requires smaller chunks
chunk_size = max(1024, chunk_size // 2)
return chunk_size
def adjust_overlap_ratio(self, chunk_size, user_overlap):
# Larger chunk size can afford smaller overlap ratio to speed up processing
# Smaller chunk size needs larger overlap ratio to maintain quality
if chunk_size >= 4096:
return min(user_overlap, 0.25) # Lower overlap
elif chunk_size >= 2048:
return min(user_overlap, 0.5)
else:
return user_overlap # Keep user selection for small chunks
def create_onnx_fp16_session(model_path, providers):
import onnx
from onnxconverter_common import float16
model_name = os.path.basename(model_path)
print(f"[PRECISION] Loading {model_name} into CPU RAM for in-memory FP16 conversion...")
model_fp32 = onnx.load(model_path, load_external_data=True)
try:
model_fp16 = float16.convert_float_to_float16(model_fp32, keep_io_types=False)
finally:
del model_fp32
gc.collect()
try:
serialized_fp16 = model_fp16.SerializeToString()
finally:
del model_fp16
gc.collect()
try:
print(f"[PRECISION] Creating ONNX Runtime FP16 session from RAM ({len(serialized_fp16) / (1024 ** 2):.2f} MiB)...")
return ort.InferenceSession(serialized_fp16, providers=providers)
finally:
del serialized_fp16
gc.collect()
def create_onnx_session(model_path, providers, precision_policy):
# Delegate to TensorRT-RTX wrapper if TensorRT-RTX was requested
is_trt_requested = any(
(isinstance(p, tuple) and "NvTensorRTRTX" in p[0]) or (isinstance(p, str) and "NvTensorRTRTX" in p)
for p in providers
) if isinstance(providers, (list, tuple)) else False
if is_trt_requested:
from engine.tensorrt_rtx import create_trt_rtx_session
dev_id = 0
cuda_options = None
if isinstance(providers, (list, tuple)):
for p in providers:
if isinstance(p, tuple):
if p[0] == 'CUDAExecutionProvider' and isinstance(p[1], dict):
cuda_options = p[1]
if isinstance(p[1], dict) and "device_id" in p[1]:
try:
dev_id = int(p[1]["device_id"])
except Exception:
dev_id = 0
model_name = os.path.basename(model_path)
if precision_policy.is_fp16:
# Respect FP16 preset on TensorRT-RTX path
import onnx
from onnxconverter_common import float16
print(f"[PRECISION] Loading {model_name} into CPU RAM for in-memory FP16 conversion...")
try:
model_fp32 = onnx.load(model_path, load_external_data=True)
try:
model_fp16 = float16.convert_float_to_float16(model_fp32, keep_io_types=False)
finally:
del model_fp32
gc.collect()
try:
serialized_fp16 = model_fp16.SerializeToString()
finally:
del model_fp16
gc.collect()
print(f"[PRECISION] Creating TensorRT-RTX FP16 session from RAM ({len(serialized_fp16) / (1024 ** 2):.2f} MiB)...")
try:
return create_trt_rtx_session(
serialized_fp16,
device_id=dev_id,
cuda_options=cuda_options,
model_name=model_name,
)
finally:
del serialized_fp16
gc.collect()
except Exception as error:
print(
f"[PRECISION] In-memory FP16 conversion failed ({type(error).__name__}: {error}). "
"Falling back to the original FP32 ONNX model without writing a temporary file."
)
return create_trt_rtx_session(
model_path,
device_id=dev_id,
cuda_options=cuda_options,
model_name=model_name,
)
else:
return create_trt_rtx_session(
model_path,
device_id=dev_id,
cuda_options=cuda_options,
model_name=model_name,
)
if not precision_policy.is_fp16:
return ort.InferenceSession(model_path, providers=providers)
try:
return create_onnx_fp16_session(model_path, providers)
except Exception as error:
print(
f"[PRECISION] In-memory FP16 conversion failed ({type(error).__name__}: {error}). "
"Falling back to the original FP32 ONNX model without writing a temporary file."
)
return ort.InferenceSession(model_path, providers=providers)
if TYPE_CHECKING:
from UVR import ModelData
# if not is_macos:
# import torch_directml
mps_available = torch.backends.mps.is_available() if is_macos else False
cuda_available = torch.cuda.is_available()
# def get_gpu_info():
# directml_device, directml_available = DIRECTML_DEVICE, False
# if not is_macos:
# directml_available = torch_directml.is_available()
# if directml_available:
# directml_device = str(torch_directml.device()).partition(":")[0]
# return directml_device, directml_available
# DIRECTML_DEVICE, directml_available = get_gpu_info()
def clear_gpu_cache():
gc.collect()
if is_macos:
torch.mps.empty_cache()
else:
torch.cuda.empty_cache()
warnings.filterwarnings("ignore")
cpu = 'cpu'
class SeperateAttributes:
def __init__(self, model_data: ModelData,
process_data: dict,
main_model_primary_stem_4_stem=None,
main_process_method=None,
is_return_dual=True,
main_model_primary=None,
vocal_stem_path=None,
master_inst_source=None,
master_vocal_source=None):
self.list_all_models: list
self.model_data = model_data
self.process_data = process_data
self.progress_value = 0
self.set_progress_bar = process_data['set_progress_bar']
self.write_to_console = process_data['write_to_console']
if vocal_stem_path:
self.audio_file, self.audio_file_base = vocal_stem_path
self.audio_file_base_voc_split = lambda stem, split:os.path.join(self.export_path, f'{self.audio_file_base.replace("_(Vocals)", "")}_({stem}_{split}).wav')
else:
self.audio_file = process_data['audio_file']
self.audio_file_base = process_data['audio_file_base']
self.audio_file_base_voc_split = None
self.export_path = process_data['export_path']
self.cached_source_callback = process_data['cached_source_callback']
self.cached_model_source_holder = process_data['cached_model_source_holder']
self.is_4_stem_ensemble = process_data['is_4_stem_ensemble']
self.list_all_models = process_data['list_all_models']
self.process_iteration = process_data['process_iteration']
self.is_return_dual = is_return_dual
self.is_pitch_change = model_data.is_pitch_change
self.semitone_shift = model_data.semitone_shift
self.is_match_frequency_pitch = model_data.is_match_frequency_pitch
self.overlap = model_data.overlap
self.overlap_mdx = model_data.overlap_mdx
self.overlap_mdx23 = model_data.overlap_mdx23
self.is_mdx_combine_stems = model_data.is_mdx_combine_stems
self.is_mdx_c = model_data.is_mdx_c
self.mdx_c_configs = model_data.mdx_c_configs
self.mdxnet_stem_select = model_data.mdxnet_stem_select
self.mixer_path = model_data.mixer_path
self.model_samplerate = model_data.model_samplerate
self.model_capacity = model_data.model_capacity
self.is_vr_51_model = model_data.is_vr_51_model
self.is_pre_proc_model = model_data.is_pre_proc_model
self.is_secondary_model_activated = model_data.is_secondary_model_activated if not self.is_pre_proc_model else False
self.is_secondary_model = model_data.is_secondary_model if not self.is_pre_proc_model else True
self.process_method = model_data.process_method
self.model_path = model_data.model_path
self.model_name = model_data.model_name
self.model_basename = model_data.model_basename
self.wav_type_set = model_data.wav_type_set
self.mp3_bit_set = model_data.mp3_bit_set
self.save_format = model_data.save_format
self.is_gpu_conversion = model_data.is_gpu_conversion
self.is_normalization = model_data.is_normalization
self.is_primary_stem_only = model_data.is_primary_stem_only if not self.is_secondary_model else model_data.is_primary_model_primary_stem_only
self.is_secondary_stem_only = model_data.is_secondary_stem_only if not self.is_secondary_model else model_data.is_primary_model_secondary_stem_only
self.is_ensemble_mode = model_data.is_ensemble_mode
self.secondary_model = model_data.secondary_model #
self.primary_model_primary_stem = model_data.primary_model_primary_stem
self.primary_stem_native = model_data.primary_stem_native
self.primary_stem = model_data.primary_stem #
self.secondary_stem = model_data.secondary_stem #
self.is_invert_spec = model_data.is_invert_spec #
self.is_deverb_vocals = model_data.is_deverb_vocals
self.is_mixer_mode = model_data.is_mixer_mode #
self.secondary_model_scale = model_data.secondary_model_scale #
self.is_demucs_pre_proc_model_inst_mix = model_data.is_demucs_pre_proc_model_inst_mix #
self.primary_source_map = {}
self.secondary_source_map = {}
self.primary_source = None
self.secondary_source = None
self.secondary_source_primary = None
self.secondary_source_secondary = None
self.main_model_primary_stem_4_stem = main_model_primary_stem_4_stem
self.main_model_primary = main_model_primary
self.ensemble_primary_stem = model_data.ensemble_primary_stem
self.is_multi_stem_ensemble = model_data.is_multi_stem_ensemble
self.is_other_gpu = False
self.is_deverb = True
self.DENOISER_MODEL = model_data.DENOISER_MODEL
self.DEVERBER_MODEL = model_data.DEVERBER_MODEL
self.is_source_swap = False
self.vocal_split_model = model_data.vocal_split_model
self.is_vocal_split_model = model_data.is_vocal_split_model
self.master_vocal_path = None
self.set_master_inst_source = None
self.master_inst_source = master_inst_source
self.master_vocal_source = master_vocal_source
self.is_save_inst_vocal_splitter = isinstance(master_inst_source, np.ndarray) and model_data.is_save_inst_vocal_splitter
self.is_inst_only_voc_splitter = model_data.is_inst_only_voc_splitter
self.is_karaoke = model_data.is_karaoke
self.is_bv_model = model_data.is_bv_model
self.is_bv_model_rebalenced = model_data.bv_model_rebalance and self.is_vocal_split_model
self.is_sec_bv_rebalance = model_data.is_sec_bv_rebalance
self.stem_path_init = os.path.join(self.export_path, f'{self.audio_file_base}_({self.secondary_stem}).wav')
self.deverb_vocal_opt = model_data.deverb_vocal_opt
self.is_save_vocal_only = model_data.is_save_vocal_only
self.device = cpu
self.run_type = ['CPUExecutionProvider']
self.is_opencl = False
self.device_set = model_data.device_set
self.is_use_opencl = model_data.is_use_opencl
if self.is_inst_only_voc_splitter or self.is_sec_bv_rebalance:
self.is_primary_stem_only = False
self.is_secondary_stem_only = False
if main_model_primary and self.is_multi_stem_ensemble:
self.primary_stem, self.secondary_stem = main_model_primary, secondary_stem(main_model_primary)
if self.is_gpu_conversion >= 0:
if mps_available:
self.device, self.is_other_gpu = 'mps', True
else:
device_prefix = None
if self.device_set != DEFAULT:
device_prefix = CUDA_DEVICE#DIRECTML_DEVICE if self.is_use_opencl and directml_available else CUDA_DEVICE
# if directml_available and self.is_use_opencl:
# self.device = torch_directml.device() if not device_prefix else f'{device_prefix}:{self.device_set}'
# self.is_other_gpu = True
if cuda_available:# and not self.is_use_opencl:
self.device = CUDA_DEVICE if not device_prefix else f'{device_prefix}:{self.device_set}'
cuda_options = {
'arena_extend_strategy': 'kSameAsRequested',
'cudnn_conv_algo_search': 'HEURISTIC',
'do_copy_in_default_stream': 'True'
}
chosen_ep = getattr(model_data, 'execution_provider', 'CUDAExecutionProvider')
if chosen_ep == 'NvTensorRTRTXExecutionProvider':
self.run_type = ['NvTensorRTRTXExecutionProvider', ('CUDAExecutionProvider', cuda_options), 'CPUExecutionProvider']
else:
self.run_type = [('CUDAExecutionProvider', cuda_options), 'CPUExecutionProvider']
if model_data.process_method == MDX_ARCH_TYPE:
self.is_mdx_ckpt = model_data.is_mdx_ckpt
self.primary_model_name, self.primary_sources = self.cached_source_callback(MDX_ARCH_TYPE, model_name=self.model_basename)
self.is_denoise = model_data.is_denoise#
self.is_denoise_model = model_data.is_denoise_model#
self.is_mdx_c_seg_def = model_data.is_mdx_c_seg_def#
self.mdx_batch_size = model_data.mdx_batch_size
self.compensate = model_data.compensate
self.mdx_segment_size = model_data.mdx_segment_size
if self.is_mdx_c:
if not self.is_4_stem_ensemble:
self.primary_stem = model_data.ensemble_primary_stem if process_data['is_ensemble_master'] else model_data.primary_stem
self.secondary_stem = model_data.ensemble_secondary_stem if process_data['is_ensemble_master'] else model_data.secondary_stem
else:
self.dim_f, self.dim_t = model_data.mdx_dim_f_set, 2**model_data.mdx_dim_t_set
self.check_label_secondary_stem_runs()
self.n_fft = model_data.mdx_n_fft_scale_set
self.chunks = model_data.chunks
self.margin = model_data.margin
self.adjust = 1
self.dim_c = 4
self.hop = 1024
if model_data.process_method == DEMUCS_ARCH_TYPE:
self.demucs_stems = model_data.demucs_stems if not main_process_method in [MDX_ARCH_TYPE, VR_ARCH_TYPE] else None
self.secondary_model_4_stem = model_data.secondary_model_4_stem
self.secondary_model_4_stem_scale = model_data.secondary_model_4_stem_scale
self.is_chunk_demucs = model_data.is_chunk_demucs
self.segment = model_data.segment
self.demucs_version = model_data.demucs_version
self.demucs_source_list = model_data.demucs_source_list
self.demucs_source_map = model_data.demucs_source_map
self.is_demucs_combine_stems = model_data.is_demucs_combine_stems
self.demucs_stem_count = model_data.demucs_stem_count
self.pre_proc_model = model_data.pre_proc_model
self.device = cpu if self.is_other_gpu and not self.demucs_version in [DEMUCS_V3, DEMUCS_V4] else self.device
self.primary_stem = model_data.ensemble_primary_stem if process_data['is_ensemble_master'] else model_data.primary_stem
self.secondary_stem = model_data.ensemble_secondary_stem if process_data['is_ensemble_master'] else model_data.secondary_stem
if (self.is_multi_stem_ensemble or self.is_4_stem_ensemble) and not self.is_secondary_model:
self.is_return_dual = False
if self.is_multi_stem_ensemble and main_model_primary:
self.is_4_stem_ensemble = False
if main_model_primary in self.demucs_source_map.keys():
self.primary_stem = main_model_primary
self.secondary_stem = secondary_stem(main_model_primary)
elif secondary_stem(main_model_primary) in self.demucs_source_map.keys():
self.primary_stem = secondary_stem(main_model_primary)
self.secondary_stem = main_model_primary
if self.is_secondary_model and not process_data['is_ensemble_master']:
if not self.demucs_stem_count == 2 and model_data.primary_model_primary_stem == INST_STEM:
self.primary_stem = VOCAL_STEM
self.secondary_stem = INST_STEM
else:
self.primary_stem = model_data.primary_model_primary_stem
self.secondary_stem = secondary_stem(self.primary_stem)
self.shifts = model_data.shifts
self.is_split_mode = model_data.is_split_mode if not self.demucs_version == DEMUCS_V4 else True
self.primary_model_name, self.primary_sources = self.cached_source_callback(DEMUCS_ARCH_TYPE, model_name=self.model_basename)
if model_data.process_method == VR_ARCH_TYPE:
self.check_label_secondary_stem_runs()
self.primary_model_name, self.primary_sources = self.cached_source_callback(VR_ARCH_TYPE, model_name=self.model_basename)
self.mp = model_data.vr_model_param
self.high_end_process = model_data.is_high_end_process
self.is_tta = model_data.is_tta
self.is_post_process = model_data.is_post_process
self.is_gpu_conversion = model_data.is_gpu_conversion
self.batch_size = model_data.batch_size
self.window_size = model_data.window_size
self.input_high_end_h = None
self.input_high_end = None
self.post_process_threshold = model_data.post_process_threshold
self.aggressiveness = {'value': model_data.aggression_setting,
'split_bin': self.mp.param['band'][1]['crop_stop'],
'aggr_correction': self.mp.param.get('aggr_correction')}
def __del__(self):
print(f"[DIAGNOSTIC] === SeperateAttributes Destructor Called for: {getattr(self, 'model_basename', 'Unknown Model')} ===")
def check_label_secondary_stem_runs(self):
# For ensemble master that's not a 4-stem ensemble, and not mdx_c
if self.process_data['is_ensemble_master'] and not self.is_4_stem_ensemble and not self.is_mdx_c:
if self.ensemble_primary_stem != self.primary_stem:
self.is_primary_stem_only, self.is_secondary_stem_only = self.is_secondary_stem_only, self.is_primary_stem_only
# For secondary models
if self.is_pre_proc_model or self.is_secondary_model:
self.is_primary_stem_only = False
self.is_secondary_stem_only = False
def start_inference_console_write(self):
if self.is_secondary_model and not self.is_pre_proc_model and not self.is_vocal_split_model:
self.write_to_console(INFERENCE_STEP_2_SEC(self.process_method, self.model_basename))
if self.is_pre_proc_model:
self.write_to_console(INFERENCE_STEP_2_PRE(self.process_method, self.model_basename))
if self.is_vocal_split_model:
self.write_to_console(INFERENCE_STEP_2_VOC_S(self.process_method, self.model_basename))
def running_inference_console_write(self, is_no_write=False):
self.write_to_console(DONE, base_text='') if not is_no_write else None
self.set_progress_bar(0.05) if not is_no_write else None
if self.is_secondary_model and not self.is_pre_proc_model and not self.is_vocal_split_model:
self.write_to_console(INFERENCE_STEP_1_SEC)
elif self.is_pre_proc_model:
self.write_to_console(INFERENCE_STEP_1_PRE)
elif self.is_vocal_split_model:
self.write_to_console(INFERENCE_STEP_1_VOC_S)
else:
self.write_to_console(INFERENCE_STEP_1)
def running_inference_progress_bar(self, length, is_match_mix=False):
if not is_match_mix:
self.progress_value += 1
if (0.8/length*self.progress_value) >= 0.8:
length = self.progress_value + 1
self.set_progress_bar(0.1, (0.8/length*self.progress_value))
def load_cached_sources(self):
if self.is_secondary_model and not self.is_pre_proc_model:
self.write_to_console(INFERENCE_STEP_2_SEC_CACHED_MODOEL(self.process_method, self.model_basename))
elif self.is_pre_proc_model:
self.write_to_console(INFERENCE_STEP_2_PRE_CACHED_MODOEL(self.process_method, self.model_basename))
else:
self.write_to_console(INFERENCE_STEP_2_PRIMARY_CACHED, "")
def cache_source(self, secondary_sources):
model_occurrences = self.list_all_models.count(self.model_basename)
if not model_occurrences <= 1:
if self.process_method == MDX_ARCH_TYPE:
self.cached_model_source_holder(MDX_ARCH_TYPE, secondary_sources, self.model_basename)
if self.process_method == VR_ARCH_TYPE:
self.cached_model_source_holder(VR_ARCH_TYPE, secondary_sources, self.model_basename)
if self.process_method == DEMUCS_ARCH_TYPE:
self.cached_model_source_holder(DEMUCS_ARCH_TYPE, secondary_sources, self.model_basename)
def process_vocal_split_chain(self, sources: dict):
def is_valid_vocal_split_condition(master_vocal_source):
"""Checks if conditions for vocal split processing are met."""
conditions = [
isinstance(master_vocal_source, np.ndarray),
self.vocal_split_model,
not self.is_ensemble_mode,
not self.is_karaoke,
not self.is_bv_model
]
return all(conditions)
# Retrieve sources from the dictionary with default fallbacks
master_inst_source = sources.get(INST_STEM, None)
master_vocal_source = sources.get(VOCAL_STEM, None)
# Process the vocal split chain if conditions are met
if is_valid_vocal_split_condition(master_vocal_source):
process_chain_model(
self.vocal_split_model,
self.process_data,
vocal_stem_path=self.master_vocal_path,
master_vocal_source=master_vocal_source,
master_inst_source=master_inst_source
)
def process_secondary_stem(self, stem_source, secondary_model_source=None, model_scale=None):
if not self.is_secondary_model:
if self.is_secondary_model_activated and isinstance(secondary_model_source, np.ndarray):
secondary_model_scale = model_scale if model_scale else self.secondary_model_scale
stem_source = spec_utils.average_dual_sources(stem_source, secondary_model_source, secondary_model_scale)
return stem_source
def final_process(self, stem_path, source, secondary_source, stem_name, samplerate):
source = self.process_secondary_stem(source, secondary_source)
self.write_audio(stem_path, source, samplerate, stem_name=stem_name)
return {stem_name: source}
def write_audio(self, stem_path: str, stem_source, samplerate, stem_name=None):
def save_audio_file(path, source):
source = spec_utils.normalize(source, self.is_normalization)
sf.write(path, source, samplerate, subtype=self.wav_type_set)
if is_not_ensemble:
save_format(path, self.save_format, self.mp3_bit_set)
def save_voc_split_instrumental(stem_name, stem_source, is_inst_invert=False):
inst_stem_name = "Instrumental (With Lead Vocals)" if stem_name == LEAD_VOCAL_STEM else "Instrumental (With Backing Vocals)"
inst_stem_path_name = LEAD_VOCAL_STEM_I if stem_name == LEAD_VOCAL_STEM else BV_VOCAL_STEM_I
inst_stem_path = self.audio_file_base_voc_split(INST_STEM, inst_stem_path_name)
stem_source = -stem_source if is_inst_invert else stem_source
inst_stem_source = spec_utils.combine_arrarys([self.master_inst_source, stem_source], is_swap=True)
save_with_message(inst_stem_path, inst_stem_name, inst_stem_source)
def save_voc_split_vocal(stem_name, stem_source):
voc_split_stem_name = LEAD_VOCAL_STEM_LABEL if stem_name == LEAD_VOCAL_STEM else BV_VOCAL_STEM_LABEL
voc_split_stem_path = self.audio_file_base_voc_split(VOCAL_STEM, stem_name)
save_with_message(voc_split_stem_path, voc_split_stem_name, stem_source)
def save_with_message(stem_path, stem_name, stem_source):
is_deverb = self.is_deverb_vocals and (
self.deverb_vocal_opt == stem_name or
(self.deverb_vocal_opt == 'ALL' and
(stem_name == VOCAL_STEM or stem_name == LEAD_VOCAL_STEM_LABEL or stem_name == BV_VOCAL_STEM_LABEL)))
self.write_to_console(f'{SAVING_STEM[0]}{stem_name}{SAVING_STEM[1]}')
if is_deverb and is_not_ensemble:
deverb_vocals(stem_path, stem_source)
save_audio_file(stem_path, stem_source)
self.write_to_console(DONE, base_text='')
def deverb_vocals(stem_path:str, stem_source):
self.write_to_console(INFERENCE_STEP_DEVERBING, base_text='')
stem_source_deverbed, stem_source_2 = vr_denoiser(stem_source, self.device, is_deverber=True, model_path=self.DEVERBER_MODEL)
save_audio_file(stem_path.replace(".wav", "_deverbed.wav"), stem_source_deverbed)
save_audio_file(stem_path.replace(".wav", "_reverb_only.wav"), stem_source_2)
is_bv_model_lead = (self.is_bv_model_rebalenced and self.is_vocal_split_model and stem_name == LEAD_VOCAL_STEM)
is_bv_rebalance_lead = (self.is_bv_model_rebalenced and self.is_vocal_split_model and stem_name == BV_VOCAL_STEM)
is_no_vocal_save = self.is_inst_only_voc_splitter and (stem_name == VOCAL_STEM or stem_name == BV_VOCAL_STEM or stem_name == LEAD_VOCAL_STEM) or is_bv_model_lead
is_not_ensemble = (not self.is_ensemble_mode or self.is_vocal_split_model)
is_do_not_save_inst = (self.is_save_vocal_only and self.is_sec_bv_rebalance and stem_name == INST_STEM)
if is_bv_rebalance_lead:
master_voc_source = spec_utils.match_array_shapes(self.master_vocal_source, stem_source, is_swap=True)
bv_rebalance_lead_source = stem_source-master_voc_source
if not is_bv_model_lead and not is_do_not_save_inst:
if self.is_vocal_split_model or not self.is_secondary_model:
if self.is_vocal_split_model and not self.is_inst_only_voc_splitter:
save_voc_split_vocal(stem_name, stem_source)
if is_bv_rebalance_lead:
save_voc_split_vocal(LEAD_VOCAL_STEM, bv_rebalance_lead_source)
else:
if not is_no_vocal_save:
save_with_message(stem_path, stem_name, stem_source)
if self.is_save_inst_vocal_splitter and not self.is_save_vocal_only:
save_voc_split_instrumental(stem_name, stem_source)
if is_bv_rebalance_lead:
save_voc_split_instrumental(LEAD_VOCAL_STEM, bv_rebalance_lead_source, is_inst_invert=True)
self.set_progress_bar(0.95)
if stem_name == VOCAL_STEM:
self.master_vocal_path = stem_path
def pitch_fix(self, source, sr_pitched, org_mix):
semitone_shift = self.semitone_shift
source = spec_utils.change_pitch_semitones(source, sr_pitched, semitone_shift=semitone_shift)[0]
source = spec_utils.match_array_shapes(source, org_mix)
return source
def match_frequency_pitch(self, mix):
source = mix
if self.is_match_frequency_pitch and self.is_pitch_change:
source, sr_pitched = spec_utils.change_pitch_semitones(mix, 44100, semitone_shift=-self.semitone_shift)
source = self.pitch_fix(source, sr_pitched, mix)
return source
class SeperateMDX(SeperateAttributes):
def seperate(self):
samplerate = 44100
if self.primary_model_name == self.model_basename and isinstance(self.primary_sources, tuple):
mix, source = self.primary_sources
self.load_cached_sources()
else:
print(f"[DIAGNOSTIC] === SeperateMDX.seperate Start ===")
if torch.cuda.is_available():
print(f"[DIAGNOSTIC] Initial GPU Alloc: {torch.cuda.memory_allocated() / (1024**2):.2f} MB, Reserved: {torch.cuda.memory_reserved() / (1024**2):.2f} MB")
print(f"[DIAGNOSTIC] Cache keys currently stored: {list(MODEL_CACHE.cache.keys())}")
self.start_inference_console_write()
precision = getattr(self.model_data, 'model_precision', MODEL_PRECISION_DEFAULT)
policy = SmartPrecisionPolicy(precision)
if not self.is_mdx_ckpt and hasattr(self, 'dim_t'):
self.mdx_segment_size = min(self.mdx_segment_size, self.dim_t)
cached_model = MODEL_CACHE.get_model(self.model_path, precision_mode=precision)
if cached_model is not None:
if self.is_mdx_ckpt:
self.model_run = cached_model
model_params = torch.load(self.model_path, map_location=lambda storage, loc: storage)['hyper_parameters']
self.dim_c, self.hop = model_params['dim_c'], model_params['hop_length']
policy.diagnose(self.model_run, 'PyTorch-CKPT', execution_provider=str(getattr(self, 'run_type', '')))
else:
if self.mdx_segment_size <= self.dim_t and not self.is_other_gpu:
ort_ = cached_model
input_type = ort_.get_inputs()[0].type
if 'float16' in input_type:
self.model_run = lambda spek: ONNX_WORKER.execute(ort_.run, None, {'input': spek.cpu().numpy().astype(np.float16)})[0].astype(np.float32)
else:
self.model_run = lambda spek: ONNX_WORKER.execute(ort_.run, None, {'input': spek.cpu().numpy()})[0]
policy.diagnose(ort_, 'ONNX', execution_provider=str(getattr(self, 'run_type', '')))
else:
self.model_run = cached_model
policy.diagnose(self.model_run, 'PyTorch-ConvertModel', execution_provider=str(getattr(self, 'run_type', '')))
else:
if self.is_mdx_ckpt:
model_params = torch.load(self.model_path, map_location=lambda storage, loc: storage)['hyper_parameters']
self.dim_c, self.hop = model_params['dim_c'], model_params['hop_length']
separator = MdxnetSet.ConvTDFNet(**model_params)
self.model_run = separator.load_from_checkpoint(self.model_path)
self.model_run = policy.cast_model(self.model_run)
self.model_run = self.model_run.to(self.device).eval()
MODEL_CACHE.set_model(self.model_path, self.model_run, size_bytes=os.path.getsize(self.model_path), precision_mode=precision)
policy.diagnose(self.model_run, 'PyTorch-CKPT', execution_provider=str(getattr(self, 'run_type', '')))
else:
if self.mdx_segment_size <= self.dim_t and not self.is_other_gpu:
ort_ = create_onnx_session(self.model_path, self.run_type, policy)
input_type = ort_.get_inputs()[0].type
if 'float16' in input_type:
self.model_run = lambda spek: ONNX_WORKER.execute(ort_.run, None, {'input': spek.cpu().numpy().astype(np.float16)})[0].astype(np.float32)
else:
self.model_run = lambda spek: ONNX_WORKER.execute(ort_.run, None, {'input': spek.cpu().numpy()})[0]
MODEL_CACHE.set_model(self.model_path, ort_, size_bytes=os.path.getsize(self.model_path), precision_mode=precision)
policy.diagnose(ort_, 'ONNX', execution_provider=str(getattr(self, 'run_type', '')))
else:
self.model_run = ConvertModel(load(self.model_path))
self.model_run = policy.cast_model(self.model_run)
self.model_run.to(self.device).eval()
MODEL_CACHE.set_model(self.model_path, self.model_run, size_bytes=os.path.getsize(self.model_path), precision_mode=precision)
policy.diagnose(self.model_run, 'PyTorch-ConvertModel', execution_provider=str(getattr(self, 'run_type', '')))
self.running_inference_console_write()
mix = prepare_mix(self.audio_file)
source = self.demix(mix)
if not self.is_vocal_split_model:
self.cache_source((mix, source))
self.write_to_console(DONE, base_text='')
mdx_net_cut = True if self.primary_stem in MDX_NET_FREQ_CUT and self.is_match_frequency_pitch else False
if self.is_secondary_model_activated and self.secondary_model:
self.secondary_source_primary, self.secondary_source_secondary = process_secondary_model(self.secondary_model, self.process_data, main_process_method=self.process_method, main_model_primary=self.primary_stem)
if not self.is_primary_stem_only:
secondary_stem_path = os.path.join(self.export_path, f'{self.audio_file_base}_({self.secondary_stem}).wav')
if not isinstance(self.secondary_source, np.ndarray):
raw_mix = self.demix(self.match_frequency_pitch(mix), is_match_mix=True) if mdx_net_cut else self.match_frequency_pitch(mix)
self.secondary_source = spec_utils.invert_stem(raw_mix, source) if self.is_invert_spec else mix.T-source.T
self.secondary_source_map = self.final_process(secondary_stem_path, self.secondary_source, self.secondary_source_secondary, self.secondary_stem, samplerate)
if not self.is_secondary_stem_only:
primary_stem_path = os.path.join(self.export_path, f'{self.audio_file_base}_({self.primary_stem}).wav')
if not isinstance(self.primary_source, np.ndarray):
self.primary_source = source.T
self.primary_source_map = self.final_process(primary_stem_path, self.primary_source, self.secondary_source_primary, self.primary_stem, samplerate)
if torch.cuda.is_available():
print(f"[DIAGNOSTIC] SeperateMDX End before clear_gpu_cache: Allocated: {torch.cuda.memory_allocated() / (1024**2):.2f} MB, Reserved: {torch.cuda.memory_reserved() / (1024**2):.2f} MB")
clear_gpu_cache()
if torch.cuda.is_available():
print(f"[DIAGNOSTIC] SeperateMDX End after clear_gpu_cache: Allocated: {torch.cuda.memory_allocated() / (1024**2):.2f} MB, Reserved: {torch.cuda.memory_reserved() / (1024**2):.2f} MB")
secondary_sources = {**self.primary_source_map, **self.secondary_source_map}
self.process_vocal_split_chain(secondary_sources)
if self.is_secondary_model or self.is_pre_proc_model:
return secondary_sources
def initialize_model_settings(self):
self.n_bins = self.n_fft//2+1
self.trim = self.n_fft//2
self.chunk_size = self.hop * (self.mdx_segment_size-1)
self.gen_size = self.chunk_size-2*self.trim
self.stft = STFT(self.n_fft, self.hop, self.dim_f, self.device)
def demix(self, mix, is_match_mix=False):
if getattr(self.model_data, 'is_adaptive_chunk', False) and not is_match_mix:
precision = getattr(self.model_data, 'model_precision', MODEL_PRECISION_DEFAULT)
scheduler = AdaptiveChunkScheduler(
device=self.device,
precision_mode=precision,
model_size_bytes=os.stat(self.model_path).st_size,
batch_size=self.mdx_batch_size
)
self.mdx_segment_size = scheduler.determine_chunk_size()
print(f"Adaptive Chunk Scheduler: Selected segment size {self.mdx_segment_size}")
while True:
try:
self.initialize_model_settings()
org_mix = mix
tar_waves_ = []
if is_match_mix:
chunk_size = self.hop * (256-1)
overlap = 0.02
else:
chunk_size = self.chunk_size
overlap = self.overlap_mdx
if getattr(self.model_data, 'is_adaptive_chunk', False):
overlap = scheduler.adjust_overlap_ratio(self.mdx_segment_size, overlap)
if self.is_pitch_change:
mix, sr_pitched = spec_utils.change_pitch_semitones(mix, 44100, semitone_shift=-self.semitone_shift)
# Convert input to PyTorch tensor
mix_tensor = torch.from_numpy(mix).to(torch.float32)
gen_size = chunk_size-2*self.trim
pad = gen_size + self.trim - ((mix_tensor.shape[-1]) % gen_size)
# PyTorch-native padding
trim_pad = torch.zeros(2, self.trim, dtype=torch.float32)
pad_pad = torch.zeros(2, pad, dtype=torch.float32)
mixture = torch.cat([trim_pad, mix_tensor, pad_pad], dim=1)
step = self.chunk_size - self.n_fft if overlap == DEFAULT else int((1 - overlap) * chunk_size)
# Setup result and divider buffers
if is_match_mix or self.device == 'cpu':
result = torch.zeros((1, 2, mixture.shape[-1]), dtype=torch.float32)
divider = torch.zeros((1, 2, mixture.shape[-1]), dtype=torch.float32)
else:
result = torch.zeros((1, 2, mixture.shape[-1]), dtype=torch.float32, device=self.device)
divider = torch.zeros((1, 2, mixture.shape[-1]), dtype=torch.float32, device=self.device)
total = 0
total_chunks = (mixture.shape[-1] + step - 1) // step
# Generate slice ranges (highly memory efficient)
slice_ranges = []
for i in range(0, mixture.shape[-1], step):
start = i
end = min(i + chunk_size, mixture.shape[-1])
slice_ranges.append((start, end))
def get_slice_tensor(idx):
if idx >= len(slice_ranges):
return None
start, end = slice_ranges[idx]
mix_part_ = mixture[:, start:end]
if end != start + chunk_size:
pad_size = (start + chunk_size) - end
pad_tensor = torch.zeros(2, pad_size, dtype=torch.float32)
mix_part_ = torch.cat([mix_part_, pad_tensor], dim=-1)
part_tensor = mix_part_.unsqueeze(0)
if self.device != 'cpu':
part_tensor = part_tensor.pin_memory()
return part_tensor
# Stream allocation
if self.device != 'cpu' and torch.cuda.is_available() and not is_match_mix:
stream_transfer = torch.cuda.Stream(device=self.device)
stream_infer = torch.cuda.Stream(device=self.device)
stream_post = torch.cuda.Stream(device=self.device)
else:
stream_transfer = None
stream_infer = None
stream_post = None
# Keep track of windows and metadata
part_metadata = []
for idx, (start, end) in enumerate(slice_ranges):
chunk_size_actual = end - start
if overlap == 0:
window = None
else:
window = np.hanning(chunk_size_actual)
window = np.tile(window[None, None, :], (1, 2, 1))
if not is_match_mix and self.device != 'cpu':
window = torch.tensor(window, dtype=torch.float32, device=self.device)
part_metadata.append((start, end, chunk_size_actual, window))
def transfer_mdx_part(part_tensor, stream):
if part_tensor is None:
return None
if stream is not None:
with torch.cuda.stream(stream):
return part_tensor.to(self.device, non_blocking=True)
else:
return part_tensor.to(self.device)
num_parts = len(slice_ranges)
total_stages = num_parts + 2
part_gpu_transferring = None
part_gpu_inferring = None
tar_waves_inferring = None
cnt_transferring = 0
cnt_inferring = 0
cnt_postprocessing = 0
for stage in range(total_stages):
# --- Stage 3: Postprocessing ---
if stage >= 2 and cnt_postprocessing < num_parts:
start, end, chunk_size_actual, window = part_metadata[cnt_postprocessing]
if stream_post is not None:
stream_post.wait_stream(stream_infer)
with torch.cuda.stream(stream_post):
if window is not None:
tar_waves_inferring[..., :chunk_size_actual] *= window
divider[..., start:end] += window
else:
divider[..., start:end] += 1
result[..., start:end] += tar_waves_inferring[..., :end-start]
else:
if window is not None:
tar_waves_inferring[..., :chunk_size_actual] *= window
divider[..., start:end] += window
else:
divider[..., start:end] += 1
result[..., start:end] += tar_waves_inferring[..., :end-start]
cnt_postprocessing += 1
# --- Stage 2: Inference ---
if stage >= 1 and cnt_inferring < num_parts:
self.running_inference_progress_bar(total_chunks, is_match_mix=is_match_mix)
if stream_infer is not None:
stream_infer.wait_stream(stream_transfer)
with torch.cuda.stream(stream_infer):
mix_waves = part_gpu_inferring.split(self.mdx_batch_size)
waves_list = []
for mix_wave in mix_waves:
waves_list.append(self.run_model(mix_wave, is_match_mix=is_match_mix))
if not is_match_mix:
tar_waves_inferring = torch.cat(waves_list, dim=0)
else:
tar_waves_inferring = np.concatenate(waves_list, axis=0)
else:
mix_waves = part_gpu_inferring.split(self.mdx_batch_size)
waves_list = []
for mix_wave in mix_waves:
waves_list.append(self.run_model(mix_wave, is_match_mix=is_match_mix))
if not is_match_mix and isinstance(waves_list[0], torch.Tensor):
tar_waves_inferring = torch.cat(waves_list, dim=0)
else:
tar_waves_inferring = np.concatenate(waves_list, axis=0)
cnt_inferring += 1
# --- Stage 1: Transfer ---
if stage < num_parts:
part_tensor_transferring = get_slice_tensor(stage)
part_gpu_transferring = transfer_mdx_part(part_tensor_transferring, stream_transfer)
else:
part_gpu_transferring = None
# Shift buffer for next stage
part_gpu_inferring = part_gpu_transferring
# Synchronize streams
if stream_post is not None:
torch.cuda.synchronize(self.device)
tar_waves = result / divider
# Trim and crop in PyTorch
tar_waves = tar_waves[:, :, self.trim:-self.trim]
tar_waves = tar_waves[0, :, :mix.shape[-1]]
source = tar_waves.cpu().detach().numpy()
if self.is_pitch_change and not is_match_mix:
source = self.pitch_fix(source, sr_pitched, org_mix)
source = source if is_match_mix else source*self.compensate
if self.is_denoise_model and not is_match_mix:
if NO_STEM in self.primary_stem_native or self.primary_stem_native == INST_STEM:
if org_mix.shape[1] != source.shape[1]:
source = spec_utils.match_array_shapes(source, org_mix)
source = org_mix - vr_denoiser(org_mix-source, self.device, model_path=self.DENOISER_MODEL)
else:
source = vr_denoiser(source, self.device, model_path=self.DENOISER_MODEL)
return source
except RuntimeError as e:
if "out of memory" in str(e).lower() and self.mdx_segment_size > 256 and not is_match_mix:
print(f"CUDA Out Of Memory with segment size {self.mdx_segment_size}. Retrying with smaller segment size...")
self.mdx_segment_size = max(256, self.mdx_segment_size // 2)
torch.cuda.empty_cache()
gc.collect()
continue
else:
raise e
def run_model(self, mix, is_match_mix=False):
with torch.no_grad():
spek = self.stft(mix.to(self.device))*self.adjust
spek[:, :, :3, :] *= 0
precision = getattr(self.model_data, 'model_precision', MODEL_PRECISION_DEFAULT)
policy = SmartPrecisionPolicy(precision)
is_pytorch_model = isinstance(self.model_run, torch.nn.Module)
# Pad spek to self.dim_t if it is smaller
pad_t = 0
if not is_match_mix and hasattr(self, 'dim_t') and spek.shape[-1] < self.dim_t:
pad_t = self.dim_t - spek.shape[-1]
spek = torch.nn.functional.pad(spek, (0, pad_t))
if is_pytorch_model and not is_match_mix:
spek = policy.prepare_spectrogram(spek)
if is_match_mix:
spec_pred = spek.to(torch.float32).cpu().numpy()
else:
with torch.cuda.amp.autocast(enabled=(policy.enabled_amp and is_pytorch_model), dtype=policy.amp_dtype):
spec_pred = -self.model_run(-spek)*0.5+self.model_run(spek)*0.5 if self.is_denoise else self.model_run(spek)
if is_pytorch_model:
spec_pred = spec_pred.to(torch.float32)
# Crop back to original length if padded
if pad_t > 0:
spec_pred = spec_pred[..., :-pad_t]
if not isinstance(spec_pred, torch.Tensor):
spec_pred = torch.tensor(spec_pred).to(self.device)
if is_match_mix:
return self.stft.inverse(spec_pred).cpu().detach().numpy()
else:
return self.stft.inverse(spec_pred)
class SeperateMDXC(SeperateAttributes):
def seperate(self):
samplerate = 44100
sources = None
if self.primary_model_name == self.model_basename and isinstance(self.primary_sources, tuple):
mix, sources = self.primary_sources
self.load_cached_sources()
else:
self.start_inference_console_write()
self.running_inference_console_write()
mix = prepare_mix(self.audio_file)
sources = self.demix(mix)
if not self.is_vocal_split_model:
self.cache_source((mix, sources))
self.write_to_console(DONE, base_text='')
stem_list = [self.mdx_c_configs.training.target_instrument] if self.mdx_c_configs.training.target_instrument else [i for i in self.mdx_c_configs.training.instruments]
if self.is_secondary_model:
if self.is_pre_proc_model:
self.mdxnet_stem_select = stem_list[0]
else:
self.mdxnet_stem_select = self.main_model_primary_stem_4_stem if self.main_model_primary_stem_4_stem else self.primary_model_primary_stem
self.primary_stem = self.mdxnet_stem_select
self.secondary_stem = secondary_stem(self.mdxnet_stem_select)
self.is_primary_stem_only, self.is_secondary_stem_only = False, False
is_all_stems = self.mdxnet_stem_select == ALL_STEMS
is_not_ensemble_master = not self.process_data['is_ensemble_master']
is_not_single_stem = not len(stem_list) <= 2
is_not_secondary_model = not self.is_secondary_model
is_ensemble_4_stem = self.is_4_stem_ensemble and is_not_single_stem
if (is_all_stems and is_not_ensemble_master and is_not_single_stem and is_not_secondary_model) or is_ensemble_4_stem and not self.is_pre_proc_model:
for stem in stem_list:
primary_stem_path = os.path.join(self.export_path, f'{self.audio_file_base}_({stem}).wav')
self.primary_source = sources[stem].T
self.write_audio(primary_stem_path, self.primary_source, samplerate, stem_name=stem)
if stem == VOCAL_STEM and not self.is_sec_bv_rebalance:
self.process_vocal_split_chain({VOCAL_STEM:stem})
else:
if len(stem_list) == 1:
source_primary = sources
else:
source_primary = sources[stem_list[0]] if self.is_multi_stem_ensemble and len(stem_list) == 2 else sources[self.mdxnet_stem_select]
if self.is_secondary_model_activated and self.secondary_model:
self.secondary_source_primary, self.secondary_source_secondary = process_secondary_model(self.secondary_model,
self.process_data,
main_process_method=self.process_method,
main_model_primary=self.primary_stem)
if not self.is_primary_stem_only:
secondary_stem_path = os.path.join(self.export_path, f'{self.audio_file_base}_({self.secondary_stem}).wav')
if not isinstance(self.secondary_source, np.ndarray):
if self.is_mdx_combine_stems and len(stem_list) >= 2:
if len(stem_list) == 2:
secondary_source = sources[self.secondary_stem]
else:
sources.pop(self.primary_stem)
next_stem = next(iter(sources))
secondary_source = np.zeros_like(sources[next_stem])
for v in sources.values():
secondary_source += v
self.secondary_source = secondary_source.T
else:
self.secondary_source, raw_mix = source_primary, self.match_frequency_pitch(mix)
self.secondary_source = spec_utils.to_shape(self.secondary_source, raw_mix.shape)
if self.is_invert_spec:
self.secondary_source = spec_utils.invert_stem(raw_mix, self.secondary_source)
else:
self.secondary_source = (-self.secondary_source.T+raw_mix.T)
self.secondary_source_map = self.final_process(secondary_stem_path, self.secondary_source, self.secondary_source_secondary, self.secondary_stem, samplerate)
if not self.is_secondary_stem_only:
primary_stem_path = os.path.join(self.export_path, f'{self.audio_file_base}_({self.primary_stem}).wav')
if not isinstance(self.primary_source, np.ndarray):
self.primary_source = source_primary.T
self.primary_source_map = self.final_process(primary_stem_path, self.primary_source, self.secondary_source_primary, self.primary_stem, samplerate)
clear_gpu_cache()
secondary_sources = {**self.primary_source_map, **self.secondary_source_map}
self.process_vocal_split_chain(secondary_sources)
if self.is_secondary_model or self.is_pre_proc_model:
return secondary_sources
def demix(self, mix):
sr_pitched = 441000
org_mix = mix
if self.is_pitch_change:
mix, sr_pitched = spec_utils.change_pitch_semitones(mix, 44100, semitone_shift=-self.semitone_shift)
precision = getattr(self.model_data, 'model_precision', MODEL_PRECISION_DEFAULT)
policy = SmartPrecisionPolicy(precision)
model = MODEL_CACHE.get_model(self.model_path, precision_mode=precision)
if model is None:
model = TFC_TDF_net(self.mdx_c_configs, device=self.device)
model.load_state_dict(torch.load(self.model_path, map_location=cpu))
model = policy.cast_model(model)
model.to(self.device).eval()
MODEL_CACHE.set_model(self.model_path, model, size_bytes=os.path.getsize(self.model_path), precision_mode=precision)
policy.diagnose(model, 'PyTorch-MDXC', execution_provider=str(getattr(self, 'run_type', '')))
try:
S = model.num_target_instruments
except Exception as e:
S = model.module.num_target_instruments
mdx_segment_size = self.mdx_c_configs.inference.dim_t if self.is_mdx_c_seg_def else self.mdx_segment_size
if getattr(self.model_data, 'is_adaptive_chunk', False):
scheduler = AdaptiveChunkScheduler(
device=self.device,
precision_mode=precision,
model_size_bytes=os.stat(self.model_path).st_size,
batch_size=self.mdx_batch_size
)
mdx_segment_size = scheduler.determine_chunk_size()
print(f"Adaptive Chunk Scheduler: Selected segment size {mdx_segment_size}")
while True:
try:
mix_tensor = torch.tensor(mix, dtype=torch.float32)
batch_size = self.mdx_batch_size
chunk_size = self.mdx_c_configs.audio.hop_length * (mdx_segment_size - 1)
overlap = self.overlap_mdx23
if getattr(self.model_data, 'is_adaptive_chunk', False):
overlap = int(scheduler.adjust_overlap_ratio(mdx_segment_size, overlap))
overlap = max(1, overlap)
hop_size = chunk_size // overlap
mix_shape = mix_tensor.shape[1]
pad_size = hop_size - (mix_shape - chunk_size) % hop_size
mix_tensor = torch.cat([torch.zeros(2, chunk_size - hop_size), mix_tensor, torch.zeros(2, pad_size + chunk_size - hop_size)], 1)
chunks = mix_tensor.unfold(1, chunk_size, hop_size).transpose(0, 1)
num_chunks = len(chunks)
num_batches = (num_chunks + batch_size - 1) // batch_size
def get_mdxc_batch(batch_idx):
if batch_idx >= num_batches:
return None
start = batch_idx * batch_size
end = min(start + batch_size, num_chunks)
batch = chunks[start:end]
if self.device != 'cpu':
batch = batch.pin_memory()
return batch
X = torch.zeros(S, *mix_tensor.shape) if S > 1 else torch.zeros_like(mix_tensor)
X = X.to(self.device)
# Stream allocation
if self.device != 'cpu' and torch.cuda.is_available():
stream_transfer = torch.cuda.Stream(device=self.device)
stream_infer = torch.cuda.Stream(device=self.device)
stream_post = torch.cuda.Stream(device=self.device)
else:
stream_transfer = None
stream_infer = None
stream_post = None
def transfer_mdxc_batch(batch_cpu, stream):
if batch_cpu is None:
return None
if stream is not None:
with torch.cuda.stream(stream):
batch_gpu = batch_cpu.to(self.device, non_blocking=True)
if policy.is_fp16:
batch_gpu = batch_gpu.half()
elif policy.is_bf16:
batch_gpu = batch_gpu.bfloat16()
return batch_gpu
else:
batch_gpu = batch_cpu.to(self.device)
if policy.is_fp16:
batch_gpu = batch_gpu.half()
elif policy.is_bf16:
batch_gpu = batch_gpu.bfloat16()
return batch_gpu
total_stages = num_batches + 2
batch_gpu_transferring = None
batch_gpu_inferring = None
x_inferring = None
cnt_transferring = 0
cnt_inferring = 0
cnt_postprocessing = 0
with torch.no_grad():
enabled_amp = policy.enabled_amp
amp_dtype = policy.amp_dtype
for stage in range(total_stages):
# --- Stage 3: Postprocessing ---
if stage >= 2 and cnt_postprocessing < num_batches:
if stream_post is not None:
stream_post.wait_stream(stream_infer)
with torch.cuda.stream(stream_post):
x_post = x_inferring.to(torch.float32)
start_idx = cnt_postprocessing * batch_size
for w_idx, w in enumerate(x_post):
c_idx = start_idx + w_idx
X[..., c_idx * hop_size : c_idx * hop_size + chunk_size] += w
else:
x_post = x_inferring.to(torch.float32)
start_idx = cnt_postprocessing * batch_size
for w_idx, w in enumerate(x_post):
c_idx = start_idx + w_idx
X[..., c_idx * hop_size : c_idx * hop_size + chunk_size] += w
cnt_postprocessing += 1
# --- Stage 2: Inference ---
if stage >= 1 and cnt_inferring < num_batches:
self.running_inference_progress_bar(num_batches)
if stream_infer is not None:
stream_infer.wait_stream(stream_transfer)
with torch.cuda.stream(stream_infer):
with torch.cuda.amp.autocast(enabled=enabled_amp, dtype=amp_dtype):
x_inferring = model(batch_gpu_inferring)
else:
with torch.cuda.amp.autocast(enabled=enabled_amp, dtype=amp_dtype):
x_inferring = model(batch_gpu_inferring)
cnt_inferring += 1
# --- Stage 1: Transfer ---
if stage < num_batches:
batch_cpu_transferring = get_mdxc_batch(stage)
batch_gpu_transferring = transfer_mdxc_batch(batch_cpu_transferring, stream_transfer)
else:
batch_gpu_transferring = None
# Shift buffer for next stage
batch_gpu_inferring = batch_gpu_transferring
if stream_post is not None:
torch.cuda.synchronize(self.device)
estimated_sources = X[..., chunk_size - hop_size:-(pad_size + chunk_size - hop_size)] / overlap
del X
break
except RuntimeError as e:
if "out of memory" in str(e).lower() and mdx_segment_size > 256:
print(f"CUDA Out Of Memory with segment size {mdx_segment_size}. Retrying with smaller segment size...")
mdx_segment_size = max(256, mdx_segment_size // 2)
torch.cuda.empty_cache()
gc.collect()
continue
else:
raise e
pitch_fix = lambda s:self.pitch_fix(s, sr_pitched, org_mix)
if S > 1:
sources = {k: pitch_fix(v) if self.is_pitch_change else v for k, v in zip(self.mdx_c_configs.training.instruments, estimated_sources.cpu().detach().numpy())}
del estimated_sources
if self.is_denoise_model:
if VOCAL_STEM in sources.keys() and INST_STEM in sources.keys():
sources[VOCAL_STEM] = vr_denoiser(sources[VOCAL_STEM], self.device, model_path=self.DENOISER_MODEL)
if sources[VOCAL_STEM].shape[1] != org_mix.shape[1]:
sources[VOCAL_STEM] = spec_utils.match_array_shapes(sources[VOCAL_STEM], org_mix)
sources[INST_STEM] = org_mix - sources[VOCAL_STEM]
return sources
else:
est_s = estimated_sources.cpu().detach().numpy()
del estimated_sources
return pitch_fix(est_s) if self.is_pitch_change else est_s
class SeperateDemucs(SeperateAttributes):
def seperate(self):
samplerate = 44100
source = None
model_scale = None
stem_source = None
stem_source_secondary = None
inst_mix = None
inst_source = None
is_no_write = False
is_no_piano_guitar = False
is_no_cache = False
if self.primary_model_name == self.model_basename and isinstance(self.primary_sources, np.ndarray) and not self.pre_proc_model:
source = self.primary_sources
self.load_cached_sources()
else:
self.start_inference_console_write()
is_no_cache = True
mix = prepare_mix(self.audio_file)
if is_no_cache:
precision = getattr(self.model_data, 'model_precision', MODEL_PRECISION_DEFAULT)
policy = SmartPrecisionPolicy(precision)
self.demucs = MODEL_CACHE.get_model(self.model_path, precision_mode=precision)
if self.demucs is None:
if self.demucs_version == DEMUCS_V1:
if str(self.model_path).endswith(".gz"):
self.model_path = gzip.open(self.model_path, "rb")
klass, args, kwargs, state = torch.load(self.model_path)
self.demucs = klass(*args, **kwargs)
self.demucs.load_state_dict(state)
elif self.demucs_version == DEMUCS_V2:
self.demucs = auto_load_demucs_model_v2(self.demucs_source_list, self.model_path)
self.demucs.load_state_dict(torch.load(self.model_path))
else:
self.demucs = HDemucs(sources=self.demucs_source_list)
self.demucs = _gm(name=os.path.splitext(os.path.basename(self.model_path))[0],
repo=Path(os.path.dirname(self.model_path)))
self.demucs = demucs_segments(self.segment, self.demucs)
self.demucs = policy.cast_model(self.demucs)
self.demucs.to(self.device)
self.demucs.eval()
MODEL_CACHE.set_model(self.model_path, self.demucs, size_bytes=os.path.getsize(self.model_path), precision_mode=precision)
policy.diagnose(self.demucs, 'PyTorch-Demucs', execution_provider=str(getattr(self, 'run_type', '')))
if self.pre_proc_model:
if self.primary_stem not in [VOCAL_STEM, INST_STEM]:
is_no_write = True
self.write_to_console(DONE, base_text='')
mix_no_voc = process_secondary_model(self.pre_proc_model, self.process_data, is_pre_proc_model=True)
inst_mix = prepare_mix(mix_no_voc[INST_STEM])
self.process_iteration()
self.running_inference_console_write(is_no_write=is_no_write)
inst_source = self.demix_demucs(inst_mix)
self.process_iteration()
self.running_inference_console_write(is_no_write=is_no_write) if not self.pre_proc_model else None
if self.primary_model_name == self.model_basename and isinstance(self.primary_sources, np.ndarray) and self.pre_proc_model:
source = self.primary_sources
else:
source = self.demix_demucs(mix)
self.write_to_console(DONE, base_text='')
del self.demucs
clear_gpu_cache()
if isinstance(inst_source, np.ndarray):
source_reshape = spec_utils.reshape_sources(inst_source[self.demucs_source_map[VOCAL_STEM]], source[self.demucs_source_map[VOCAL_STEM]])
inst_source[self.demucs_source_map[VOCAL_STEM]] = source_reshape
source = inst_source
if isinstance(source, np.ndarray):
if len(source) == 2:
self.demucs_source_map = DEMUCS_2_SOURCE_MAPPER
else:
self.demucs_source_map = DEMUCS_6_SOURCE_MAPPER if len(source) == 6 else DEMUCS_4_SOURCE_MAPPER
if len(source) == 6 and self.process_data['is_ensemble_master'] or len(source) == 6 and self.is_secondary_model:
is_no_piano_guitar = True
six_stem_other_source = list(source)
six_stem_other_source = [i for n, i in enumerate(source) if n in [self.demucs_source_map[OTHER_STEM], self.demucs_source_map[GUITAR_STEM], self.demucs_source_map[PIANO_STEM]]]
other_source = np.zeros_like(six_stem_other_source[0])
for i in six_stem_other_source:
other_source += i
source_reshape = spec_utils.reshape_sources(source[self.demucs_source_map[OTHER_STEM]], other_source)
source[self.demucs_source_map[OTHER_STEM]] = source_reshape
if not self.is_vocal_split_model:
self.cache_source(source)
if (self.demucs_stems == ALL_STEMS and not self.process_data['is_ensemble_master']) or self.is_4_stem_ensemble and not self.is_return_dual:
for stem_name, stem_value in self.demucs_source_map.items():
if self.is_secondary_model_activated and not self.is_secondary_model and not stem_value >= 4:
if self.secondary_model_4_stem[stem_value]:
model_scale = self.secondary_model_4_stem_scale[stem_value]
stem_source_secondary = process_secondary_model(self.secondary_model_4_stem[stem_value], self.process_data, main_model_primary_stem_4_stem=stem_name, is_source_load=True, is_return_dual=False)
if isinstance(stem_source_secondary, np.ndarray):
stem_source_secondary = stem_source_secondary[1 if self.secondary_model_4_stem[stem_value].demucs_stem_count == 2 else stem_value].T
elif type(stem_source_secondary) is dict:
stem_source_secondary = stem_source_secondary[stem_name]
stem_source_secondary = None if stem_value >= 4 else stem_source_secondary
stem_path = os.path.join(self.export_path, f'{self.audio_file_base}_({stem_name}).wav')
stem_source = source[stem_value].T
stem_source = self.process_secondary_stem(stem_source, secondary_model_source=stem_source_secondary, model_scale=model_scale)
self.write_audio(stem_path, stem_source, samplerate, stem_name=stem_name)
if stem_name == VOCAL_STEM and not self.is_sec_bv_rebalance:
self.process_vocal_split_chain({VOCAL_STEM:stem_source})
if self.is_secondary_model:
return source
else:
if self.is_secondary_model_activated and self.secondary_model:
self.secondary_source_primary, self.secondary_source_secondary = process_secondary_model(self.secondary_model, self.process_data, main_process_method=self.process_method)
if not self.is_primary_stem_only:
def secondary_save(sec_stem_name, source, raw_mixture=None, is_inst_mixture=False):
secondary_source = self.secondary_source if not is_inst_mixture else None
secondary_stem_path = os.path.join(self.export_path, f'{self.audio_file_base}_({sec_stem_name}).wav')
secondary_source_secondary = None
if not isinstance(secondary_source, np.ndarray):
if self.is_demucs_combine_stems:
source = list(source)
if is_inst_mixture:
source = [i for n, i in enumerate(source) if not n in [self.demucs_source_map[self.primary_stem], self.demucs_source_map[VOCAL_STEM]]]
else:
source.pop(self.demucs_source_map[self.primary_stem])
source = source[:len(source) - 2] if is_no_piano_guitar else source
secondary_source = np.zeros_like(source[0])
for i in source:
secondary_source += i
secondary_source = secondary_source.T
else:
if not isinstance(raw_mixture, np.ndarray):
raw_mixture = prepare_mix(self.audio_file)
secondary_source = source[self.demucs_source_map[self.primary_stem]]
if self.is_invert_spec:
secondary_source = spec_utils.invert_stem(raw_mixture, secondary_source)
else:
raw_mixture = spec_utils.reshape_sources(secondary_source, raw_mixture)
secondary_source = (-secondary_source.T+raw_mixture.T)
if not is_inst_mixture:
self.secondary_source = secondary_source
secondary_source_secondary = self.secondary_source_secondary
self.secondary_source = self.process_secondary_stem(secondary_source, secondary_source_secondary)
self.secondary_source_map = {self.secondary_stem: self.secondary_source}
self.write_audio(secondary_stem_path, secondary_source, samplerate, stem_name=sec_stem_name)
secondary_save(self.secondary_stem, source, raw_mixture=mix)
if self.is_demucs_pre_proc_model_inst_mix and self.pre_proc_model and not self.is_4_stem_ensemble:
secondary_save(f"{self.secondary_stem} {INST_STEM}", source, raw_mixture=inst_mix, is_inst_mixture=True)
if not self.is_secondary_stem_only:
primary_stem_path = os.path.join(self.export_path, f'{self.audio_file_base}_({self.primary_stem}).wav')
if not isinstance(self.primary_source, np.ndarray):
self.primary_source = source[self.demucs_source_map[self.primary_stem]].T
self.primary_source_map = self.final_process(primary_stem_path, self.primary_source, self.secondary_source_primary, self.primary_stem, samplerate)
secondary_sources = {**self.primary_source_map, **self.secondary_source_map}
self.process_vocal_split_chain(secondary_sources)
if self.is_secondary_model:
return secondary_sources
def demix_demucs(self, mix):
org_mix = mix
if self.is_pitch_change:
mix, sr_pitched = spec_utils.change_pitch_semitones(mix, 44100, semitone_shift=-self.semitone_shift)
processed = {}
mix = torch.tensor(mix, dtype=torch.float32)
ref = mix.mean(0)
mix = (mix - ref.mean()) / ref.std()
mix_infer = mix
precision = getattr(self.model_data, 'model_precision', MODEL_PRECISION_DEFAULT)
policy = SmartPrecisionPolicy(precision)
enabled_amp = policy.enabled_amp
amp_dtype = policy.amp_dtype
mix_infer_dev = mix_infer.to(self.device)
if policy.is_fp16:
mix_infer_dev = mix_infer_dev.half()
elif policy.is_bf16:
mix_infer_dev = mix_infer_dev.bfloat16()
with torch.no_grad():
with torch.cuda.amp.autocast(enabled=enabled_amp, dtype=amp_dtype):
if self.demucs_version == DEMUCS_V1:
sources = apply_model_v1(self.demucs,
mix_infer_dev,
self.shifts,
self.is_split_mode,
set_progress_bar=self.set_progress_bar)
elif self.demucs_version == DEMUCS_V2:
sources = apply_model_v2(self.demucs,
mix_infer_dev,
self.shifts,
self.is_split_mode,
self.overlap,
set_progress_bar=self.set_progress_bar)
else:
sources = apply_model(self.demucs,
mix_infer_dev[None],
self.shifts,
self.is_split_mode,
self.overlap,
static_shifts=1 if self.shifts == 0 else self.shifts,
set_progress_bar=self.set_progress_bar,
device=self.device)[0]
sources = sources.to(torch.float32)
sources = (sources * ref.std() + ref.mean()).cpu().numpy()
sources[[0,1]] = sources[[1,0]]
processed[mix] = sources[:,:,0:None].copy()
sources = list(processed.values())
sources = [s[:,:,0:None] for s in sources]
#sources = [self.pitch_fix(s[:,:,0:None], sr_pitched, org_mix) if self.is_pitch_change else s[:,:,0:None] for s in sources]
sources = np.concatenate(sources, axis=-1)
if self.is_pitch_change:
sources = np.stack([self.pitch_fix(stem, sr_pitched, org_mix) for stem in sources])
return sources
class SeperateVR(SeperateAttributes):
def seperate(self):
if self.primary_model_name == self.model_basename and isinstance(self.primary_sources, tuple):
y_spec, v_spec = self.primary_sources
self.load_cached_sources()
else:
self.start_inference_console_write()
device = self.device
nn_arch_sizes = [
31191, # default
33966, 56817, 123821, 123812, 129605, 218409, 537238, 537227]
vr_5_1_models = [56817, 218409]
model_size = math.ceil(os.stat(self.model_path).st_size / 1024)
nn_arch_size = min(nn_arch_sizes, key=lambda x:abs(x-model_size))
precision = getattr(self.model_data, 'model_precision', MODEL_PRECISION_DEFAULT)
policy = SmartPrecisionPolicy(precision)
self.model_run = MODEL_CACHE.get_model(self.model_path, precision_mode=precision)
if self.model_run is None:
if nn_arch_size in vr_5_1_models or self.is_vr_51_model:
self.model_run = nets_new.CascadedNet(self.mp.param['bins'] * 2,
nn_arch_size,
nout=self.model_capacity[0],
nout_lstm=self.model_capacity[1])
self.is_vr_51_model = True
else:
self.model_run = nets.determine_model_capacity(self.mp.param['bins'] * 2, nn_arch_size)
self.model_run.load_state_dict(torch.load(self.model_path, map_location=cpu))
self.model_run = policy.cast_model(self.model_run)
self.model_run.to(device)
MODEL_CACHE.set_model(self.model_path, self.model_run, size_bytes=os.path.getsize(self.model_path), precision_mode=precision)
policy.diagnose(self.model_run, 'PyTorch-VR', execution_provider=str(getattr(self, 'run_type', '')))
self.running_inference_console_write()
if getattr(self.model_data, 'is_adaptive_chunk', False):
scheduler = AdaptiveChunkScheduler(
device=device,
precision_mode=precision,
model_size_bytes=os.stat(self.model_path).st_size,
batch_size=self.batch_size
)
sched_chunk = scheduler.determine_chunk_size()
if sched_chunk >= 4096:
self.window_size = 1024
elif sched_chunk >= 2048:
self.window_size = 512
else:
self.window_size = 320
print(f"Adaptive Chunk Scheduler: Selected window_size {self.window_size}")
y_spec, v_spec = self.inference_vr(self.loading_mix(), device, self.aggressiveness)
if not self.is_vocal_split_model:
self.cache_source((y_spec, v_spec))
self.write_to_console(DONE, base_text='')
if self.is_secondary_model_activated and self.secondary_model:
self.secondary_source_primary, self.secondary_source_secondary = process_secondary_model(self.secondary_model, self.process_data, main_process_method=self.process_method, main_model_primary=self.primary_stem)
if not self.is_secondary_stem_only:
primary_stem_path = os.path.join(self.export_path, f'{self.audio_file_base}_({self.primary_stem}).wav')
if not isinstance(self.primary_source, np.ndarray):
self.primary_source = self.spec_to_wav(y_spec).T
if not self.model_samplerate == 44100:
self.primary_source = librosa.resample(self.primary_source.T, orig_sr=self.model_samplerate, target_sr=44100).T
self.primary_source_map = self.final_process(primary_stem_path, self.primary_source, self.secondary_source_primary, self.primary_stem, 44100)
if not self.is_primary_stem_only:
secondary_stem_path = os.path.join(self.export_path, f'{self.audio_file_base}_({self.secondary_stem}).wav')
if not isinstance(self.secondary_source, np.ndarray):
self.secondary_source = self.spec_to_wav(v_spec).T
if not self.model_samplerate == 44100:
self.secondary_source = librosa.resample(self.secondary_source.T, orig_sr=self.model_samplerate, target_sr=44100).T
self.secondary_source_map = self.final_process(secondary_stem_path, self.secondary_source, self.secondary_source_secondary, self.secondary_stem, 44100)
clear_gpu_cache()
secondary_sources = {**self.primary_source_map, **self.secondary_source_map}
self.process_vocal_split_chain(secondary_sources)
if self.is_secondary_model:
return secondary_sources
def loading_mix(self):
X_wave, X_spec_s = {}, {}
bands_n = len(self.mp.param['band'])
audio_file = spec_utils.write_array_to_mem(self.audio_file, subtype=self.wav_type_set)
is_mp3 = audio_file.endswith('.mp3') if isinstance(audio_file, str) else False
for d in range(bands_n, 0, -1):
bp = self.mp.param['band'][d]
if OPERATING_SYSTEM == 'Darwin':
wav_resolution = 'polyphase' if SYSTEM_PROC == ARM or ARM in SYSTEM_ARCH else bp['res_type']
else:
wav_resolution = bp['res_type']
if d == bands_n: # high-end band
X_wave[d], _ = librosa.load(audio_file, bp['sr'], False, dtype=np.float32, res_type=wav_resolution)
X_spec_s[d] = spec_utils.wave_to_spectrogram(X_wave[d], bp['hl'], bp['n_fft'], self.mp, band=d, is_v51_model=self.is_vr_51_model)
if not np.any(X_wave[d]) and is_mp3:
X_wave[d] = rerun_mp3(audio_file, bp['sr'])
if X_wave[d].ndim == 1:
X_wave[d] = np.asarray([X_wave[d], X_wave[d]])
else: # lower bands
X_wave[d] = librosa.resample(X_wave[d+1], self.mp.param['band'][d+1]['sr'], bp['sr'], res_type=wav_resolution)
X_spec_s[d] = spec_utils.wave_to_spectrogram(X_wave[d], bp['hl'], bp['n_fft'], self.mp, band=d, is_v51_model=self.is_vr_51_model)
if d == bands_n and self.high_end_process != 'none':
self.input_high_end_h = (bp['n_fft']//2 - bp['crop_stop']) + (self.mp.param['pre_filter_stop'] - self.mp.param['pre_filter_start'])
self.input_high_end = X_spec_s[d][:, bp['n_fft']//2-self.input_high_end_h:bp['n_fft']//2, :]
X_spec = spec_utils.combine_spectrograms(X_spec_s, self.mp, is_v51_model=self.is_vr_51_model)
del X_wave, X_spec_s, audio_file
return X_spec
def inference_vr(self, X_spec, device, aggressiveness):
def _execute(X_mag_pad, roi_size):
X_dataset = []
patches = (X_mag_pad.shape[2] - 2 * self.model_run.offset) // roi_size
total_iterations = patches//self.batch_size if not self.is_tta else (patches//self.batch_size)*2
for i in range(patches):
start = i * roi_size
X_mag_window = X_mag_pad[:, :, start:start + self.window_size]
X_dataset.append(X_mag_window)
X_dataset = np.asarray(X_dataset)
self.model_run.eval()
with torch.no_grad():
mask = []
for i in range(0, patches, self.batch_size):
self.progress_value += 1
if self.progress_value >= total_iterations:
self.progress_value = total_iterations
self.set_progress_bar(0.1, 0.8/total_iterations*self.progress_value)
precision = getattr(self.model_data, 'model_precision', MODEL_PRECISION_DEFAULT)
policy = SmartPrecisionPolicy(precision)
enabled_amp = policy.enabled_amp
amp_dtype = policy.amp_dtype
X_batch = X_dataset[i: i + self.batch_size]
X_batch = torch.from_numpy(X_batch).to(device)
if policy.is_fp16:
X_batch = X_batch.half()
elif policy.is_bf16:
X_batch = X_batch.bfloat16()
with torch.cuda.amp.autocast(enabled=enabled_amp, dtype=amp_dtype):
pred = self.model_run.predict_mask(X_batch)
if not pred.size()[3] > 0:
raise Exception(ERROR_MAPPER[WINDOW_SIZE_ERROR])
pred = pred.to(torch.float32).detach().cpu().numpy()
pred = np.concatenate(pred, axis=2)
mask.append(pred)
if len(mask) == 0:
raise Exception(ERROR_MAPPER[WINDOW_SIZE_ERROR])
mask = np.concatenate(mask, axis=2)
return mask
def postprocess(mask, X_mag, X_phase):
is_non_accom_stem = False
for stem in NON_ACCOM_STEMS:
if stem == self.primary_stem:
is_non_accom_stem = True
mask = spec_utils.adjust_aggr(mask, is_non_accom_stem, aggressiveness)
if self.is_post_process:
mask = spec_utils.merge_artifacts(mask, thres=self.post_process_threshold)
y_spec = mask * X_mag * np.exp(1.j * X_phase)
v_spec = (1 - mask) * X_mag * np.exp(1.j * X_phase)
return y_spec, v_spec
X_mag, X_phase = spec_utils.preprocess(X_spec)
n_frame = X_mag.shape[2]
pad_l, pad_r, roi_size = spec_utils.make_padding(n_frame, self.window_size, self.model_run.offset)
X_mag_pad = np.pad(X_mag, ((0, 0), (0, 0), (pad_l, pad_r)), mode='constant')
X_mag_pad /= X_mag_pad.max()
mask = _execute(X_mag_pad, roi_size)
if self.is_tta:
pad_l += roi_size // 2
pad_r += roi_size // 2
X_mag_pad = np.pad(X_mag, ((0, 0), (0, 0), (pad_l, pad_r)), mode='constant')
X_mag_pad /= X_mag_pad.max()
mask_tta = _execute(X_mag_pad, roi_size)
mask_tta = mask_tta[:, :, roi_size // 2:]
mask = (mask[:, :, :n_frame] + mask_tta[:, :, :n_frame]) * 0.5
else:
mask = mask[:, :, :n_frame]
y_spec, v_spec = postprocess(mask, X_mag, X_phase)
return y_spec, v_spec
def spec_to_wav(self, spec):
if self.high_end_process.startswith('mirroring') and isinstance(self.input_high_end, np.ndarray) and self.input_high_end_h:
input_high_end_ = spec_utils.mirroring(self.high_end_process, spec, self.input_high_end, self.mp)
wav = spec_utils.cmb_spectrogram_to_wave(spec, self.mp, self.input_high_end_h, input_high_end_, is_v51_model=self.is_vr_51_model)
else:
wav = spec_utils.cmb_spectrogram_to_wave(spec, self.mp, is_v51_model=self.is_vr_51_model)
return wav
def process_secondary_model(secondary_model: ModelData,
process_data,
main_model_primary_stem_4_stem=None,
is_source_load=False,
main_process_method=None,
is_pre_proc_model=False,
is_return_dual=True,
main_model_primary=None):
if not is_pre_proc_model:
process_iteration = process_data['process_iteration']
process_iteration()
if secondary_model.process_method == VR_ARCH_TYPE:
seperator = SeperateVR(secondary_model, process_data, main_model_primary_stem_4_stem=main_model_primary_stem_4_stem, main_process_method=main_process_method, main_model_primary=main_model_primary)
if secondary_model.process_method == MDX_ARCH_TYPE:
if secondary_model.is_mdx_c:
seperator = SeperateMDXC(secondary_model, process_data, main_model_primary_stem_4_stem=main_model_primary_stem_4_stem, main_process_method=main_process_method, is_return_dual=is_return_dual, main_model_primary=main_model_primary)
else:
seperator = SeperateMDX(secondary_model, process_data, main_model_primary_stem_4_stem=main_model_primary_stem_4_stem, main_process_method=main_process_method, main_model_primary=main_model_primary)
if secondary_model.process_method == DEMUCS_ARCH_TYPE:
seperator = SeperateDemucs(secondary_model, process_data, main_model_primary_stem_4_stem=main_model_primary_stem_4_stem, main_process_method=main_process_method, is_return_dual=is_return_dual, main_model_primary=main_model_primary)
secondary_sources = seperator.seperate()
if type(secondary_sources) is dict and not is_source_load and not is_pre_proc_model:
return gather_sources(secondary_model.primary_model_primary_stem, secondary_stem(secondary_model.primary_model_primary_stem), secondary_sources)
else:
return secondary_sources
def process_chain_model(secondary_model: ModelData,
process_data,
vocal_stem_path,
master_vocal_source,
master_inst_source=None):
process_iteration = process_data['process_iteration']
process_iteration()
if secondary_model.bv_model_rebalance:
vocal_source = spec_utils.reduce_mix_bv(master_inst_source, master_vocal_source, reduction_rate=secondary_model.bv_model_rebalance)
else:
vocal_source = master_vocal_source
vocal_stem_path = [vocal_source, os.path.splitext(os.path.basename(vocal_stem_path))[0]]
if secondary_model.process_method == VR_ARCH_TYPE:
seperator = SeperateVR(secondary_model, process_data, vocal_stem_path=vocal_stem_path, master_inst_source=master_inst_source, master_vocal_source=master_vocal_source)
if secondary_model.process_method == MDX_ARCH_TYPE:
if secondary_model.is_mdx_c:
seperator = SeperateMDXC(secondary_model, process_data, vocal_stem_path=vocal_stem_path, master_inst_source=master_inst_source, master_vocal_source=master_vocal_source)
else:
seperator = SeperateMDX(secondary_model, process_data, vocal_stem_path=vocal_stem_path, master_inst_source=master_inst_source, master_vocal_source=master_vocal_source)
if secondary_model.process_method == DEMUCS_ARCH_TYPE:
seperator = SeperateDemucs(secondary_model, process_data, vocal_stem_path=vocal_stem_path, master_inst_source=master_inst_source, master_vocal_source=master_vocal_source)
secondary_sources = seperator.seperate()
if type(secondary_sources) is dict:
return secondary_sources
else:
return None
def gather_sources(primary_stem_name, secondary_stem_name, secondary_sources: dict):
source_primary = False
source_secondary = False
for key, value in secondary_sources.items():
if key in primary_stem_name:
source_primary = value
if key in secondary_stem_name:
source_secondary = value
return source_primary, source_secondary
def prepare_mix(mix):
audio_path = mix
if not isinstance(mix, np.ndarray):
mix, sr = librosa.load(mix, mono=False, sr=44100)
else:
mix = mix.T
if isinstance(audio_path, str):
if not np.any(mix) and audio_path.endswith('.mp3'):
mix = rerun_mp3(audio_path)
if mix.ndim == 1:
mix = np.asfortranarray([mix,mix])
return mix
def rerun_mp3(audio_file, sample_rate=44100):
with audioread.audio_open(audio_file) as f:
track_length = int(f.duration)
return librosa.load(audio_file, duration=track_length, mono=False, sr=sample_rate)[0]
def save_format(audio_path, save_format, mp3_bit_set):
if not save_format == WAV:
if OPERATING_SYSTEM == 'Darwin':
FFMPEG_PATH = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'ffmpeg')
pydub.AudioSegment.converter = FFMPEG_PATH
musfile = pydub.AudioSegment.from_wav(audio_path)
if save_format == FLAC:
audio_path_flac = audio_path.replace(".wav", ".flac")
musfile.export(audio_path_flac, format="flac")
if save_format == MP3:
audio_path_mp3 = audio_path.replace(".wav", ".mp3")
try:
musfile.export(audio_path_mp3, format="mp3", bitrate=mp3_bit_set, codec="libmp3lame")
except Exception as e:
print(e)
musfile.export(audio_path_mp3, format="mp3", bitrate=mp3_bit_set)
try:
os.remove(audio_path)
except Exception as e:
print(e)
def pitch_shift(mix):
new_sr = 31183
# Resample audio file
resampled_audio = signal.resample_poly(mix, new_sr, 44100)
return resampled_audio
def list_to_dictionary(lst):
dictionary = {item: index for index, item in enumerate(lst)}
return dictionary
def vr_denoiser(X, device, hop_length=1024, n_fft=2048, cropsize=256, is_deverber=False, model_path=None):
batchsize = 4
if is_deverber:
nout, nout_lstm = 64, 128
mp = ModelParameters(os.path.join('lib_v5', 'vr_network', 'modelparams', '4band_v3.json'))
n_fft = mp.param['bins'] * 2
else:
mp = None
hop_length=1024
nout, nout_lstm = 16, 128
model = nets_new.CascadedNet(n_fft, nout=nout, nout_lstm=nout_lstm)
model.load_state_dict(torch.load(model_path, map_location=cpu))
model.to(device)
if mp is None:
X_spec = spec_utils.wave_to_spectrogram_old(X, hop_length, n_fft)
else:
X_spec = loading_mix(X.T, mp)
#PreProcess
X_mag = np.abs(X_spec)
X_phase = np.angle(X_spec)
#Sep
n_frame = X_mag.shape[2]
pad_l, pad_r, roi_size = spec_utils.make_padding(n_frame, cropsize, model.offset)
X_mag_pad = np.pad(X_mag, ((0, 0), (0, 0), (pad_l, pad_r)), mode='constant')
X_mag_pad /= X_mag_pad.max()
X_dataset = []
patches = (X_mag_pad.shape[2] - 2 * model.offset) // roi_size
for i in range(patches):
start = i * roi_size
X_mag_crop = X_mag_pad[:, :, start:start + cropsize]
X_dataset.append(X_mag_crop)
X_dataset = np.asarray(X_dataset)
model.eval()
with torch.no_grad():
mask = []
# To reduce the overhead, dataloader is not used.
for i in range(0, patches, batchsize):
X_batch = X_dataset[i: i + batchsize]
X_batch = torch.from_numpy(X_batch).to(device)
pred = model.predict_mask(X_batch)
pred = pred.detach().cpu().numpy()
pred = np.concatenate(pred, axis=2)
mask.append(pred)
mask = np.concatenate(mask, axis=2)
mask = mask[:, :, :n_frame]
#Post Proc
if is_deverber:
v_spec = mask * X_mag * np.exp(1.j * X_phase)
y_spec = (1 - mask) * X_mag * np.exp(1.j * X_phase)
else:
v_spec = (1 - mask) * X_mag * np.exp(1.j * X_phase)
if mp is None:
wave = spec_utils.spectrogram_to_wave_old(v_spec, hop_length=1024)
else:
wave = spec_utils.cmb_spectrogram_to_wave(v_spec, mp, is_v51_model=True).T
wave = spec_utils.match_array_shapes(wave, X)
if is_deverber:
wave_2 = spec_utils.cmb_spectrogram_to_wave(y_spec, mp, is_v51_model=True).T
wave_2 = spec_utils.match_array_shapes(wave_2, X)
return wave, wave_2
else:
return wave
def loading_mix(X, mp):
X_wave, X_spec_s = {}, {}
bands_n = len(mp.param['band'])
for d in range(bands_n, 0, -1):
bp = mp.param['band'][d]
if OPERATING_SYSTEM == 'Darwin':
wav_resolution = 'polyphase' if SYSTEM_PROC == ARM or ARM in SYSTEM_ARCH else bp['res_type']
else:
wav_resolution = 'polyphase'#bp['res_type']
if d == bands_n: # high-end band
X_wave[d] = X
else: # lower bands
X_wave[d] = librosa.resample(X_wave[d+1], mp.param['band'][d+1]['sr'], bp['sr'], res_type=wav_resolution)
X_spec_s[d] = spec_utils.wave_to_spectrogram(X_wave[d], bp['hl'], bp['n_fft'], mp, band=d, is_v51_model=True)
# if d == bands_n and is_high_end_process:
# input_high_end_h = (bp['n_fft']//2 - bp['crop_stop']) + (mp.param['pre_filter_stop'] - mp.param['pre_filter_start'])
# input_high_end = X_spec_s[d][:, bp['n_fft']//2-input_high_end_h:bp['n_fft']//2, :]
X_spec = spec_utils.combine_spectrograms(X_spec_s, mp)
del X_wave, X_spec_s
return X_spec