Duplicate from ImPavloh/RVC-TTS-Demo
Browse filesCo-authored-by: Pablo <ImPavloh@users.noreply.huggingface.co>
This view is limited to 50 files because it contains too many changes.
See raw diff
- .gitattributes +48 -0
- README.md +14 -0
- app.py +200 -0
- config.py +31 -0
- hubert_base.pt +3 -0
- lib/infer_pack/__pycache__/attentions.cpython-310.pyc +0 -0
- lib/infer_pack/__pycache__/commons.cpython-310.pyc +0 -0
- lib/infer_pack/__pycache__/models.cpython-310.pyc +0 -0
- lib/infer_pack/__pycache__/modules.cpython-310.pyc +0 -0
- lib/infer_pack/__pycache__/transforms.cpython-310.pyc +0 -0
- lib/infer_pack/attentions.py +253 -0
- lib/infer_pack/commons.py +112 -0
- lib/infer_pack/models.py +711 -0
- lib/infer_pack/models_onnx.py +582 -0
- lib/infer_pack/modules.py +315 -0
- lib/infer_pack/onnx_inference.py +67 -0
- lib/infer_pack/transforms.py +115 -0
- models/FernandoAlonso/FernandoAlonso/added_IVF582_Flat_nprobe_1_fernando2_v2.index +3 -0
- models/FernandoAlonso/FernandoAlonso/fernando2.pth +3 -0
- models/HolasoyGerman/HolasoyGerman/HolasoyGerman.pth +3 -0
- models/HolasoyGerman/HolasoyGerman/added_IVF3117_Flat_nprobe_1_HolasoyGerman_v2.index +3 -0
- models/HolasoyGerman/model_info.json +9 -0
- models/Homer/Homer/HomerEsp.pth +3 -0
- models/Homer/Homer/added_IVF360_Flat_nprobe_1_HomerEsp_v1.index +3 -0
- models/Homer/model_info.json +9 -0
- models/Ibai/Ibai/Ibai.pth +3 -0
- models/Ibai/Ibai/added_IVF4601_Flat_nprobe_1_Ibai_v2.index +3 -0
- models/Ibai/model_info.json +9 -0
- models/IlloJuan/IlloJuan/IlloJuan.pth +3 -0
- models/IlloJuan/IlloJuan/added_IVF593_Flat_nprobe_1_IlloJuan_v2.index +3 -0
- models/IlloJuan/model_info.json +9 -0
- models/Quevedo/Quevedo/added_IVF2301_Flat_nprobe_10.index +3 -0
- models/Quevedo/Quevedo/quevedo.pth +3 -0
- models/Quevedo/model_info.json +9 -0
- models/Shadoune666/Shadoune666/Shadoune666.pth +3 -0
- models/Shadoune666/Shadoune666/added_IVF716_Flat_nprobe_1_Shadoune666_v2.index +3 -0
- models/Shadoune666/model_info.json +9 -0
- models/Spreen/Spreen/Spreen.pth +3 -0
- models/Spreen/Spreen/added_IVF4737_Flat_nprobe_1_Spreen_v2.index +3 -0
- models/Spreen/model_info.json +9 -0
- models/Totakeke/Totakeke/added_IVF256_Flat_nprobe_1_full_totakeke_v2.index +3 -0
- models/Totakeke/Totakeke/full_totakeke_e150_s44400.pth +3 -0
- models/Totakeke/model_info.json +9 -0
- models/Vegetta777/Vegetta777/Vegetta777.pth +3 -0
- models/Vegetta777/Vegetta777/added_IVF2021_Flat_nprobe_1_Vegetta777_v2.index +3 -0
- models/Vegetta777/model_info.json +9 -0
- models/Villager/Villager/added_IVF81_Flat_nprobe_1_v2.index +3 -0
- models/Villager/Villager/villager.pth +3 -0
- models/Villager/model_info.json +9 -0
- models/WalterWhite/WalterWhite/added_IVF458_Flat_nprobe_1_ww2_v2.index +3 -0
.gitattributes
ADDED
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
*.7z filter=lfs diff=lfs merge=lfs -text
|
2 |
+
*.arrow filter=lfs diff=lfs merge=lfs -text
|
3 |
+
*.bin filter=lfs diff=lfs merge=lfs -text
|
4 |
+
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
5 |
+
*.ckpt filter=lfs diff=lfs merge=lfs -text
|
6 |
+
*.ftz filter=lfs diff=lfs merge=lfs -text
|
7 |
+
*.gz filter=lfs diff=lfs merge=lfs -text
|
8 |
+
*.h5 filter=lfs diff=lfs merge=lfs -text
|
9 |
+
*.joblib filter=lfs diff=lfs merge=lfs -text
|
10 |
+
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
11 |
+
*.mlmodel filter=lfs diff=lfs merge=lfs -text
|
12 |
+
*.model filter=lfs diff=lfs merge=lfs -text
|
13 |
+
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
14 |
+
*.npy filter=lfs diff=lfs merge=lfs -text
|
15 |
+
*.npz filter=lfs diff=lfs merge=lfs -text
|
16 |
+
*.onnx filter=lfs diff=lfs merge=lfs -text
|
17 |
+
*.ot filter=lfs diff=lfs merge=lfs -text
|
18 |
+
*.parquet filter=lfs diff=lfs merge=lfs -text
|
19 |
+
*.pb filter=lfs diff=lfs merge=lfs -text
|
20 |
+
*.pickle filter=lfs diff=lfs merge=lfs -text
|
21 |
+
*.pkl filter=lfs diff=lfs merge=lfs -text
|
22 |
+
*.pt filter=lfs diff=lfs merge=lfs -text
|
23 |
+
*.pth filter=lfs diff=lfs merge=lfs -text
|
24 |
+
*.rar filter=lfs diff=lfs merge=lfs -text
|
25 |
+
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
26 |
+
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
27 |
+
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
28 |
+
*.tar filter=lfs diff=lfs merge=lfs -text
|
29 |
+
*.tflite filter=lfs diff=lfs merge=lfs -text
|
30 |
+
*.tgz filter=lfs diff=lfs merge=lfs -text
|
31 |
+
*.wasm filter=lfs diff=lfs merge=lfs -text
|
32 |
+
*.xz filter=lfs diff=lfs merge=lfs -text
|
33 |
+
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
+
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
+
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
models/HolasoyGerman/HolasoyGerman/added_IVF3117_Flat_nprobe_1_HolasoyGerman_v2.index filter=lfs diff=lfs merge=lfs -text
|
37 |
+
models/Homer/Homer/added_IVF360_Flat_nprobe_1_HomerEsp_v1.index filter=lfs diff=lfs merge=lfs -text
|
38 |
+
models/Ibai/Ibai/added_IVF4601_Flat_nprobe_1_Ibai_v2.index filter=lfs diff=lfs merge=lfs -text
|
39 |
+
models/IlloJuan/IlloJuan/added_IVF593_Flat_nprobe_1_IlloJuan_v2.index filter=lfs diff=lfs merge=lfs -text
|
40 |
+
models/Joseju/Joseju/added_IVF256_Flat_nprobe_1_Joseju_v2.index filter=lfs diff=lfs merge=lfs -text
|
41 |
+
models/Quevedo/Quevedo/added_IVF2301_Flat_nprobe_10.index filter=lfs diff=lfs merge=lfs -text
|
42 |
+
models/Shadoune666/Shadoune666/added_IVF716_Flat_nprobe_1_Shadoune666_v2.index filter=lfs diff=lfs merge=lfs -text
|
43 |
+
models/Spreen/Spreen/added_IVF4737_Flat_nprobe_1_Spreen_v2.index filter=lfs diff=lfs merge=lfs -text
|
44 |
+
models/Totakeke/Totakeke/added_IVF256_Flat_nprobe_1_full_totakeke_v2.index filter=lfs diff=lfs merge=lfs -text
|
45 |
+
models/Vegetta777/Vegetta777/added_IVF2021_Flat_nprobe_1_Vegetta777_v2.index filter=lfs diff=lfs merge=lfs -text
|
46 |
+
models/Villager/Villager/added_IVF81_Flat_nprobe_1_v2.index filter=lfs diff=lfs merge=lfs -text
|
47 |
+
models/WalterWhite/WalterWhite/added_IVF458_Flat_nprobe_1_ww2_v2.index filter=lfs diff=lfs merge=lfs -text
|
48 |
+
models/FernandoAlonso/FernandoAlonso/added_IVF582_Flat_nprobe_1_fernando2_v2.index filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
title: RVC TTS Demo
|
3 |
+
emoji: 🚀
|
4 |
+
colorFrom: red
|
5 |
+
colorTo: pink
|
6 |
+
sdk: gradio
|
7 |
+
sdk_version: 3.36.1
|
8 |
+
app_file: app.py
|
9 |
+
pinned: false
|
10 |
+
license: gpl-3.0
|
11 |
+
duplicated_from: ImPavloh/RVC-TTS-Demo
|
12 |
+
---
|
13 |
+
|
14 |
+
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
app.py
ADDED
@@ -0,0 +1,200 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import json
|
3 |
+
import torch
|
4 |
+
import asyncio
|
5 |
+
import librosa
|
6 |
+
import hashlib
|
7 |
+
import edge_tts
|
8 |
+
import gradio as gr
|
9 |
+
from config import Config
|
10 |
+
from vc_infer_pipeline import VC
|
11 |
+
from fairseq import checkpoint_utils
|
12 |
+
from lib.infer_pack.models import (SynthesizerTrnMs256NSFsid, SynthesizerTrnMs256NSFsid_nono, SynthesizerTrnMs768NSFsid, SynthesizerTrnMs768NSFsid_nono,)
|
13 |
+
|
14 |
+
config = Config()
|
15 |
+
|
16 |
+
def load_json_file(filepath):
|
17 |
+
with open(filepath, "r", encoding="utf-8") as f: content = json.load(f)
|
18 |
+
return content
|
19 |
+
|
20 |
+
def file_checksum(file_path):
|
21 |
+
with open(file_path, 'rb') as f:
|
22 |
+
file_data = f.read()
|
23 |
+
return hashlib.md5(file_data).hexdigest()
|
24 |
+
|
25 |
+
def get_existing_model_info(category_directory):
|
26 |
+
model_info_path = os.path.join(category_directory, 'model_info.json')
|
27 |
+
if os.path.exists(model_info_path):
|
28 |
+
with open(model_info_path, 'r') as f: return json.load(f)
|
29 |
+
return None
|
30 |
+
|
31 |
+
def generate_model_info_files():
|
32 |
+
folder_info = {}
|
33 |
+
model_directory = "models/"
|
34 |
+
for category_name in os.listdir(model_directory):
|
35 |
+
category_directory = os.path.join(model_directory, category_name)
|
36 |
+
if not os.path.isdir(category_directory): continue
|
37 |
+
|
38 |
+
folder_info[category_name] = {"title": category_name, "folder_path": category_name}
|
39 |
+
existing_model_info = get_existing_model_info(category_directory)
|
40 |
+
model_info = {}
|
41 |
+
regenerate_model_info = False
|
42 |
+
|
43 |
+
for model_name in os.listdir(category_directory):
|
44 |
+
model_path = os.path.join(category_directory, model_name)
|
45 |
+
if not os.path.isdir(model_path): continue
|
46 |
+
|
47 |
+
model_data, regenerate = gather_model_info(category_directory, model_name, model_path, existing_model_info)
|
48 |
+
if model_data is not None:
|
49 |
+
model_info[model_name] = model_data
|
50 |
+
regenerate_model_info |= regenerate
|
51 |
+
|
52 |
+
if regenerate_model_info:
|
53 |
+
with open(os.path.join(category_directory, 'model_info.json'), 'w') as f: json.dump(model_info, f, indent=4)
|
54 |
+
|
55 |
+
folder_info_path = os.path.join(model_directory, 'folder_info.json')
|
56 |
+
with open(folder_info_path, 'w') as f: json.dump(folder_info, f, indent=4)
|
57 |
+
|
58 |
+
def should_regenerate_model_info(existing_model_info, model_name, pth_checksum, index_checksum):
|
59 |
+
if existing_model_info is None or model_name not in existing_model_info: return True
|
60 |
+
return (existing_model_info[model_name]['model_path_checksum'] != pth_checksum or existing_model_info[model_name]['index_path_checksum'] != index_checksum)
|
61 |
+
|
62 |
+
def get_model_files(model_path): return [f for f in os.listdir(model_path) if f.endswith('.pth') or f.endswith('.index')]
|
63 |
+
|
64 |
+
def gather_model_info(category_directory, model_name, model_path, existing_model_info):
|
65 |
+
model_files = get_model_files(model_path)
|
66 |
+
if len(model_files) != 2: return None, False
|
67 |
+
|
68 |
+
pth_file = [f for f in model_files if f.endswith('.pth')][0]
|
69 |
+
index_file = [f for f in model_files if f.endswith('.index')][0]
|
70 |
+
pth_checksum = file_checksum(os.path.join(model_path, pth_file))
|
71 |
+
index_checksum = file_checksum(os.path.join(model_path, index_file))
|
72 |
+
regenerate = should_regenerate_model_info(existing_model_info, model_name, pth_checksum, index_checksum)
|
73 |
+
|
74 |
+
return {"title": model_name, "model_path": pth_file, "feature_retrieval_library": index_file, "model_path_checksum": pth_checksum, "index_path_checksum": index_checksum}, regenerate
|
75 |
+
|
76 |
+
def create_vc_fn(model_name, tgt_sr, net_g, vc, if_f0, version, file_index):
|
77 |
+
def vc_fn(tts_text, tts_voice):
|
78 |
+
try:
|
79 |
+
if len(tts_text) > 100: return None
|
80 |
+
if tts_text is None or tts_voice is None: return None
|
81 |
+
asyncio.run(edge_tts.Communicate(tts_text, "-".join(tts_voice.split('-')[:-1])).save("tts.mp3"))
|
82 |
+
audio, sr = librosa.load("tts.mp3", sr=16000, mono=True)
|
83 |
+
vc_input = "tts.mp3"
|
84 |
+
times = [0, 0, 0]
|
85 |
+
audio_opt = vc.pipeline(hubert_model, net_g, 0, audio, vc_input, times, 0, "pm", file_index, 0.7, if_f0, 3, tgt_sr, 0, 1, version, 0.5, f0_file=None)
|
86 |
+
return (tgt_sr, audio_opt)
|
87 |
+
except Exception: return None
|
88 |
+
return vc_fn
|
89 |
+
|
90 |
+
def load_model_parameters(category_folder, character_name, info):
|
91 |
+
model_index = f"models/{category_folder}/{character_name}/{info['feature_retrieval_library']}"
|
92 |
+
cpt = torch.load(f"models/{category_folder}/{character_name}/{info['model_path']}", map_location="cpu")
|
93 |
+
return model_index, cpt
|
94 |
+
|
95 |
+
def select_net_g(cpt, version, if_f0):
|
96 |
+
if version == "v1":
|
97 |
+
if if_f0 == 1: net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=config.is_half)
|
98 |
+
else: net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
|
99 |
+
elif version == "v2":
|
100 |
+
if if_f0 == 1: net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=config.is_half)
|
101 |
+
else: net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
|
102 |
+
return net_g
|
103 |
+
|
104 |
+
def load_model_and_prepare(cpt, net_g):
|
105 |
+
del net_g.enc_q
|
106 |
+
net_g.load_state_dict(cpt["weight"], strict=False)
|
107 |
+
net_g.eval().to(config.device)
|
108 |
+
net_g = net_g.half() if config.is_half else net_g.float()
|
109 |
+
return net_g
|
110 |
+
|
111 |
+
def create_and_append_model(models, model_functions, character_name, model_title, version, vc_fn):
|
112 |
+
models.append((character_name, model_title, version, vc_fn))
|
113 |
+
model_functions[character_name] = vc_fn
|
114 |
+
return models, model_functions
|
115 |
+
|
116 |
+
def load_model():
|
117 |
+
categories = []
|
118 |
+
model_functions = {}
|
119 |
+
folder_info = load_json_file("models/folder_info.json")
|
120 |
+
for category_name, category_info in folder_info.items():
|
121 |
+
models = []
|
122 |
+
models_info = load_json_file(f"models/{category_info['folder_path']}/model_info.json")
|
123 |
+
for character_name, info in models_info.items():
|
124 |
+
model_index, cpt = load_model_parameters(category_info['folder_path'], character_name, info)
|
125 |
+
net_g = select_net_g(cpt, cpt.get("version", "v1"), cpt.get("f0", 1))
|
126 |
+
cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
|
127 |
+
net_g = load_model_and_prepare(cpt, net_g)
|
128 |
+
vc = VC(cpt["config"][-1], config)
|
129 |
+
vc_fn = create_vc_fn(info['model_path'], cpt["config"][-1], net_g, vc, cpt.get("f0", 1), cpt.get("version", "v1"), model_index)
|
130 |
+
models, model_functions = create_and_append_model(models, model_functions, character_name, info['title'], cpt.get("version", "v1"), vc_fn)
|
131 |
+
categories.append([category_info['title'], category_info['folder_path'], models])
|
132 |
+
return categories, model_functions
|
133 |
+
|
134 |
+
generate_model_info_files()
|
135 |
+
|
136 |
+
css = """
|
137 |
+
.gradio-container { font-family: 'IBM Plex Sans', sans-serif; }
|
138 |
+
footer { visibility: hidden; display: none; }
|
139 |
+
.center-container { display: flex; flex-direction: column; align-items: center; justify-content: center;}
|
140 |
+
"""
|
141 |
+
|
142 |
+
if __name__ == '__main__':
|
143 |
+
global hubert_model
|
144 |
+
models, _, _ = checkpoint_utils.load_model_ensemble_and_task(["hubert_base.pt"], suffix="")
|
145 |
+
hubert_model = models[0]
|
146 |
+
hubert_model = hubert_model.to(config.device)
|
147 |
+
hubert_model = hubert_model.half() if config.is_half else hubert_model.float()
|
148 |
+
hubert_model.eval()
|
149 |
+
categories, model_functions = load_model()
|
150 |
+
tts_voice_list = asyncio.get_event_loop().run_until_complete(edge_tts.list_voices())
|
151 |
+
voices = [f"{v['ShortName']}-{v['Gender']}" for v in tts_voice_list]
|
152 |
+
with gr.Blocks(css=css, title="Demo RVC TTS - Pavloh", theme=gr.themes.Soft(primary_hue="cyan", secondary_hue="blue", radius_size="lg", text_size="lg")
|
153 |
+
.set(loader_color="#0B0F19", shadow_drop='*shadow_drop_lg', block_border_width="3px")) as pavloh:
|
154 |
+
gr.HTML("""
|
155 |
+
<div class="center-container">
|
156 |
+
<div style="display: flex; justify-content: center;">
|
157 |
+
<a href="https://github.com/ImPavloh/rvc-tts/blob/main/LICENSE" target="_blank">
|
158 |
+
<img src="https://img.shields.io/github/license/impavloh/voiceit?style=for-the-badge&logo=github&logoColor=white" alt="License">
|
159 |
+
</a>
|
160 |
+
<a href="https://github.com/ImPavloh/rvc-tts" target="_blank">
|
161 |
+
<img src="https://img.shields.io/badge/repository-%23121011.svg?style=for-the-badge&logo=github&logoColor=white" alt="GitHub">
|
162 |
+
</a>
|
163 |
+
<form action="https://www.paypal.com/donate" method="post" target="_blank">
|
164 |
+
<input type="hidden" name="hosted_button_id" value="6FPWP9AWEKSWJ" />
|
165 |
+
<input type="image" src="https://img.shields.io/badge/support-%2300457C.svg?style=for-the-badge&logo=paypal&logoColor=white" border="0" name="submit" alt="Donate with PayPal" />
|
166 |
+
<img alt="" border="0" src="https://www.paypal.com/es_ES/i/scr/pixel.gif" width="1" height="1" />
|
167 |
+
</form>
|
168 |
+
<a href="https://twitter.com/impavloh" target="_blank">
|
169 |
+
<img src="https://img.shields.io/badge/follow-%231DA1F2.svg?style=for-the-badge&logo=twitter&logoColor=white" alt="Twitter">
|
170 |
+
</a>
|
171 |
+
</div>
|
172 |
+
<div style="display: inline-flex; align-items: center; gap: 0.8rem; font-size: 1.75rem;">
|
173 |
+
<h1 style="font-weight: 900; margin-bottom: 7px; margin-top: 5px">🗣️ RVC TTS Demo - <a style="text-decoration: underline;" href="https://twitter.com/impavloh">Pavloh</a></h1>
|
174 |
+
</div>
|
175 |
+
<p style="margin-bottom: 10px; font-size: 94%; line-height: 23px;">An AI-Powered Text-to-Speech</p>
|
176 |
+
<p><b>Try out the <a style="text-decoration: underline;" href="https://github.com/ImPavloh/rvc-tts-discord-bot">RVC Text-to-Speech Discord Bot</a></b></p>
|
177 |
+
</div>
|
178 |
+
""")
|
179 |
+
|
180 |
+
with gr.Row():
|
181 |
+
with gr.Column():
|
182 |
+
m1 = gr.Dropdown(label="📦 Voice Model", choices=list(model_functions.keys()), allow_custom_value=False, value="Ibai")
|
183 |
+
t2 = gr.Dropdown(label="⚙️ Voice style and language [Edge-TTS]", choices=voices, allow_custom_value=False, value="es-ES-AlvaroNeural-Male")
|
184 |
+
t1 = gr.Textbox(label="📝 Text to convert")
|
185 |
+
c1 = gr.Button("Convert", variant="primary")
|
186 |
+
a1 = gr.Audio(label="🔉 Converted Text", interactive=False)
|
187 |
+
|
188 |
+
def call_selected_model_fn(selected_model, t1, t2):
|
189 |
+
vc_fn = model_functions[selected_model]
|
190 |
+
return vc_fn(t1, t2)
|
191 |
+
|
192 |
+
c1.click(fn=call_selected_model_fn, inputs=[m1, t1, t2], outputs=[a1])
|
193 |
+
|
194 |
+
gr.HTML("""
|
195 |
+
<center>
|
196 |
+
<p><i> By using this website, you agree to the <a style="text-decoration: underline;" href="https://github.com/ImPavloh/rvc-tts/blob/main/LICENSE">license</a>. </i></p>
|
197 |
+
</center>
|
198 |
+
""")
|
199 |
+
|
200 |
+
pavloh.queue(concurrency_count=1).launch()
|
config.py
ADDED
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import argparse
|
2 |
+
from multiprocessing import cpu_count
|
3 |
+
|
4 |
+
class Config:
|
5 |
+
def __init__(self):
|
6 |
+
self.device = "cpu"
|
7 |
+
self.is_half = False
|
8 |
+
self.n_cpu = cpu_count()
|
9 |
+
(self.python_cmd, self.colab, self.noparallel, self.noautoopen, self.api) = self.arg_parse()
|
10 |
+
self.listen_port = 7860
|
11 |
+
self.x_pad, self.x_query, self.x_center, self.x_max = self.device_config()
|
12 |
+
|
13 |
+
@staticmethod
|
14 |
+
def arg_parse() -> tuple:
|
15 |
+
parser = argparse.ArgumentParser()
|
16 |
+
parser.add_argument("--pycmd", type=str, default="python")
|
17 |
+
parser.add_argument("--colab", action="store_true")
|
18 |
+
parser.add_argument("--noparallel", action="store_true")
|
19 |
+
parser.add_argument("--noautoopen", action="store_true")
|
20 |
+
parser.add_argument("--api", action="store_true")
|
21 |
+
cmd_opts = parser.parse_args()
|
22 |
+
|
23 |
+
return (cmd_opts.pycmd, cmd_opts.colab, cmd_opts.noparallel, cmd_opts.noautoopen, cmd_opts.api)
|
24 |
+
|
25 |
+
def device_config(self) -> tuple:
|
26 |
+
x_pad = 1
|
27 |
+
x_query = 6
|
28 |
+
x_center = 38
|
29 |
+
x_max = 41
|
30 |
+
|
31 |
+
return x_pad, x_query, x_center, x_max
|
hubert_base.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f54b40fd2802423a5643779c4861af1e9ee9c1564dc9d32f54f20b5ffba7db96
|
3 |
+
size 189507909
|
lib/infer_pack/__pycache__/attentions.cpython-310.pyc
ADDED
Binary file (9.24 kB). View file
|
|
lib/infer_pack/__pycache__/commons.cpython-310.pyc
ADDED
Binary file (5.08 kB). View file
|
|
lib/infer_pack/__pycache__/models.cpython-310.pyc
ADDED
Binary file (19.5 kB). View file
|
|
lib/infer_pack/__pycache__/modules.cpython-310.pyc
ADDED
Binary file (11.4 kB). View file
|
|
lib/infer_pack/__pycache__/transforms.cpython-310.pyc
ADDED
Binary file (4.35 kB). View file
|
|
lib/infer_pack/attentions.py
ADDED
@@ -0,0 +1,253 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import copy
|
2 |
+
import math
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
from torch import nn
|
6 |
+
from torch.nn import functional as F
|
7 |
+
from lib.infer_pack import commons
|
8 |
+
from lib.infer_pack import modules
|
9 |
+
from lib.infer_pack.modules import LayerNorm
|
10 |
+
|
11 |
+
class Encoder(nn.Module):
|
12 |
+
def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0.0, window_size=10, **kwargs):
|
13 |
+
super().__init__()
|
14 |
+
self.hidden_channels = hidden_channels
|
15 |
+
self.filter_channels = filter_channels
|
16 |
+
self.n_heads = n_heads
|
17 |
+
self.n_layers = n_layers
|
18 |
+
self.kernel_size = kernel_size
|
19 |
+
self.p_dropout = p_dropout
|
20 |
+
self.window_size = window_size
|
21 |
+
self.drop = nn.Dropout(p_dropout)
|
22 |
+
self.attn_layers = nn.ModuleList()
|
23 |
+
self.norm_layers_1 = nn.ModuleList()
|
24 |
+
self.ffn_layers = nn.ModuleList()
|
25 |
+
self.norm_layers_2 = nn.ModuleList()
|
26 |
+
for _ in range(self.n_layers):
|
27 |
+
self.attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, window_size=window_size))
|
28 |
+
self.norm_layers_1.append(LayerNorm(hidden_channels))
|
29 |
+
self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout))
|
30 |
+
self.norm_layers_2.append(LayerNorm(hidden_channels))
|
31 |
+
|
32 |
+
def forward(self, x, x_mask):
|
33 |
+
attn_mask = x_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
|
34 |
+
x = x * x_mask
|
35 |
+
for i in range(self.n_layers):
|
36 |
+
y = self.attn_layers[i](x, x, attn_mask)
|
37 |
+
y = self.drop(y)
|
38 |
+
x = self.norm_layers_1[i](x + y)
|
39 |
+
y = self.ffn_layers[i](x, x_mask)
|
40 |
+
y = self.drop(y)
|
41 |
+
x = self.norm_layers_2[i](x + y)
|
42 |
+
x = x * x_mask
|
43 |
+
return x
|
44 |
+
|
45 |
+
class Decoder(nn.Module):
|
46 |
+
def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0.0, proximal_bias=False, proximal_init=True, **kwargs):
|
47 |
+
super().__init__()
|
48 |
+
self.hidden_channels = hidden_channels
|
49 |
+
self.filter_channels = filter_channels
|
50 |
+
self.n_heads = n_heads
|
51 |
+
self.n_layers = n_layers
|
52 |
+
self.kernel_size = kernel_size
|
53 |
+
self.p_dropout = p_dropout
|
54 |
+
self.proximal_bias = proximal_bias
|
55 |
+
self.proximal_init = proximal_init
|
56 |
+
self.drop = nn.Dropout(p_dropout)
|
57 |
+
self.self_attn_layers = nn.ModuleList()
|
58 |
+
self.norm_layers_0 = nn.ModuleList()
|
59 |
+
self.encdec_attn_layers = nn.ModuleList()
|
60 |
+
self.norm_layers_1 = nn.ModuleList()
|
61 |
+
self.ffn_layers = nn.ModuleList()
|
62 |
+
self.norm_layers_2 = nn.ModuleList()
|
63 |
+
for _ in range(self.n_layers):
|
64 |
+
self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))
|
65 |
+
self.norm_layers_0.append(LayerNorm(hidden_channels))
|
66 |
+
self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))
|
67 |
+
self.norm_layers_1.append(LayerNorm(hidden_channels))
|
68 |
+
self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
|
69 |
+
self.norm_layers_2.append(LayerNorm(hidden_channels))
|
70 |
+
|
71 |
+
def forward(self, x, x_mask, h, h_mask):
|
72 |
+
self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)
|
73 |
+
encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
|
74 |
+
x = x * x_mask
|
75 |
+
for i in range(self.n_layers):
|
76 |
+
y = self.self_attn_layers[i](x, x, self_attn_mask)
|
77 |
+
y = self.drop(y)
|
78 |
+
x = self.norm_layers_0[i](x + y)
|
79 |
+
|
80 |
+
y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
|
81 |
+
y = self.drop(y)
|
82 |
+
x = self.norm_layers_1[i](x + y)
|
83 |
+
|
84 |
+
y = self.ffn_layers[i](x, x_mask)
|
85 |
+
y = self.drop(y)
|
86 |
+
x = self.norm_layers_2[i](x + y)
|
87 |
+
x = x * x_mask
|
88 |
+
return x
|
89 |
+
|
90 |
+
|
91 |
+
class MultiHeadAttention(nn.Module):
|
92 |
+
def __init__(self, channels, out_channels, n_heads, p_dropout=0.0, window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False):
|
93 |
+
super().__init__()
|
94 |
+
assert channels % n_heads == 0
|
95 |
+
|
96 |
+
self.channels = channels
|
97 |
+
self.out_channels = out_channels
|
98 |
+
self.n_heads = n_heads
|
99 |
+
self.p_dropout = p_dropout
|
100 |
+
self.window_size = window_size
|
101 |
+
self.heads_share = heads_share
|
102 |
+
self.block_length = block_length
|
103 |
+
self.proximal_bias = proximal_bias
|
104 |
+
self.proximal_init = proximal_init
|
105 |
+
self.attn = None
|
106 |
+
|
107 |
+
self.k_channels = channels // n_heads
|
108 |
+
self.conv_q = nn.Conv1d(channels, channels, 1)
|
109 |
+
self.conv_k = nn.Conv1d(channels, channels, 1)
|
110 |
+
self.conv_v = nn.Conv1d(channels, channels, 1)
|
111 |
+
self.conv_o = nn.Conv1d(channels, out_channels, 1)
|
112 |
+
self.drop = nn.Dropout(p_dropout)
|
113 |
+
|
114 |
+
if window_size is not None:
|
115 |
+
n_heads_rel = 1 if heads_share else n_heads
|
116 |
+
rel_stddev = self.k_channels**-0.5
|
117 |
+
self.emb_rel_k = nn.Parameter(
|
118 |
+
torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
|
119 |
+
* rel_stddev
|
120 |
+
)
|
121 |
+
self.emb_rel_v = nn.Parameter(
|
122 |
+
torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels)
|
123 |
+
* rel_stddev
|
124 |
+
)
|
125 |
+
|
126 |
+
nn.init.xavier_uniform_(self.conv_q.weight)
|
127 |
+
nn.init.xavier_uniform_(self.conv_k.weight)
|
128 |
+
nn.init.xavier_uniform_(self.conv_v.weight)
|
129 |
+
if proximal_init:
|
130 |
+
with torch.no_grad():
|
131 |
+
self.conv_k.weight.copy_(self.conv_q.weight)
|
132 |
+
self.conv_k.bias.copy_(self.conv_q.bias)
|
133 |
+
|
134 |
+
def forward(self, x, c, attn_mask=None):
|
135 |
+
q = self.conv_q(x)
|
136 |
+
k = self.conv_k(c)
|
137 |
+
v = self.conv_v(c)
|
138 |
+
|
139 |
+
x, self.attn = self.attention(q, k, v, mask=attn_mask)
|
140 |
+
|
141 |
+
x = self.conv_o(x)
|
142 |
+
return x
|
143 |
+
|
144 |
+
def attention(self, query, key, value, mask=None):
|
145 |
+
b, d, t_s, t_t = (*key.size(), query.size(2))
|
146 |
+
query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
|
147 |
+
key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
|
148 |
+
value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
|
149 |
+
|
150 |
+
scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
|
151 |
+
if self.window_size is not None:
|
152 |
+
assert (
|
153 |
+
t_s == t_t
|
154 |
+
), "Relative attention is only available for self-attention."
|
155 |
+
key_relative_embeddings = self._get_relative_embeddings(self.emb_rel_k, t_s)
|
156 |
+
rel_logits = self._matmul_with_relative_keys(query / math.sqrt(self.k_channels), key_relative_embeddings)
|
157 |
+
scores_local = self._relative_position_to_absolute_position(rel_logits)
|
158 |
+
scores = scores + scores_local
|
159 |
+
if self.proximal_bias:
|
160 |
+
assert t_s == t_t, "Proximal bias is only available for self-attention."
|
161 |
+
scores = scores + self._attention_bias_proximal(t_s).to(device=scores.device, dtype=scores.dtype)
|
162 |
+
if mask is not None:
|
163 |
+
scores = scores.masked_fill(mask == 0, -1e4)
|
164 |
+
if self.block_length is not None:
|
165 |
+
assert (t_s == t_t), "Local attention is only available for self-attention."
|
166 |
+
block_mask = (
|
167 |
+
torch.ones_like(scores)
|
168 |
+
.triu(-self.block_length)
|
169 |
+
.tril(self.block_length)
|
170 |
+
)
|
171 |
+
scores = scores.masked_fill(block_mask == 0, -1e4)
|
172 |
+
p_attn = F.softmax(scores, dim=-1)
|
173 |
+
p_attn = self.drop(p_attn)
|
174 |
+
output = torch.matmul(p_attn, value)
|
175 |
+
if self.window_size is not None:
|
176 |
+
relative_weights = self._absolute_position_to_relative_position(p_attn)
|
177 |
+
value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
|
178 |
+
output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
|
179 |
+
output = (output.transpose(2, 3).contiguous().view(b, d, t_t))
|
180 |
+
return output, p_attn
|
181 |
+
|
182 |
+
def _matmul_with_relative_values(self, x, y): return torch.matmul(x, y.unsqueeze(0))
|
183 |
+
def _matmul_with_relative_keys(self, x, y): return torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
|
184 |
+
|
185 |
+
def _get_relative_embeddings(self, relative_embeddings, length):
|
186 |
+
max_relative_position = 2 * self.window_size + 1
|
187 |
+
pad_length = max(length - (self.window_size + 1), 0)
|
188 |
+
slice_start_position = max((self.window_size + 1) - length, 0)
|
189 |
+
slice_end_position = slice_start_position + 2 * length - 1
|
190 |
+
if pad_length > 0:padded_relative_embeddings = F.pad(relative_embeddings, commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))
|
191 |
+
else:
|
192 |
+
padded_relative_embeddings = relative_embeddings
|
193 |
+
return padded_relative_embeddings[:, slice_start_position:slice_end_position]
|
194 |
+
|
195 |
+
def _relative_position_to_absolute_position(self, x):
|
196 |
+
batch, heads, length, _ = x.size()
|
197 |
+
x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, 1]]))
|
198 |
+
x_flat = x.view([batch, heads, length * 2 * length])
|
199 |
+
x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [0, length - 1]]))
|
200 |
+
|
201 |
+
return x_flat.view([batch, heads, length + 1, 2 * length - 1])[:, :, :length, length - 1 :]
|
202 |
+
|
203 |
+
def _absolute_position_to_relative_position(self, x):
|
204 |
+
batch, heads, length, _ = x.size()
|
205 |
+
x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length - 1]]))
|
206 |
+
x_flat = x.view([batch, heads, length**2 + length * (length - 1)])
|
207 |
+
x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
|
208 |
+
return x_flat.view([batch, heads, length, 2 * length])[:, :, :, 1:]
|
209 |
+
|
210 |
+
def _attention_bias_proximal(self, length):
|
211 |
+
r = torch.arange(length, dtype=torch.float32)
|
212 |
+
diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
|
213 |
+
return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
|
214 |
+
|
215 |
+
class FFN(nn.Module):
|
216 |
+
def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0.0, activation=None, causal=False):
|
217 |
+
super().__init__()
|
218 |
+
self.in_channels = in_channels
|
219 |
+
self.out_channels = out_channels
|
220 |
+
self.filter_channels = filter_channels
|
221 |
+
self.kernel_size = kernel_size
|
222 |
+
self.p_dropout = p_dropout
|
223 |
+
self.activation = activation
|
224 |
+
self.causal = causal
|
225 |
+
self.padding = self._causal_padding if causal else self._same_padding
|
226 |
+
self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
|
227 |
+
self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
|
228 |
+
self.drop = nn.Dropout(p_dropout)
|
229 |
+
|
230 |
+
def forward(self, x, x_mask):
|
231 |
+
x = self.conv_1(self.padding(x * x_mask))
|
232 |
+
if self.activation == "gelu": x = x * torch.sigmoid(1.702 * x)
|
233 |
+
else: x = torch.relu(x)
|
234 |
+
x = self.drop(x)
|
235 |
+
x = self.conv_2(self.padding(x * x_mask))
|
236 |
+
return x * x_mask
|
237 |
+
|
238 |
+
def _causal_padding(self, x):
|
239 |
+
if self.kernel_size == 1: return x
|
240 |
+
pad_l = self.kernel_size - 1
|
241 |
+
pad_r = 0
|
242 |
+
return self._extracted_from__same_padding_5(pad_l, pad_r, x)
|
243 |
+
|
244 |
+
def _same_padding(self, x):
|
245 |
+
if self.kernel_size == 1: return x
|
246 |
+
pad_l = (self.kernel_size - 1) // 2
|
247 |
+
pad_r = self.kernel_size // 2
|
248 |
+
return self._extracted_from__same_padding_5(pad_l, pad_r, x)
|
249 |
+
|
250 |
+
def _extracted_from__same_padding_5(self, pad_l, pad_r, x):
|
251 |
+
padding = [[0, 0], [0, 0], [pad_l, pad_r]]
|
252 |
+
x = F.pad(x, commons.convert_pad_shape(padding))
|
253 |
+
return x
|
lib/infer_pack/commons.py
ADDED
@@ -0,0 +1,112 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import numpy as np
|
3 |
+
import torch
|
4 |
+
from torch import nn
|
5 |
+
from torch.nn import functional as F
|
6 |
+
|
7 |
+
def init_weights(m, mean=0.0, std=0.01):
|
8 |
+
classname = m.__class__.__name__
|
9 |
+
if "Conv" in classname: m.weight.data.normal_(mean, std)
|
10 |
+
|
11 |
+
def get_padding(kernel_size, dilation=1):
|
12 |
+
return (kernel_size * dilation - dilation) // 2
|
13 |
+
|
14 |
+
def convert_pad_shape(pad_shape):
|
15 |
+
l = pad_shape[::-1]
|
16 |
+
return [item for sublist in l for item in sublist]
|
17 |
+
|
18 |
+
def kl_divergence(m_p, logs_p, m_q, logs_q):
|
19 |
+
kl = logs_q - logs_p - 0.5
|
20 |
+
kl += 0.5 * (torch.exp(2.0 * logs_p) + (m_p - m_q) ** 2) * torch.exp(-2.0 * logs_q)
|
21 |
+
return kl
|
22 |
+
|
23 |
+
def rand_gumbel(shape):
|
24 |
+
return -torch.log(-torch.log(torch.rand(shape) + 1e-5))
|
25 |
+
|
26 |
+
def rand_gumbel_like(x):
|
27 |
+
return rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
|
28 |
+
|
29 |
+
def slice_segments(x, ids_str, segment_size=4, slice_dim=2):
|
30 |
+
if slice_dim == 1: ret = torch.zeros_like(x[:, :segment_size])
|
31 |
+
else: ret = torch.zeros_like(x[:, :, :segment_size])
|
32 |
+
|
33 |
+
for i in range(x.size(0)):
|
34 |
+
idx_str = ids_str[i]
|
35 |
+
idx_end = idx_str + segment_size
|
36 |
+
ret[i] = x[i, ..., idx_str:idx_end]
|
37 |
+
return ret
|
38 |
+
|
39 |
+
def rand_slice_segments(x, x_lengths=None, segment_size=4):
|
40 |
+
b, d, t = x.size()
|
41 |
+
|
42 |
+
if x_lengths is None: x_lengths = t
|
43 |
+
|
44 |
+
ids_str_max = x_lengths - segment_size + 1
|
45 |
+
ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
|
46 |
+
ret = slice_segments(x, ids_str, segment_size)
|
47 |
+
return ret, ids_str
|
48 |
+
|
49 |
+
def get_timing_signal_1d(length, channels, min_timescale=1.0, max_timescale=1.0e4):
|
50 |
+
position = torch.arange(length, dtype=torch.float)
|
51 |
+
num_timescales = channels // 2
|
52 |
+
log_timescale_increment = math.log(float(max_timescale) / float(min_timescale)) / (num_timescales - 1)
|
53 |
+
inv_timescales = min_timescale * torch.exp(torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment)
|
54 |
+
scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
|
55 |
+
signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
|
56 |
+
signal = F.pad(signal, [0, 0, 0, channels % 2])
|
57 |
+
signal = signal.view(1, channels, length)
|
58 |
+
return signal
|
59 |
+
|
60 |
+
def apply_timing_signal_1d(x, operation='add', min_timescale=1.0, max_timescale=1.0e4):
|
61 |
+
b, channels, length = x.size()
|
62 |
+
signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
|
63 |
+
signal = signal.to(dtype=x.dtype, device=x.device)
|
64 |
+
|
65 |
+
if operation == 'add': return x + signal
|
66 |
+
elif operation == 'cat': return torch.cat([x, signal], axis=1)
|
67 |
+
|
68 |
+
def subsequent_mask(length):
|
69 |
+
return torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
|
70 |
+
|
71 |
+
@torch.jit.script
|
72 |
+
def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
|
73 |
+
n_channels_int = n_channels[0]
|
74 |
+
in_act = input_a + input_b
|
75 |
+
t_act = torch.tanh(in_act[:, :n_channels_int, :])
|
76 |
+
s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
|
77 |
+
return t_act * s_act
|
78 |
+
|
79 |
+
def shift_1d(x):
|
80 |
+
x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
|
81 |
+
return x
|
82 |
+
|
83 |
+
def sequence_mask(length, max_length=None):
|
84 |
+
if max_length is None: max_length = length.max()
|
85 |
+
x = torch.arange(max_length, dtype=length.dtype, device=length.device)
|
86 |
+
return x.unsqueeze(0) < length.unsqueeze(1)
|
87 |
+
|
88 |
+
def generate_path(duration, mask):
|
89 |
+
device = duration.device
|
90 |
+
b, _, t_y, t_x = mask.shape
|
91 |
+
cum_duration = torch.cumsum(duration, -1)
|
92 |
+
cum_duration_flat = cum_duration.view(b * t_x)
|
93 |
+
path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
|
94 |
+
path = path.view(b, t_x, t_y)
|
95 |
+
path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
|
96 |
+
path = path.unsqueeze(1).transpose(2, 3) * mask
|
97 |
+
return path
|
98 |
+
|
99 |
+
def clip_grad_value_(parameters, clip_value, norm_type=2):
|
100 |
+
if isinstance(parameters, torch.Tensor): parameters = [parameters]
|
101 |
+
parameters = list(filter(lambda p: p.grad is not None, parameters))
|
102 |
+
norm_type = float(norm_type)
|
103 |
+
|
104 |
+
if clip_value is not None: clip_value = float(clip_value)
|
105 |
+
|
106 |
+
total_norm = 0
|
107 |
+
for p in parameters:
|
108 |
+
param_norm = p.grad.data.norm(norm_type)
|
109 |
+
total_norm += param_norm.item() ** norm_type
|
110 |
+
if clip_value is not None: p.grad.data.clamp_(min=-clip_value, max=clip_value)
|
111 |
+
total_norm = total_norm ** (1.0 / norm_type)
|
112 |
+
return total_norm
|
lib/infer_pack/models.py
ADDED
@@ -0,0 +1,711 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import torch
|
3 |
+
from torch import nn
|
4 |
+
from torch.nn import functional as F
|
5 |
+
from lib.infer_pack import modules
|
6 |
+
from lib.infer_pack import attentions
|
7 |
+
from lib.infer_pack import commons
|
8 |
+
from lib.infer_pack.commons import init_weights, get_padding
|
9 |
+
from torch.nn import Conv1d, ConvTranspose1d, Conv2d
|
10 |
+
from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
|
11 |
+
from lib.infer_pack.commons import init_weights
|
12 |
+
import numpy as np
|
13 |
+
from lib.infer_pack import commons
|
14 |
+
|
15 |
+
class TextEncoder256(nn.Module):
|
16 |
+
def __init__(self, out_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, f0=True):
|
17 |
+
super().__init__()
|
18 |
+
self.out_channels = out_channels
|
19 |
+
self.hidden_channels = hidden_channels
|
20 |
+
self.filter_channels = filter_channels
|
21 |
+
self.n_heads = n_heads
|
22 |
+
self.n_layers = n_layers
|
23 |
+
self.kernel_size = kernel_size
|
24 |
+
self.p_dropout = p_dropout
|
25 |
+
self.emb_phone = nn.Linear(256, hidden_channels)
|
26 |
+
self.lrelu = nn.LeakyReLU(0.1, inplace=True)
|
27 |
+
if f0 == True: self.emb_pitch = nn.Embedding(256, hidden_channels)
|
28 |
+
self.encoder = attentions.Encoder(hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout)
|
29 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
|
30 |
+
|
31 |
+
def forward(self, phone, pitch, lengths):
|
32 |
+
if pitch is None: x = self.emb_phone(phone)
|
33 |
+
else: x = self.emb_phone(phone) + self.emb_pitch(pitch)
|
34 |
+
x = x * math.sqrt(self.hidden_channels)
|
35 |
+
x = self.lrelu(x)
|
36 |
+
x = torch.transpose(x, 1, -1)
|
37 |
+
x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(x.dtype)
|
38 |
+
x = self.encoder(x * x_mask, x_mask)
|
39 |
+
stats = self.proj(x) * x_mask
|
40 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
41 |
+
return m, logs, x_mask
|
42 |
+
|
43 |
+
class TextEncoder768(nn.Module):
|
44 |
+
def __init__(self, out_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, f0=True):
|
45 |
+
super().__init__()
|
46 |
+
self.out_channels = out_channels
|
47 |
+
self.hidden_channels = hidden_channels
|
48 |
+
self.filter_channels = filter_channels
|
49 |
+
self.n_heads = n_heads
|
50 |
+
self.n_layers = n_layers
|
51 |
+
self.kernel_size = kernel_size
|
52 |
+
self.p_dropout = p_dropout
|
53 |
+
self.emb_phone = nn.Linear(768, hidden_channels)
|
54 |
+
self.lrelu = nn.LeakyReLU(0.1, inplace=True)
|
55 |
+
if f0 == True: self.emb_pitch = nn.Embedding(256, hidden_channels)
|
56 |
+
self.encoder = attentions.Encoder(hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout)
|
57 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
|
58 |
+
|
59 |
+
def forward(self, phone, pitch, lengths):
|
60 |
+
if pitch is None: x = self.emb_phone(phone)
|
61 |
+
else: x = self.emb_phone(phone) + self.emb_pitch(pitch)
|
62 |
+
x = x * math.sqrt(self.hidden_channels)
|
63 |
+
x = self.lrelu(x)
|
64 |
+
x = torch.transpose(x, 1, -1)
|
65 |
+
x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(x.dtype)
|
66 |
+
x = self.encoder(x * x_mask, x_mask)
|
67 |
+
stats = self.proj(x) * x_mask
|
68 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
69 |
+
return m, logs, x_mask
|
70 |
+
|
71 |
+
class ResidualCouplingBlock(nn.Module):
|
72 |
+
def __init__(self, channels, hidden_channels, kernel_size, dilation_rate, n_layers, n_flows=4, gin_channels=0):
|
73 |
+
super().__init__()
|
74 |
+
self.channels = channels
|
75 |
+
self.hidden_channels = hidden_channels
|
76 |
+
self.kernel_size = kernel_size
|
77 |
+
self.dilation_rate = dilation_rate
|
78 |
+
self.n_layers = n_layers
|
79 |
+
self.n_flows = n_flows
|
80 |
+
self.gin_channels = gin_channels
|
81 |
+
self.flows = nn.ModuleList()
|
82 |
+
for _ in range(n_flows):
|
83 |
+
self.flows.append(modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels, mean_only=True))
|
84 |
+
self.flows.append(modules.Flip())
|
85 |
+
|
86 |
+
def forward(self, x, x_mask, g=None, reverse=False):
|
87 |
+
if not reverse:
|
88 |
+
for flow in self.flows: x, _ = flow(x, x_mask, g=g, reverse=reverse)
|
89 |
+
else:
|
90 |
+
for flow in reversed(self.flows): x = flow(x, x_mask, g=g, reverse=reverse)
|
91 |
+
return x
|
92 |
+
|
93 |
+
def remove_weight_norm(self):
|
94 |
+
for i in range(self.n_flows): self.flows[i * 2].remove_weight_norm()
|
95 |
+
|
96 |
+
class PosteriorEncoder(nn.Module):
|
97 |
+
def __init__(
|
98 |
+
self,
|
99 |
+
in_channels,
|
100 |
+
out_channels,
|
101 |
+
hidden_channels,
|
102 |
+
kernel_size,
|
103 |
+
dilation_rate,
|
104 |
+
n_layers,
|
105 |
+
gin_channels=0,
|
106 |
+
):
|
107 |
+
super().__init__()
|
108 |
+
self.in_channels = in_channels
|
109 |
+
self.out_channels = out_channels
|
110 |
+
self.hidden_channels = hidden_channels
|
111 |
+
self.kernel_size = kernel_size
|
112 |
+
self.dilation_rate = dilation_rate
|
113 |
+
self.n_layers = n_layers
|
114 |
+
self.gin_channels = gin_channels
|
115 |
+
|
116 |
+
self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
|
117 |
+
self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
|
118 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
|
119 |
+
|
120 |
+
def forward(self, x, x_lengths, g=None):
|
121 |
+
x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
|
122 |
+
x = self.pre(x) * x_mask
|
123 |
+
x = self.enc(x, x_mask, g=g)
|
124 |
+
stats = self.proj(x) * x_mask
|
125 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
126 |
+
z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
|
127 |
+
return z, m, logs, x_mask
|
128 |
+
|
129 |
+
def remove_weight_norm(self):
|
130 |
+
self.enc.remove_weight_norm()
|
131 |
+
|
132 |
+
class Generator(torch.nn.Module):
|
133 |
+
def __init__(
|
134 |
+
self,
|
135 |
+
initial_channel,
|
136 |
+
resblock,
|
137 |
+
resblock_kernel_sizes,
|
138 |
+
resblock_dilation_sizes,
|
139 |
+
upsample_rates,
|
140 |
+
upsample_initial_channel,
|
141 |
+
upsample_kernel_sizes,
|
142 |
+
gin_channels=0,
|
143 |
+
):
|
144 |
+
super(Generator, self).__init__()
|
145 |
+
self.num_kernels = len(resblock_kernel_sizes)
|
146 |
+
self.num_upsamples = len(upsample_rates)
|
147 |
+
self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
|
148 |
+
resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
|
149 |
+
|
150 |
+
self.ups = nn.ModuleList()
|
151 |
+
for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)): self.ups.append(weight_norm(ConvTranspose1d(upsample_initial_channel // (2**i), upsample_initial_channel // (2 ** (i + 1)), k, u, padding=(k - u) // 2)))
|
152 |
+
|
153 |
+
self.resblocks = nn.ModuleList()
|
154 |
+
for i in range(len(self.ups)):
|
155 |
+
ch = upsample_initial_channel // (2 ** (i + 1))
|
156 |
+
for k, d in zip(resblock_kernel_sizes, resblock_dilation_sizes): self.resblocks.append(resblock(ch, k, d))
|
157 |
+
|
158 |
+
self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
|
159 |
+
self.ups.apply(init_weights)
|
160 |
+
|
161 |
+
if gin_channels != 0: self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
|
162 |
+
|
163 |
+
def forward(self, x, g=None):
|
164 |
+
x = self.conv_pre(x)
|
165 |
+
if g is not None: x = x + self.cond(g)
|
166 |
+
|
167 |
+
for i in range(self.num_upsamples):
|
168 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
169 |
+
x = self.ups[i](x)
|
170 |
+
xs = None
|
171 |
+
for j in range(self.num_kernels):
|
172 |
+
if xs is None: xs = self.resblocks[i * self.num_kernels + j](x)
|
173 |
+
else: xs += self.resblocks[i * self.num_kernels + j](x)
|
174 |
+
x = xs / self.num_kernels
|
175 |
+
x = F.leaky_relu(x)
|
176 |
+
x = self.conv_post(x)
|
177 |
+
x = torch.tanh(x)
|
178 |
+
return x
|
179 |
+
|
180 |
+
def remove_weight_norm(self):
|
181 |
+
for l in self.ups: remove_weight_norm(l)
|
182 |
+
for l in self.resblocks: l.remove_weight_norm()
|
183 |
+
|
184 |
+
class SineGen(torch.nn.Module):
|
185 |
+
def __init__(self, samp_rate, harmonic_num=0, sine_amp=0.1, noise_std=0.003, voiced_threshold=0):
|
186 |
+
super(SineGen, self).__init__()
|
187 |
+
self.sine_amp = sine_amp
|
188 |
+
self.noise_std = noise_std
|
189 |
+
self.harmonic_num = harmonic_num
|
190 |
+
self.dim = self.harmonic_num + 1
|
191 |
+
self.sampling_rate = samp_rate
|
192 |
+
self.voiced_threshold = voiced_threshold
|
193 |
+
|
194 |
+
def _f02uv(self, f0):
|
195 |
+
uv = torch.ones_like(f0)
|
196 |
+
uv = uv * (f0 > self.voiced_threshold)
|
197 |
+
return uv
|
198 |
+
|
199 |
+
def forward(self, f0, upp):
|
200 |
+
with torch.no_grad():
|
201 |
+
f0 = f0[:, None].transpose(1, 2)
|
202 |
+
f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device)
|
203 |
+
f0_buf[:, :, 0] = f0[:, :, 0]
|
204 |
+
for idx in np.arange(self.harmonic_num): f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (idx + 2)
|
205 |
+
rad_values = (f0_buf / self.sampling_rate) % 1
|
206 |
+
rand_ini = torch.rand(f0_buf.shape[0], f0_buf.shape[2], device=f0_buf.device)
|
207 |
+
rand_ini[:, 0] = 0
|
208 |
+
rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
|
209 |
+
tmp_over_one = torch.cumsum(rad_values, 1)
|
210 |
+
tmp_over_one *= upp
|
211 |
+
tmp_over_one = F.interpolate(tmp_over_one.transpose(2, 1), scale_factor=upp, mode="linear", align_corners=True).transpose(2, 1)
|
212 |
+
rad_values = F.interpolate(rad_values.transpose(2, 1), scale_factor=upp, mode="nearest").transpose(2, 1)
|
213 |
+
tmp_over_one %= 1
|
214 |
+
tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0
|
215 |
+
cumsum_shift = torch.zeros_like(rad_values)
|
216 |
+
cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
|
217 |
+
sine_waves = torch.sin(torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi)
|
218 |
+
sine_waves = sine_waves * self.sine_amp
|
219 |
+
uv = self._f02uv(f0)
|
220 |
+
uv = F.interpolate(uv.transpose(2, 1), scale_factor=upp, mode="nearest").transpose(2, 1)
|
221 |
+
noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
|
222 |
+
noise = noise_amp * torch.randn_like(sine_waves)
|
223 |
+
sine_waves = sine_waves * uv + noise
|
224 |
+
return sine_waves, uv, noise
|
225 |
+
|
226 |
+
|
227 |
+
class SourceModuleHnNSF(torch.nn.Module):
|
228 |
+
def __init__(
|
229 |
+
self,
|
230 |
+
sampling_rate,
|
231 |
+
harmonic_num=0,
|
232 |
+
sine_amp=0.1,
|
233 |
+
add_noise_std=0.003,
|
234 |
+
voiced_threshod=0,
|
235 |
+
is_half=True,
|
236 |
+
):
|
237 |
+
super(SourceModuleHnNSF, self).__init__()
|
238 |
+
|
239 |
+
self.sine_amp = sine_amp
|
240 |
+
self.noise_std = add_noise_std
|
241 |
+
self.is_half = is_half
|
242 |
+
self.l_sin_gen = SineGen(sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod)
|
243 |
+
self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
|
244 |
+
self.l_tanh = torch.nn.Tanh()
|
245 |
+
|
246 |
+
def forward(self, x, upp=None):
|
247 |
+
sine_wavs, uv, _ = self.l_sin_gen(x, upp)
|
248 |
+
if self.is_half: sine_wavs = sine_wavs.half()
|
249 |
+
sine_merge = self.l_tanh(self.l_linear(sine_wavs))
|
250 |
+
return sine_merge, None, None
|
251 |
+
|
252 |
+
|
253 |
+
class GeneratorNSF(torch.nn.Module):
|
254 |
+
def __init__(
|
255 |
+
self,
|
256 |
+
initial_channel,
|
257 |
+
resblock,
|
258 |
+
resblock_kernel_sizes,
|
259 |
+
resblock_dilation_sizes,
|
260 |
+
upsample_rates,
|
261 |
+
upsample_initial_channel,
|
262 |
+
upsample_kernel_sizes,
|
263 |
+
gin_channels,
|
264 |
+
sr,
|
265 |
+
is_half=False,
|
266 |
+
):
|
267 |
+
super(GeneratorNSF, self).__init__()
|
268 |
+
self.num_kernels = len(resblock_kernel_sizes)
|
269 |
+
self.num_upsamples = len(upsample_rates)
|
270 |
+
self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates))
|
271 |
+
self.m_source = SourceModuleHnNSF(sampling_rate=sr, harmonic_num=0, is_half=is_half)
|
272 |
+
self.noise_convs = nn.ModuleList()
|
273 |
+
self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
|
274 |
+
resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
|
275 |
+
|
276 |
+
self.ups = nn.ModuleList()
|
277 |
+
for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
|
278 |
+
c_cur = upsample_initial_channel // (2 ** (i + 1))
|
279 |
+
self.ups.append(
|
280 |
+
weight_norm(ConvTranspose1d(upsample_initial_channel // (2**i), upsample_initial_channel // (2 ** (i + 1)), k, u, padding=(k - u) // 2)))
|
281 |
+
if i + 1 < len(upsample_rates):
|
282 |
+
stride_f0 = np.prod(upsample_rates[i + 1 :])
|
283 |
+
self.noise_convs.append(Conv1d(1, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=stride_f0 // 2))
|
284 |
+
else: self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1))
|
285 |
+
|
286 |
+
self.resblocks = nn.ModuleList()
|
287 |
+
for i in range(len(self.ups)):
|
288 |
+
ch = upsample_initial_channel // (2 ** (i + 1))
|
289 |
+
for k, d in zip(resblock_kernel_sizes, resblock_dilation_sizes): self.resblocks.append(resblock(ch, k, d))
|
290 |
+
|
291 |
+
self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
|
292 |
+
self.ups.apply(init_weights)
|
293 |
+
|
294 |
+
if gin_channels != 0: self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
|
295 |
+
|
296 |
+
self.upp = np.prod(upsample_rates)
|
297 |
+
|
298 |
+
def forward(self, x, f0, g=None):
|
299 |
+
har_source, noi_source, uv = self.m_source(f0, self.upp)
|
300 |
+
har_source = har_source.transpose(1, 2)
|
301 |
+
x = self.conv_pre(x)
|
302 |
+
if g is not None: x = x + self.cond(g)
|
303 |
+
|
304 |
+
for i in range(self.num_upsamples):
|
305 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
306 |
+
x = self.ups[i](x)
|
307 |
+
x_source = self.noise_convs[i](har_source)
|
308 |
+
x = x + x_source
|
309 |
+
xs = None
|
310 |
+
for j in range(self.num_kernels):
|
311 |
+
if xs is None: xs = self.resblocks[i * self.num_kernels + j](x)
|
312 |
+
else: xs += self.resblocks[i * self.num_kernels + j](x)
|
313 |
+
x = xs / self.num_kernels
|
314 |
+
x = F.leaky_relu(x)
|
315 |
+
x = self.conv_post(x)
|
316 |
+
x = torch.tanh(x)
|
317 |
+
return x
|
318 |
+
|
319 |
+
def remove_weight_norm(self):
|
320 |
+
for l in self.ups: remove_weight_norm(l)
|
321 |
+
for l in self.resblocks: l.remove_weight_norm()
|
322 |
+
|
323 |
+
sr2sr = {
|
324 |
+
"32k": 32000,
|
325 |
+
"40k": 40000,
|
326 |
+
"48k": 48000,
|
327 |
+
}
|
328 |
+
|
329 |
+
class SynthesizerTrnMs256NSFsid(nn.Module):
|
330 |
+
def __init__(
|
331 |
+
self,
|
332 |
+
spec_channels,
|
333 |
+
segment_size,
|
334 |
+
inter_channels,
|
335 |
+
hidden_channels,
|
336 |
+
filter_channels,
|
337 |
+
n_heads,
|
338 |
+
n_layers,
|
339 |
+
kernel_size,
|
340 |
+
p_dropout,
|
341 |
+
resblock,
|
342 |
+
resblock_kernel_sizes,
|
343 |
+
resblock_dilation_sizes,
|
344 |
+
upsample_rates,
|
345 |
+
upsample_initial_channel,
|
346 |
+
upsample_kernel_sizes,
|
347 |
+
spk_embed_dim,
|
348 |
+
gin_channels,
|
349 |
+
sr,
|
350 |
+
**kwargs
|
351 |
+
):
|
352 |
+
super().__init__()
|
353 |
+
if type(sr) == type("strr"): sr = sr2sr[sr]
|
354 |
+
self.spec_channels = spec_channels
|
355 |
+
self.inter_channels = inter_channels
|
356 |
+
self.hidden_channels = hidden_channels
|
357 |
+
self.filter_channels = filter_channels
|
358 |
+
self.n_heads = n_heads
|
359 |
+
self.n_layers = n_layers
|
360 |
+
self.kernel_size = kernel_size
|
361 |
+
self.p_dropout = p_dropout
|
362 |
+
self.resblock = resblock
|
363 |
+
self.resblock_kernel_sizes = resblock_kernel_sizes
|
364 |
+
self.resblock_dilation_sizes = resblock_dilation_sizes
|
365 |
+
self.upsample_rates = upsample_rates
|
366 |
+
self.upsample_initial_channel = upsample_initial_channel
|
367 |
+
self.upsample_kernel_sizes = upsample_kernel_sizes
|
368 |
+
self.segment_size = segment_size
|
369 |
+
self.gin_channels = gin_channels
|
370 |
+
self.spk_embed_dim = spk_embed_dim
|
371 |
+
self.enc_p = TextEncoder256(inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout)
|
372 |
+
self.dec = GeneratorNSF(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels, sr=sr, is_half=kwargs["is_half"])
|
373 |
+
self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
|
374 |
+
self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels)
|
375 |
+
self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
|
376 |
+
|
377 |
+
def remove_weight_norm(self):
|
378 |
+
self.dec.remove_weight_norm()
|
379 |
+
self.flow.remove_weight_norm()
|
380 |
+
self.enc_q.remove_weight_norm()
|
381 |
+
|
382 |
+
def forward(self, phone, phone_lengths, pitch, pitchf, y, y_lengths, ds):
|
383 |
+
g = self.emb_g(ds).unsqueeze(-1)
|
384 |
+
m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
|
385 |
+
z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
|
386 |
+
z_p = self.flow(z, y_mask, g=g)
|
387 |
+
z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
|
388 |
+
pitchf = commons.slice_segments2(pitchf, ids_slice, self.segment_size)
|
389 |
+
o = self.dec(z_slice, pitchf, g=g)
|
390 |
+
return o, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
|
391 |
+
|
392 |
+
def infer(self, phone, phone_lengths, pitch, nsff0, sid, max_len=None):
|
393 |
+
g = self.emb_g(sid).unsqueeze(-1)
|
394 |
+
m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
|
395 |
+
z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
|
396 |
+
z = self.flow(z_p, x_mask, g=g, reverse=True)
|
397 |
+
o = self.dec((z * x_mask)[:, :, :max_len], nsff0, g=g)
|
398 |
+
return o, x_mask, (z, z_p, m_p, logs_p)
|
399 |
+
|
400 |
+
|
401 |
+
class SynthesizerTrnMs768NSFsid(nn.Module):
|
402 |
+
def __init__(
|
403 |
+
self,
|
404 |
+
spec_channels,
|
405 |
+
segment_size,
|
406 |
+
inter_channels,
|
407 |
+
hidden_channels,
|
408 |
+
filter_channels,
|
409 |
+
n_heads,
|
410 |
+
n_layers,
|
411 |
+
kernel_size,
|
412 |
+
p_dropout,
|
413 |
+
resblock,
|
414 |
+
resblock_kernel_sizes,
|
415 |
+
resblock_dilation_sizes,
|
416 |
+
upsample_rates,
|
417 |
+
upsample_initial_channel,
|
418 |
+
upsample_kernel_sizes,
|
419 |
+
spk_embed_dim,
|
420 |
+
gin_channels,
|
421 |
+
sr,
|
422 |
+
**kwargs
|
423 |
+
):
|
424 |
+
super().__init__()
|
425 |
+
if type(sr) == type("strr"): sr = sr2sr[sr]
|
426 |
+
self.spec_channels = spec_channels
|
427 |
+
self.inter_channels = inter_channels
|
428 |
+
self.hidden_channels = hidden_channels
|
429 |
+
self.filter_channels = filter_channels
|
430 |
+
self.n_heads = n_heads
|
431 |
+
self.n_layers = n_layers
|
432 |
+
self.kernel_size = kernel_size
|
433 |
+
self.p_dropout = p_dropout
|
434 |
+
self.resblock = resblock
|
435 |
+
self.resblock_kernel_sizes = resblock_kernel_sizes
|
436 |
+
self.resblock_dilation_sizes = resblock_dilation_sizes
|
437 |
+
self.upsample_rates = upsample_rates
|
438 |
+
self.upsample_initial_channel = upsample_initial_channel
|
439 |
+
self.upsample_kernel_sizes = upsample_kernel_sizes
|
440 |
+
self.segment_size = segment_size
|
441 |
+
self.gin_channels = gin_channels
|
442 |
+
self.spk_embed_dim = spk_embed_dim
|
443 |
+
self.enc_p = TextEncoder768(inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout)
|
444 |
+
self.dec = GeneratorNSF(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels, sr=sr, is_half=kwargs["is_half"])
|
445 |
+
self.enc_q = PosteriorEncoder(spec_channels,inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
|
446 |
+
self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels)
|
447 |
+
self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
|
448 |
+
|
449 |
+
def remove_weight_norm(self):
|
450 |
+
self.dec.remove_weight_norm()
|
451 |
+
self.flow.remove_weight_norm()
|
452 |
+
self.enc_q.remove_weight_norm()
|
453 |
+
|
454 |
+
def forward(self, phone, phone_lengths, pitch, pitchf, y, y_lengths, ds):
|
455 |
+
g = self.emb_g(ds).unsqueeze(-1)
|
456 |
+
m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
|
457 |
+
z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
|
458 |
+
z_p = self.flow(z, y_mask, g=g)
|
459 |
+
z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
|
460 |
+
pitchf = commons.slice_segments2(pitchf, ids_slice, self.segment_size)
|
461 |
+
o = self.dec(z_slice, pitchf, g=g)
|
462 |
+
return o, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
|
463 |
+
|
464 |
+
def infer(self, phone, phone_lengths, pitch, nsff0, sid, max_len=None):
|
465 |
+
g = self.emb_g(sid).unsqueeze(-1)
|
466 |
+
m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
|
467 |
+
z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
|
468 |
+
z = self.flow(z_p, x_mask, g=g, reverse=True)
|
469 |
+
o = self.dec((z * x_mask)[:, :, :max_len], nsff0, g=g)
|
470 |
+
return o, x_mask, (z, z_p, m_p, logs_p)
|
471 |
+
|
472 |
+
class SynthesizerTrnMs256NSFsid_nono(nn.Module):
|
473 |
+
def __init__(
|
474 |
+
self,
|
475 |
+
spec_channels,
|
476 |
+
segment_size,
|
477 |
+
inter_channels,
|
478 |
+
hidden_channels,
|
479 |
+
filter_channels,
|
480 |
+
n_heads,
|
481 |
+
n_layers,
|
482 |
+
kernel_size,
|
483 |
+
p_dropout,
|
484 |
+
resblock,
|
485 |
+
resblock_kernel_sizes,
|
486 |
+
resblock_dilation_sizes,
|
487 |
+
upsample_rates,
|
488 |
+
upsample_initial_channel,
|
489 |
+
upsample_kernel_sizes,
|
490 |
+
spk_embed_dim,
|
491 |
+
gin_channels,
|
492 |
+
sr=None,
|
493 |
+
**kwargs
|
494 |
+
):
|
495 |
+
super().__init__()
|
496 |
+
self.spec_channels = spec_channels
|
497 |
+
self.inter_channels = inter_channels
|
498 |
+
self.hidden_channels = hidden_channels
|
499 |
+
self.filter_channels = filter_channels
|
500 |
+
self.n_heads = n_heads
|
501 |
+
self.n_layers = n_layers
|
502 |
+
self.kernel_size = kernel_size
|
503 |
+
self.p_dropout = p_dropout
|
504 |
+
self.resblock = resblock
|
505 |
+
self.resblock_kernel_sizes = resblock_kernel_sizes
|
506 |
+
self.resblock_dilation_sizes = resblock_dilation_sizes
|
507 |
+
self.upsample_rates = upsample_rates
|
508 |
+
self.upsample_initial_channel = upsample_initial_channel
|
509 |
+
self.upsample_kernel_sizes = upsample_kernel_sizes
|
510 |
+
self.segment_size = segment_size
|
511 |
+
self.gin_channels = gin_channels
|
512 |
+
self.spk_embed_dim = spk_embed_dim
|
513 |
+
self.enc_p = TextEncoder256(inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, f0=False)
|
514 |
+
self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
|
515 |
+
self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
|
516 |
+
self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels)
|
517 |
+
self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
|
518 |
+
|
519 |
+
def remove_weight_norm(self):
|
520 |
+
self.dec.remove_weight_norm()
|
521 |
+
self.flow.remove_weight_norm()
|
522 |
+
self.enc_q.remove_weight_norm()
|
523 |
+
|
524 |
+
def forward(self, phone, phone_lengths, y, y_lengths, ds):
|
525 |
+
g = self.emb_g(ds).unsqueeze(-1)
|
526 |
+
m_p, logs_p, x_mask = self.enc_p(phone, None, phone_lengths)
|
527 |
+
z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
|
528 |
+
z_p = self.flow(z, y_mask, g=g)
|
529 |
+
z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
|
530 |
+
o = self.dec(z_slice, g=g)
|
531 |
+
return o, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
|
532 |
+
|
533 |
+
def infer(self, phone, phone_lengths, sid, max_len=None):
|
534 |
+
g = self.emb_g(sid).unsqueeze(-1)
|
535 |
+
m_p, logs_p, x_mask = self.enc_p(phone, None, phone_lengths)
|
536 |
+
z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
|
537 |
+
z = self.flow(z_p, x_mask, g=g, reverse=True)
|
538 |
+
o = self.dec((z * x_mask)[:, :, :max_len], g=g)
|
539 |
+
return o, x_mask, (z, z_p, m_p, logs_p)
|
540 |
+
|
541 |
+
class SynthesizerTrnMs768NSFsid_nono(nn.Module):
|
542 |
+
def __init__(
|
543 |
+
self,
|
544 |
+
spec_channels,
|
545 |
+
segment_size,
|
546 |
+
inter_channels,
|
547 |
+
hidden_channels,
|
548 |
+
filter_channels,
|
549 |
+
n_heads,
|
550 |
+
n_layers,
|
551 |
+
kernel_size,
|
552 |
+
p_dropout,
|
553 |
+
resblock,
|
554 |
+
resblock_kernel_sizes,
|
555 |
+
resblock_dilation_sizes,
|
556 |
+
upsample_rates,
|
557 |
+
upsample_initial_channel,
|
558 |
+
upsample_kernel_sizes,
|
559 |
+
spk_embed_dim,
|
560 |
+
gin_channels,
|
561 |
+
sr=None,
|
562 |
+
**kwargs
|
563 |
+
):
|
564 |
+
super().__init__()
|
565 |
+
self.spec_channels = spec_channels
|
566 |
+
self.inter_channels = inter_channels
|
567 |
+
self.hidden_channels = hidden_channels
|
568 |
+
self.filter_channels = filter_channels
|
569 |
+
self.n_heads = n_heads
|
570 |
+
self.n_layers = n_layers
|
571 |
+
self.kernel_size = kernel_size
|
572 |
+
self.p_dropout = p_dropout
|
573 |
+
self.resblock = resblock
|
574 |
+
self.resblock_kernel_sizes = resblock_kernel_sizes
|
575 |
+
self.resblock_dilation_sizes = resblock_dilation_sizes
|
576 |
+
self.upsample_rates = upsample_rates
|
577 |
+
self.upsample_initial_channel = upsample_initial_channel
|
578 |
+
self.upsample_kernel_sizes = upsample_kernel_sizes
|
579 |
+
self.segment_size = segment_size
|
580 |
+
self.gin_channels = gin_channels
|
581 |
+
self.spk_embed_dim = spk_embed_dim
|
582 |
+
self.enc_p = TextEncoder768(inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout, f0=False)
|
583 |
+
self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
|
584 |
+
self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
|
585 |
+
self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels)
|
586 |
+
self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
|
587 |
+
|
588 |
+
def remove_weight_norm(self):
|
589 |
+
self.dec.remove_weight_norm()
|
590 |
+
self.flow.remove_weight_norm()
|
591 |
+
self.enc_q.remove_weight_norm()
|
592 |
+
|
593 |
+
def forward(self, phone, phone_lengths, y, y_lengths, ds):
|
594 |
+
g = self.emb_g(ds).unsqueeze(-1)
|
595 |
+
m_p, logs_p, x_mask = self.enc_p(phone, None, phone_lengths)
|
596 |
+
z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
|
597 |
+
z_p = self.flow(z, y_mask, g=g)
|
598 |
+
z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
|
599 |
+
o = self.dec(z_slice, g=g)
|
600 |
+
return o, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
|
601 |
+
|
602 |
+
def infer(self, phone, phone_lengths, sid, max_len=None):
|
603 |
+
g = self.emb_g(sid).unsqueeze(-1)
|
604 |
+
m_p, logs_p, x_mask = self.enc_p(phone, None, phone_lengths)
|
605 |
+
z_p = (m_p + torch.exp(logs_p) * torch.randn_like(m_p) * 0.66666) * x_mask
|
606 |
+
z = self.flow(z_p, x_mask, g=g, reverse=True)
|
607 |
+
o = self.dec((z * x_mask)[:, :, :max_len], g=g)
|
608 |
+
return o, x_mask, (z, z_p, m_p, logs_p)
|
609 |
+
|
610 |
+
class MultiPeriodDiscriminator(torch.nn.Module):
|
611 |
+
def __init__(self, use_spectral_norm=False):
|
612 |
+
super(MultiPeriodDiscriminator, self).__init__()
|
613 |
+
periods = [2, 3, 5, 7, 11, 17]
|
614 |
+
discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
|
615 |
+
discs += [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
|
616 |
+
self.discriminators = nn.ModuleList(discs)
|
617 |
+
|
618 |
+
def forward(self, y, y_hat):
|
619 |
+
y_d_rs = []
|
620 |
+
y_d_gs = []
|
621 |
+
fmap_rs = []
|
622 |
+
fmap_gs = []
|
623 |
+
for d in self.discriminators:
|
624 |
+
y_d_r, fmap_r = d(y)
|
625 |
+
y_d_g, fmap_g = d(y_hat)
|
626 |
+
y_d_rs.append(y_d_r)
|
627 |
+
y_d_gs.append(y_d_g)
|
628 |
+
fmap_rs.append(fmap_r)
|
629 |
+
fmap_gs.append(fmap_g)
|
630 |
+
|
631 |
+
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
632 |
+
|
633 |
+
class MultiPeriodDiscriminatorV2(torch.nn.Module):
|
634 |
+
def __init__(self, use_spectral_norm=False):
|
635 |
+
super(MultiPeriodDiscriminatorV2, self).__init__()
|
636 |
+
periods = [2, 3, 5, 7, 11, 17, 23, 37]
|
637 |
+
|
638 |
+
discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
|
639 |
+
discs += [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
|
640 |
+
self.discriminators = nn.ModuleList(discs)
|
641 |
+
|
642 |
+
def forward(self, y, y_hat):
|
643 |
+
y_d_rs = []
|
644 |
+
y_d_gs = []
|
645 |
+
fmap_rs = []
|
646 |
+
fmap_gs = []
|
647 |
+
for d in self.discriminators:
|
648 |
+
y_d_r, fmap_r = d(y)
|
649 |
+
y_d_g, fmap_g = d(y_hat)
|
650 |
+
y_d_rs.append(y_d_r)
|
651 |
+
y_d_gs.append(y_d_g)
|
652 |
+
fmap_rs.append(fmap_r)
|
653 |
+
fmap_gs.append(fmap_g)
|
654 |
+
|
655 |
+
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
656 |
+
|
657 |
+
class DiscriminatorS(torch.nn.Module):
|
658 |
+
def __init__(self, use_spectral_norm=False):
|
659 |
+
super(DiscriminatorS, self).__init__()
|
660 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
661 |
+
self.convs = nn.ModuleList([norm_f(Conv1d(1, 16, 15, 1, padding=7)), norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)), norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)), norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)), norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)), norm_f(Conv1d(1024, 1024, 5, 1, padding=2))])
|
662 |
+
self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
|
663 |
+
|
664 |
+
def forward(self, x):
|
665 |
+
fmap = []
|
666 |
+
for l in self.convs:
|
667 |
+
x = l(x)
|
668 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
669 |
+
fmap.append(x)
|
670 |
+
x = self.conv_post(x)
|
671 |
+
fmap.append(x)
|
672 |
+
x = torch.flatten(x, 1, -1)
|
673 |
+
|
674 |
+
return x, fmap
|
675 |
+
|
676 |
+
|
677 |
+
class DiscriminatorP(torch.nn.Module):
|
678 |
+
def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
|
679 |
+
super(DiscriminatorP, self).__init__()
|
680 |
+
self.period = period
|
681 |
+
self.use_spectral_norm = use_spectral_norm
|
682 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
683 |
+
self.convs = nn.ModuleList(
|
684 |
+
[
|
685 |
+
norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
|
686 |
+
norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
|
687 |
+
norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
|
688 |
+
norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
|
689 |
+
norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
|
690 |
+
]
|
691 |
+
)
|
692 |
+
self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
|
693 |
+
|
694 |
+
def forward(self, x):
|
695 |
+
fmap = []
|
696 |
+
b, c, t = x.shape
|
697 |
+
if t % self.period != 0:
|
698 |
+
n_pad = self.period - (t % self.period)
|
699 |
+
x = F.pad(x, (0, n_pad), "reflect")
|
700 |
+
t = t + n_pad
|
701 |
+
x = x.view(b, c, t // self.period, self.period)
|
702 |
+
|
703 |
+
for l in self.convs:
|
704 |
+
x = l(x)
|
705 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
706 |
+
fmap.append(x)
|
707 |
+
x = self.conv_post(x)
|
708 |
+
fmap.append(x)
|
709 |
+
x = torch.flatten(x, 1, -1)
|
710 |
+
|
711 |
+
return x, fmap
|
lib/infer_pack/models_onnx.py
ADDED
@@ -0,0 +1,582 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import torch
|
3 |
+
import numpy as np
|
4 |
+
from torch import nn
|
5 |
+
from torch.nn import functional as F
|
6 |
+
from torch.nn import Conv1d, ConvTranspose1d, Conv2d
|
7 |
+
from lib.infer_pack import modules, attentions, commons
|
8 |
+
from torch.nn.utils import weight_norm, remove_weight_norm
|
9 |
+
from lib.infer_pack.commons import init_weights, get_padding
|
10 |
+
from lib.infer_pack.commons import init_weights, sequence_mask
|
11 |
+
from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
|
12 |
+
from lib.infer_pack.modules import ResidualCouplingLayer, WN, ResBlock1, ResBlock2, LRELU_SLOPE
|
13 |
+
|
14 |
+
class TextEncoder(nn.Module):
|
15 |
+
def __init__(
|
16 |
+
self,
|
17 |
+
input_dim,
|
18 |
+
out_channels,
|
19 |
+
hidden_channels,
|
20 |
+
filter_channels,
|
21 |
+
n_heads,
|
22 |
+
n_layers,
|
23 |
+
kernel_size,
|
24 |
+
p_dropout,
|
25 |
+
f0=True,
|
26 |
+
):
|
27 |
+
super().__init__()
|
28 |
+
self.out_channels = out_channels
|
29 |
+
self.hidden_channels = hidden_channels
|
30 |
+
self.filter_channels = filter_channels
|
31 |
+
self.n_heads = n_heads
|
32 |
+
self.n_layers = n_layers
|
33 |
+
self.kernel_size = kernel_size
|
34 |
+
self.p_dropout = p_dropout
|
35 |
+
self.emb_phone = nn.Linear(input_dim, hidden_channels)
|
36 |
+
self.lrelu = nn.LeakyReLU(0.1, inplace=True)
|
37 |
+
|
38 |
+
if f0:self.emb_pitch = nn.Embedding(256, hidden_channels)
|
39 |
+
self.encoder = attentions.Encoder(hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout)
|
40 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
|
41 |
+
|
42 |
+
def forward(self, phone, pitch, lengths):
|
43 |
+
x = self.emb_phone(phone) + self.emb_pitch(pitch) if pitch is not None else self.emb_phone(phone)
|
44 |
+
x *= math.sqrt(self.hidden_channels)
|
45 |
+
x = self.lrelu(x)
|
46 |
+
x = torch.transpose(x, 1, -1)
|
47 |
+
x_mask = torch.unsqueeze(sequence_mask(lengths, x.size(2)), 1).to(x.dtype)
|
48 |
+
x = self.encoder(x * x_mask, x_mask)
|
49 |
+
stats = self.proj(x) * x_mask
|
50 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
51 |
+
return m, logs, x_mask
|
52 |
+
|
53 |
+
class TextEncoder768(nn.Module):
|
54 |
+
def __init__(
|
55 |
+
self,
|
56 |
+
out_channels,
|
57 |
+
hidden_channels,
|
58 |
+
filter_channels,
|
59 |
+
n_heads,
|
60 |
+
n_layers,
|
61 |
+
kernel_size,
|
62 |
+
p_dropout,
|
63 |
+
f0=True,
|
64 |
+
):
|
65 |
+
super().__init__()
|
66 |
+
self.out_channels = out_channels
|
67 |
+
self.hidden_channels = hidden_channels
|
68 |
+
self.filter_channels = filter_channels
|
69 |
+
self.n_heads = n_heads
|
70 |
+
self.n_layers = n_layers
|
71 |
+
self.kernel_size = kernel_size
|
72 |
+
self.p_dropout = p_dropout
|
73 |
+
self.emb_phone = nn.Linear(768, hidden_channels)
|
74 |
+
self.lrelu = nn.LeakyReLU(0.1, inplace=True)
|
75 |
+
if f0 == True:self.emb_pitch = nn.Embedding(256, hidden_channels)
|
76 |
+
self.encoder = attentions.Encoder(hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout)
|
77 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
|
78 |
+
|
79 |
+
def forward(self, phone, pitch, lengths):
|
80 |
+
if pitch is None: x = self.emb_phone(phone)
|
81 |
+
else: x = self.emb_phone(phone) + self.emb_pitch(pitch)
|
82 |
+
x = x * math.sqrt(self.hidden_channels)
|
83 |
+
x = self.lrelu(x)
|
84 |
+
x = torch.transpose(x, 1, -1)
|
85 |
+
x_mask = torch.unsqueeze(commons.sequence_mask(lengths, x.size(2)), 1).to(x.dtype)
|
86 |
+
x = self.encoder(x * x_mask, x_mask)
|
87 |
+
stats = self.proj(x) * x_mask
|
88 |
+
|
89 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
90 |
+
return m, logs, x_mask
|
91 |
+
|
92 |
+
class ResidualCouplingBlock(nn.Module):
|
93 |
+
def __init__(
|
94 |
+
self,
|
95 |
+
channels,
|
96 |
+
hidden_channels,
|
97 |
+
kernel_size,
|
98 |
+
dilation_rate,
|
99 |
+
n_layers,
|
100 |
+
n_flows=4,
|
101 |
+
gin_channels=0,
|
102 |
+
):
|
103 |
+
super().__init__()
|
104 |
+
self.channels = channels
|
105 |
+
self.hidden_channels = hidden_channels
|
106 |
+
self.kernel_size = kernel_size
|
107 |
+
self.dilation_rate = dilation_rate
|
108 |
+
self.n_layers = n_layers
|
109 |
+
self.n_flows = n_flows
|
110 |
+
self.gin_channels = gin_channels
|
111 |
+
|
112 |
+
self.flows = nn.ModuleList()
|
113 |
+
for _ in range(n_flows):
|
114 |
+
self.flows.append(
|
115 |
+
modules.ResidualCouplingLayer(
|
116 |
+
channels,
|
117 |
+
hidden_channels,
|
118 |
+
kernel_size,
|
119 |
+
dilation_rate,
|
120 |
+
n_layers,
|
121 |
+
gin_channels=gin_channels,
|
122 |
+
mean_only=True,
|
123 |
+
)
|
124 |
+
)
|
125 |
+
self.flows.append(modules.Flip())
|
126 |
+
|
127 |
+
def forward(self, x, x_mask, g=None, reverse=False):
|
128 |
+
if not reverse:
|
129 |
+
for flow in self.flows: x, _ = flow(x, x_mask, g=g, reverse=reverse)
|
130 |
+
else:
|
131 |
+
for flow in reversed(self.flows): x = flow(x, x_mask, g=g, reverse=reverse)
|
132 |
+
return x
|
133 |
+
|
134 |
+
def remove_weight_norm(self):
|
135 |
+
for i in range(self.n_flows):
|
136 |
+
self.flows[i * 2].remove_weight_norm()
|
137 |
+
|
138 |
+
class PosteriorEncoder(nn.Module):
|
139 |
+
def __init__(
|
140 |
+
self,
|
141 |
+
in_channels,
|
142 |
+
out_channels,
|
143 |
+
hidden_channels,
|
144 |
+
kernel_size,
|
145 |
+
dilation_rate,
|
146 |
+
n_layers,
|
147 |
+
gin_channels=0,
|
148 |
+
):
|
149 |
+
super().__init__()
|
150 |
+
self.in_channels = in_channels
|
151 |
+
self.out_channels = out_channels
|
152 |
+
self.hidden_channels = hidden_channels
|
153 |
+
self.kernel_size = kernel_size
|
154 |
+
self.dilation_rate = dilation_rate
|
155 |
+
self.n_layers = n_layers
|
156 |
+
self.gin_channels = gin_channels
|
157 |
+
|
158 |
+
self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
|
159 |
+
self.enc = modules.WN(
|
160 |
+
hidden_channels,
|
161 |
+
kernel_size,
|
162 |
+
dilation_rate,
|
163 |
+
n_layers,
|
164 |
+
gin_channels=gin_channels,
|
165 |
+
)
|
166 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
|
167 |
+
|
168 |
+
def forward(self, x, x_lengths, g=None):
|
169 |
+
x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(
|
170 |
+
x.dtype
|
171 |
+
)
|
172 |
+
x = self.pre(x) * x_mask
|
173 |
+
x = self.enc(x, x_mask, g=g)
|
174 |
+
stats = self.proj(x) * x_mask
|
175 |
+
m, logs = torch.split(stats, self.out_channels, dim=1)
|
176 |
+
z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
|
177 |
+
return z, m, logs, x_mask
|
178 |
+
|
179 |
+
def remove_weight_norm(self):
|
180 |
+
self.enc.remove_weight_norm()
|
181 |
+
|
182 |
+
class Generator(torch.nn.Module):
|
183 |
+
def __init__(
|
184 |
+
self,
|
185 |
+
initial_channel,
|
186 |
+
resblock,
|
187 |
+
resblock_kernel_sizes,
|
188 |
+
resblock_dilation_sizes,
|
189 |
+
upsample_rates,
|
190 |
+
upsample_initial_channel,
|
191 |
+
upsample_kernel_sizes,
|
192 |
+
gin_channels=0,
|
193 |
+
):
|
194 |
+
super(Generator, self).__init__()
|
195 |
+
self.num_kernels = len(resblock_kernel_sizes)
|
196 |
+
self.num_upsamples = len(upsample_rates)
|
197 |
+
self.conv_pre = Conv1d(
|
198 |
+
initial_channel, upsample_initial_channel, 7, 1, padding=3
|
199 |
+
)
|
200 |
+
resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
|
201 |
+
|
202 |
+
self.ups = nn.ModuleList()
|
203 |
+
for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
|
204 |
+
self.ups.append(
|
205 |
+
weight_norm(
|
206 |
+
ConvTranspose1d(
|
207 |
+
upsample_initial_channel // (2**i),
|
208 |
+
upsample_initial_channel // (2 ** (i + 1)),
|
209 |
+
k,
|
210 |
+
u,
|
211 |
+
padding=(k - u) // 2,
|
212 |
+
)
|
213 |
+
)
|
214 |
+
)
|
215 |
+
|
216 |
+
self.resblocks = nn.ModuleList()
|
217 |
+
for i in range(len(self.ups)):
|
218 |
+
ch = upsample_initial_channel // (2 ** (i + 1))
|
219 |
+
for k, d in zip(resblock_kernel_sizes, resblock_dilation_sizes): self.resblocks.append(resblock(ch, k, d))
|
220 |
+
|
221 |
+
self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
|
222 |
+
self.ups.apply(init_weights)
|
223 |
+
|
224 |
+
if gin_channels != 0: self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
|
225 |
+
|
226 |
+
def forward(self, x, g=None):
|
227 |
+
x = self.conv_pre(x)
|
228 |
+
if g is not None: x = x + self.cond(g)
|
229 |
+
|
230 |
+
for i in range(self.num_upsamples):
|
231 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
232 |
+
x = self.ups[i](x)
|
233 |
+
xs = None
|
234 |
+
for j in range(self.num_kernels):
|
235 |
+
if xs is None: xs = self.resblocks[i * self.num_kernels + j](x)
|
236 |
+
else: xs += self.resblocks[i * self.num_kernels + j](x)
|
237 |
+
x = xs / self.num_kernels
|
238 |
+
x = F.leaky_relu(x)
|
239 |
+
x = self.conv_post(x)
|
240 |
+
x = torch.tanh(x)
|
241 |
+
|
242 |
+
return x
|
243 |
+
|
244 |
+
def remove_weight_norm(self):
|
245 |
+
for l in self.ups: remove_weight_norm(l)
|
246 |
+
for l in self.resblocks: l.remove_weight_norm()
|
247 |
+
|
248 |
+
class SineGen(torch.nn.Module):
|
249 |
+
def __init__(
|
250 |
+
self,
|
251 |
+
samp_rate,
|
252 |
+
harmonic_num=0,
|
253 |
+
sine_amp=0.1,
|
254 |
+
noise_std=0.003,
|
255 |
+
voiced_threshold=0,
|
256 |
+
):
|
257 |
+
super(SineGen, self).__init__()
|
258 |
+
self.sine_amp = sine_amp
|
259 |
+
self.noise_std = noise_std
|
260 |
+
self.harmonic_num = harmonic_num
|
261 |
+
self.dim = self.harmonic_num + 1
|
262 |
+
self.sampling_rate = samp_rate
|
263 |
+
self.voiced_threshold = voiced_threshold
|
264 |
+
|
265 |
+
def _f02uv(self, f0):
|
266 |
+
uv = torch.ones_like(f0)
|
267 |
+
uv = uv * (f0 > self.voiced_threshold)
|
268 |
+
return uv
|
269 |
+
|
270 |
+
def forward(self, f0, upp):
|
271 |
+
with torch.no_grad():
|
272 |
+
f0 = f0[:, None].transpose(1, 2)
|
273 |
+
f0_buf = torch.zeros(f0.shape[0], f0.shape[1], self.dim, device=f0.device)
|
274 |
+
f0_buf[:, :, 0] = f0[:, :, 0]
|
275 |
+
for idx in np.arange(self.harmonic_num): f0_buf[:, :, idx + 1] = f0_buf[:, :, 0] * (idx + 2)
|
276 |
+
rad_values = (f0_buf / self.sampling_rate) % 1
|
277 |
+
rand_ini = torch.rand(f0_buf.shape[0], f0_buf.shape[2], device=f0_buf.device)
|
278 |
+
rand_ini[:, 0] = 0
|
279 |
+
rad_values[:, 0, :] = rad_values[:, 0, :] + rand_ini
|
280 |
+
tmp_over_one = torch.cumsum(rad_values, 1)
|
281 |
+
tmp_over_one *= upp
|
282 |
+
tmp_over_one = F.interpolate(tmp_over_one.transpose(2, 1), scale_factor=upp, mode="linear", align_corners=True).transpose(2, 1)
|
283 |
+
rad_values = F.interpolate(rad_values.transpose(2, 1), scale_factor=upp, mode="nearest").transpose(2, 1)
|
284 |
+
tmp_over_one %= 1
|
285 |
+
tmp_over_one_idx = (tmp_over_one[:, 1:, :] - tmp_over_one[:, :-1, :]) < 0
|
286 |
+
cumsum_shift = torch.zeros_like(rad_values)
|
287 |
+
cumsum_shift[:, 1:, :] = tmp_over_one_idx * -1.0
|
288 |
+
sine_waves = torch.sin(torch.cumsum(rad_values + cumsum_shift, dim=1) * 2 * np.pi)
|
289 |
+
sine_waves = sine_waves * self.sine_amp
|
290 |
+
uv = self._f02uv(f0)
|
291 |
+
uv = F.interpolate(uv.transpose(2, 1), scale_factor=upp, mode="nearest").transpose(2, 1)
|
292 |
+
noise_amp = uv * self.noise_std + (1 - uv) * self.sine_amp / 3
|
293 |
+
noise = noise_amp * torch.randn_like(sine_waves)
|
294 |
+
sine_waves = sine_waves * uv + noise
|
295 |
+
return sine_waves, uv, noise
|
296 |
+
|
297 |
+
class SourceModuleHnNSF(torch.nn.Module):
|
298 |
+
def __init__(
|
299 |
+
self,
|
300 |
+
sampling_rate,
|
301 |
+
harmonic_num=0,
|
302 |
+
sine_amp=0.1,
|
303 |
+
add_noise_std=0.003,
|
304 |
+
voiced_threshod=0,
|
305 |
+
is_half=True,
|
306 |
+
):
|
307 |
+
super(SourceModuleHnNSF, self).__init__()
|
308 |
+
|
309 |
+
self.sine_amp = sine_amp
|
310 |
+
self.noise_std = add_noise_std
|
311 |
+
self.is_half = is_half
|
312 |
+
self.l_sin_gen = SineGen(sampling_rate, harmonic_num, sine_amp, add_noise_std, voiced_threshod)
|
313 |
+
self.l_linear = torch.nn.Linear(harmonic_num + 1, 1)
|
314 |
+
self.l_tanh = torch.nn.Tanh()
|
315 |
+
|
316 |
+
def forward(self, x, upp=None):
|
317 |
+
sine_wavs, uv, _ = self.l_sin_gen(x, upp)
|
318 |
+
if self.is_half: sine_wavs = sine_wavs.half()
|
319 |
+
sine_merge = self.l_tanh(self.l_linear(sine_wavs))
|
320 |
+
return sine_merge, None, None
|
321 |
+
|
322 |
+
|
323 |
+
class GeneratorNSF(torch.nn.Module):
|
324 |
+
def __init__(
|
325 |
+
self,
|
326 |
+
initial_channel,
|
327 |
+
resblock,
|
328 |
+
resblock_kernel_sizes,
|
329 |
+
resblock_dilation_sizes,
|
330 |
+
upsample_rates,
|
331 |
+
upsample_initial_channel,
|
332 |
+
upsample_kernel_sizes,
|
333 |
+
gin_channels,
|
334 |
+
sr,
|
335 |
+
is_half=False,
|
336 |
+
):
|
337 |
+
super(GeneratorNSF, self).__init__()
|
338 |
+
self.num_kernels = len(resblock_kernel_sizes)
|
339 |
+
self.num_upsamples = len(upsample_rates)
|
340 |
+
self.f0_upsamp = torch.nn.Upsample(scale_factor=np.prod(upsample_rates))
|
341 |
+
self.m_source = SourceModuleHnNSF(sampling_rate=sr, harmonic_num=0, is_half=is_half)
|
342 |
+
self.noise_convs = nn.ModuleList()
|
343 |
+
self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
|
344 |
+
resblock = modules.ResBlock1 if resblock == "1" else modules.ResBlock2
|
345 |
+
|
346 |
+
self.ups = nn.ModuleList()
|
347 |
+
for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
|
348 |
+
c_cur = upsample_initial_channel // (2 ** (i + 1))
|
349 |
+
self.ups.append(weight_norm(ConvTranspose1d(upsample_initial_channel // (2**i), upsample_initial_channel // (2 ** (i + 1)), k, u, padding=(k - u) // 2,)))
|
350 |
+
if i + 1 < len(upsample_rates):
|
351 |
+
stride_f0 = np.prod(upsample_rates[i + 1 :])
|
352 |
+
self.noise_convs.append(Conv1d(1, c_cur, kernel_size=stride_f0 * 2, stride=stride_f0, padding=stride_f0 // 2,))
|
353 |
+
else: self.noise_convs.append(Conv1d(1, c_cur, kernel_size=1))
|
354 |
+
|
355 |
+
self.resblocks = nn.ModuleList()
|
356 |
+
for i in range(len(self.ups)):
|
357 |
+
ch = upsample_initial_channel // (2 ** (i + 1))
|
358 |
+
for k, d in zip(resblock_kernel_sizes, resblock_dilation_sizes): self.resblocks.append(resblock(ch, k, d))
|
359 |
+
|
360 |
+
self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
|
361 |
+
self.ups.apply(init_weights)
|
362 |
+
|
363 |
+
if gin_channels != 0: self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
|
364 |
+
|
365 |
+
self.upp = np.prod(upsample_rates)
|
366 |
+
|
367 |
+
def forward(self, x, f0, g=None):
|
368 |
+
har_source, noi_source, uv = self.m_source(f0, self.upp)
|
369 |
+
har_source = har_source.transpose(1, 2)
|
370 |
+
x = self.conv_pre(x)
|
371 |
+
if g is not None: x = x + self.cond(g)
|
372 |
+
|
373 |
+
for i in range(self.num_upsamples):
|
374 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
375 |
+
x = self.ups[i](x)
|
376 |
+
x_source = self.noise_convs[i](har_source)
|
377 |
+
x = x + x_source
|
378 |
+
xs = None
|
379 |
+
for j in range(self.num_kernels):
|
380 |
+
if xs is None: xs = self.resblocks[i * self.num_kernels + j](x)
|
381 |
+
else: xs += self.resblocks[i * self.num_kernels + j](x)
|
382 |
+
x = xs / self.num_kernels
|
383 |
+
x = F.leaky_relu(x)
|
384 |
+
x = self.conv_post(x)
|
385 |
+
x = torch.tanh(x)
|
386 |
+
return x
|
387 |
+
|
388 |
+
def remove_weight_norm(self):
|
389 |
+
for l in self.ups:
|
390 |
+
remove_weight_norm(l)
|
391 |
+
for l in self.resblocks:
|
392 |
+
l.remove_weight_norm()
|
393 |
+
|
394 |
+
sr2sr = {"32k": 32000,"40k": 40000,"48k": 48000,}
|
395 |
+
|
396 |
+
class SynthesizerTrnMsNSFsidM(nn.Module):
|
397 |
+
def __init__(
|
398 |
+
self,
|
399 |
+
spec_channels,
|
400 |
+
segment_size,
|
401 |
+
inter_channels,
|
402 |
+
hidden_channels,
|
403 |
+
filter_channels,
|
404 |
+
n_heads,
|
405 |
+
n_layers,
|
406 |
+
kernel_size,
|
407 |
+
p_dropout,
|
408 |
+
resblock,
|
409 |
+
resblock_kernel_sizes,
|
410 |
+
resblock_dilation_sizes,
|
411 |
+
upsample_rates,
|
412 |
+
upsample_initial_channel,
|
413 |
+
upsample_kernel_sizes,
|
414 |
+
spk_embed_dim,
|
415 |
+
gin_channels,
|
416 |
+
sr,
|
417 |
+
version,
|
418 |
+
**kwargs
|
419 |
+
):
|
420 |
+
super().__init__()
|
421 |
+
if type(sr) == type("strr"): sr = sr2sr[sr]
|
422 |
+
self.spec_channels = spec_channels
|
423 |
+
self.inter_channels = inter_channels
|
424 |
+
self.hidden_channels = hidden_channels
|
425 |
+
self.filter_channels = filter_channels
|
426 |
+
self.n_heads = n_heads
|
427 |
+
self.n_layers = n_layers
|
428 |
+
self.kernel_size = kernel_size
|
429 |
+
self.p_dropout = p_dropout
|
430 |
+
self.resblock = resblock
|
431 |
+
self.resblock_kernel_sizes = resblock_kernel_sizes
|
432 |
+
self.resblock_dilation_sizes = resblock_dilation_sizes
|
433 |
+
self.upsample_rates = upsample_rates
|
434 |
+
self.upsample_initial_channel = upsample_initial_channel
|
435 |
+
self.upsample_kernel_sizes = upsample_kernel_sizes
|
436 |
+
self.segment_size = segment_size
|
437 |
+
self.gin_channels = gin_channels
|
438 |
+
self.spk_embed_dim = spk_embed_dim
|
439 |
+
if version == "v1": self.enc_p = TextEncoder(inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout)
|
440 |
+
else: self.enc_p = TextEncoder768(inter_channels, hidden_channels, filter_channels, n_heads, n_layers, kernel_size, p_dropout)
|
441 |
+
self.dec = GeneratorNSF(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates, upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels, sr=sr, is_half=kwargs["is_half"])
|
442 |
+
self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16, gin_channels=gin_channels)
|
443 |
+
self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 3, gin_channels=gin_channels)
|
444 |
+
self.emb_g = nn.Embedding(self.spk_embed_dim, gin_channels)
|
445 |
+
self.speaker_map = None
|
446 |
+
|
447 |
+
def remove_weight_norm(self):
|
448 |
+
self.dec.remove_weight_norm()
|
449 |
+
self.flow.remove_weight_norm()
|
450 |
+
self.enc_q.remove_weight_norm()
|
451 |
+
|
452 |
+
def construct_spkmixmap(self, n_speaker):
|
453 |
+
self.speaker_map = torch.zeros((n_speaker, 1, 1, self.gin_channels))
|
454 |
+
for i in range(n_speaker): self.speaker_map[i] = self.emb_g(torch.LongTensor([[i]]))
|
455 |
+
self.speaker_map = self.speaker_map.unsqueeze(0)
|
456 |
+
|
457 |
+
def forward(self, phone, phone_lengths, pitch, nsff0, g, rnd, max_len=None):
|
458 |
+
if self.speaker_map is not None:
|
459 |
+
g = g.reshape((g.shape[0], g.shape[1], 1, 1, 1))
|
460 |
+
g = g * self.speaker_map
|
461 |
+
g = torch.sum(g, dim=1)
|
462 |
+
g = g.transpose(0, -1).transpose(0, -2).squeeze(0)
|
463 |
+
else:
|
464 |
+
g = g.unsqueeze(0)
|
465 |
+
g = self.emb_g(g).transpose(1, 2)
|
466 |
+
|
467 |
+
m_p, logs_p, x_mask = self.enc_p(phone, pitch, phone_lengths)
|
468 |
+
z_p = (m_p + torch.exp(logs_p) * rnd) * x_mask
|
469 |
+
z = self.flow(z_p, x_mask, g=g, reverse=True)
|
470 |
+
return self.dec((z * x_mask)[:, :, :max_len], nsff0, g=g)
|
471 |
+
|
472 |
+
class MultiPeriodDiscriminator(torch.nn.Module):
|
473 |
+
def __init__(self, use_spectral_norm=False):
|
474 |
+
super(MultiPeriodDiscriminator, self).__init__()
|
475 |
+
periods = [2, 3, 5, 7, 11, 17]
|
476 |
+
|
477 |
+
discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
|
478 |
+
discs += [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
|
479 |
+
self.discriminators = nn.ModuleList(discs)
|
480 |
+
|
481 |
+
def forward(self, y, y_hat):
|
482 |
+
y_d_rs = []
|
483 |
+
y_d_gs = []
|
484 |
+
fmap_rs = []
|
485 |
+
fmap_gs = []
|
486 |
+
for d in self.discriminators:
|
487 |
+
y_d_r, fmap_r = d(y)
|
488 |
+
y_d_g, fmap_g = d(y_hat)
|
489 |
+
y_d_rs.append(y_d_r)
|
490 |
+
y_d_gs.append(y_d_g)
|
491 |
+
fmap_rs.append(fmap_r)
|
492 |
+
fmap_gs.append(fmap_g)
|
493 |
+
|
494 |
+
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
495 |
+
|
496 |
+
class MultiPeriodDiscriminatorV2(torch.nn.Module):
|
497 |
+
def __init__(self, use_spectral_norm=False):
|
498 |
+
super(MultiPeriodDiscriminatorV2, self).__init__()
|
499 |
+
periods = [2, 3, 5, 7, 11, 17, 23, 37]
|
500 |
+
discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
|
501 |
+
discs += [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
|
502 |
+
self.discriminators = nn.ModuleList(discs)
|
503 |
+
|
504 |
+
def forward(self, y, y_hat):
|
505 |
+
y_d_rs = []
|
506 |
+
y_d_gs = []
|
507 |
+
fmap_rs = []
|
508 |
+
fmap_gs = []
|
509 |
+
for d in self.discriminators:
|
510 |
+
y_d_r, fmap_r = d(y)
|
511 |
+
y_d_g, fmap_g = d(y_hat)
|
512 |
+
y_d_rs.append(y_d_r)
|
513 |
+
y_d_gs.append(y_d_g)
|
514 |
+
fmap_rs.append(fmap_r)
|
515 |
+
fmap_gs.append(fmap_g)
|
516 |
+
|
517 |
+
return y_d_rs, y_d_gs, fmap_rs, fmap_gs
|
518 |
+
|
519 |
+
class DiscriminatorS(torch.nn.Module):
|
520 |
+
def __init__(self, use_spectral_norm=False):
|
521 |
+
super(DiscriminatorS, self).__init__()
|
522 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
523 |
+
self.convs = nn.ModuleList(
|
524 |
+
[
|
525 |
+
norm_f(Conv1d(1, 16, 15, 1, padding=7)),
|
526 |
+
norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
|
527 |
+
norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
|
528 |
+
norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
|
529 |
+
norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
|
530 |
+
norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
|
531 |
+
]
|
532 |
+
)
|
533 |
+
self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
|
534 |
+
|
535 |
+
def forward(self, x):
|
536 |
+
fmap = []
|
537 |
+
|
538 |
+
for l in self.convs:
|
539 |
+
x = l(x)
|
540 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
541 |
+
fmap.append(x)
|
542 |
+
x = self.conv_post(x)
|
543 |
+
fmap.append(x)
|
544 |
+
x = torch.flatten(x, 1, -1)
|
545 |
+
|
546 |
+
return x, fmap
|
547 |
+
|
548 |
+
class DiscriminatorP(torch.nn.Module):
|
549 |
+
def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
|
550 |
+
super(DiscriminatorP, self).__init__()
|
551 |
+
self.period = period
|
552 |
+
self.use_spectral_norm = use_spectral_norm
|
553 |
+
norm_f = weight_norm if use_spectral_norm == False else spectral_norm
|
554 |
+
self.convs = nn.ModuleList(
|
555 |
+
[
|
556 |
+
norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0),)),
|
557 |
+
norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0),)),
|
558 |
+
norm_f(Conv2d( 128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0),)),
|
559 |
+
norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0),)),
|
560 |
+
norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0),)),
|
561 |
+
]
|
562 |
+
)
|
563 |
+
self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
|
564 |
+
|
565 |
+
def forward(self, x):
|
566 |
+
fmap = []
|
567 |
+
b, c, t = x.shape
|
568 |
+
if t % self.period != 0:
|
569 |
+
n_pad = self.period - (t % self.period)
|
570 |
+
x = F.pad(x, (0, n_pad), "reflect")
|
571 |
+
t = t + n_pad
|
572 |
+
x = x.view(b, c, t // self.period, self.period)
|
573 |
+
|
574 |
+
for l in self.convs:
|
575 |
+
x = l(x)
|
576 |
+
x = F.leaky_relu(x, modules.LRELU_SLOPE)
|
577 |
+
fmap.append(x)
|
578 |
+
x = self.conv_post(x)
|
579 |
+
fmap.append(x)
|
580 |
+
x = torch.flatten(x, 1, -1)
|
581 |
+
|
582 |
+
return x, fmap
|
lib/infer_pack/modules.py
ADDED
@@ -0,0 +1,315 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import copy
|
2 |
+
import math
|
3 |
+
import numpy as np
|
4 |
+
import scipy
|
5 |
+
import torch
|
6 |
+
from torch import nn
|
7 |
+
from torch.nn import functional as F
|
8 |
+
from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
|
9 |
+
from torch.nn.utils import weight_norm, remove_weight_norm
|
10 |
+
from lib.infer_pack import commons
|
11 |
+
from lib.infer_pack.commons import init_weights, get_padding
|
12 |
+
from lib.infer_pack.transforms import piecewise_rational_quadratic_transform
|
13 |
+
|
14 |
+
LRELU_SLOPE = 0.1
|
15 |
+
|
16 |
+
class LayerNorm(nn.Module):
|
17 |
+
def __init__(self, channels, eps=1e-5):
|
18 |
+
super().__init__()
|
19 |
+
self.channels = channels
|
20 |
+
self.eps = eps
|
21 |
+
self.gamma = nn.Parameter(torch.ones(channels))
|
22 |
+
self.beta = nn.Parameter(torch.zeros(channels))
|
23 |
+
|
24 |
+
def forward(self, x):
|
25 |
+
x = x.transpose(1, -1)
|
26 |
+
x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
|
27 |
+
return x.transpose(1, -1)
|
28 |
+
|
29 |
+
class ConvReluNorm(nn.Module):
|
30 |
+
def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout):
|
31 |
+
super().__init__()
|
32 |
+
self.in_channels = in_channels
|
33 |
+
self.hidden_channels = hidden_channels
|
34 |
+
self.out_channels = out_channels
|
35 |
+
self.kernel_size = kernel_size
|
36 |
+
self.n_layers = n_layers
|
37 |
+
self.p_dropout = p_dropout
|
38 |
+
assert n_layers > 1, "Number of layers should be larger than 0."
|
39 |
+
self.conv_layers = nn.ModuleList()
|
40 |
+
self.norm_layers = nn.ModuleList()
|
41 |
+
self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size // 2))
|
42 |
+
self.norm_layers.append(LayerNorm(hidden_channels))
|
43 |
+
self.relu_drop = nn.Sequential(nn.ReLU(), nn.Dropout(p_dropout))
|
44 |
+
for _ in range(n_layers - 1):
|
45 |
+
self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size // 2))
|
46 |
+
self.norm_layers.append(LayerNorm(hidden_channels))
|
47 |
+
self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
|
48 |
+
self.proj.weight.data.zero_()
|
49 |
+
self.proj.bias.data.zero_()
|
50 |
+
|
51 |
+
def forward(self, x, x_mask):
|
52 |
+
x_org = x
|
53 |
+
for i in range(self.n_layers):
|
54 |
+
x = self.conv_layers[i](x * x_mask)
|
55 |
+
x = self.norm_layers[i](x)
|
56 |
+
x = self.relu_drop(x)
|
57 |
+
x = x_org + self.proj(x)
|
58 |
+
return x * x_mask
|
59 |
+
|
60 |
+
class DDSConv(nn.Module):
|
61 |
+
def __init__(self, channels, kernel_size, n_layers, p_dropout=0.0):
|
62 |
+
super().__init__()
|
63 |
+
self.channels = channels
|
64 |
+
self.kernel_size = kernel_size
|
65 |
+
self.n_layers = n_layers
|
66 |
+
self.p_dropout = p_dropout
|
67 |
+
self.drop = nn.Dropout(p_dropout)
|
68 |
+
self.convs_sep = nn.ModuleList()
|
69 |
+
self.convs_1x1 = nn.ModuleList()
|
70 |
+
self.norms_1 = nn.ModuleList()
|
71 |
+
self.norms_2 = nn.ModuleList()
|
72 |
+
for i in range(n_layers):
|
73 |
+
dilation = kernel_size**i
|
74 |
+
padding = (kernel_size * dilation - dilation)
|
75 |
+
self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size, groups=channels, dilation=dilation, padding=padding))
|
76 |
+
self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
|
77 |
+
self.norms_1.append(LayerNorm(channels))
|
78 |
+
self.norms_2.append(LayerNorm(channels))
|
79 |
+
|
80 |
+
def forward(self, x, x_mask, g=None):
|
81 |
+
if g is not None: x = x + g
|
82 |
+
for i in range(self.n_layers):
|
83 |
+
y = self.convs_sep[i](x * x_mask)
|
84 |
+
y = self.norms_1[i](y)
|
85 |
+
y = F.gelu(y)
|
86 |
+
y = self.convs_1x1[i](y)
|
87 |
+
y = self.norms_2[i](y)
|
88 |
+
y = F.gelu(y)
|
89 |
+
y = self.drop(y)
|
90 |
+
x = x + y
|
91 |
+
return x * x_mask
|
92 |
+
|
93 |
+
class WN(torch.nn.Module):
|
94 |
+
def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
|
95 |
+
super(WN, self).__init__()
|
96 |
+
assert kernel_size % 2 == 1
|
97 |
+
self.hidden_channels = hidden_channels
|
98 |
+
self.kernel_size = (kernel_size,)
|
99 |
+
self.dilation_rate = dilation_rate
|
100 |
+
self.n_layers = n_layers
|
101 |
+
self.gin_channels = gin_channels
|
102 |
+
self.p_dropout = p_dropout
|
103 |
+
|
104 |
+
self.in_layers = torch.nn.ModuleList()
|
105 |
+
self.res_skip_layers = torch.nn.ModuleList()
|
106 |
+
self.drop = nn.Dropout(p_dropout)
|
107 |
+
|
108 |
+
if gin_channels != 0:
|
109 |
+
cond_layer = torch.nn.Conv1d(gin_channels, 2 * hidden_channels * n_layers, 1)
|
110 |
+
self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name="weight")
|
111 |
+
|
112 |
+
for i in range(n_layers):
|
113 |
+
dilation = dilation_rate**i
|
114 |
+
padding = int((kernel_size * dilation - dilation) / 2)
|
115 |
+
in_layer = torch.nn.Conv1d(hidden_channels, 2 * hidden_channels, kernel_size, dilation=dilation, padding=padding)
|
116 |
+
in_layer = torch.nn.utils.weight_norm(in_layer, name="weight")
|
117 |
+
self.in_layers.append(in_layer)
|
118 |
+
|
119 |
+
if i < n_layers - 1: res_skip_channels = 2 * hidden_channels
|
120 |
+
else: res_skip_channels = hidden_channels
|
121 |
+
|
122 |
+
res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
|
123 |
+
res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name="weight")
|
124 |
+
self.res_skip_layers.append(res_skip_layer)
|
125 |
+
|
126 |
+
def forward(self, x, x_mask, g=None, **kwargs):
|
127 |
+
output = torch.zeros_like(x)
|
128 |
+
n_channels_tensor = torch.IntTensor([self.hidden_channels])
|
129 |
+
|
130 |
+
if g is not None: g = self.cond_layer(g)
|
131 |
+
|
132 |
+
for i in range(self.n_layers):
|
133 |
+
x_in = self.in_layers[i](x)
|
134 |
+
if g is not None:
|
135 |
+
cond_offset = i * 2 * self.hidden_channels
|
136 |
+
g_l = g[:, cond_offset : cond_offset + 2 * self.hidden_channels, :]
|
137 |
+
else: g_l = torch.zeros_like(x_in)
|
138 |
+
|
139 |
+
acts = commons.fused_add_tanh_sigmoid_multiply(x_in, g_l, n_channels_tensor)
|
140 |
+
acts = self.drop(acts)
|
141 |
+
|
142 |
+
res_skip_acts = self.res_skip_layers[i](acts)
|
143 |
+
if i < self.n_layers - 1:
|
144 |
+
res_acts = res_skip_acts[:, : self.hidden_channels, :]
|
145 |
+
x = (x + res_acts) * x_mask
|
146 |
+
output = output + res_skip_acts[:, self.hidden_channels :, :]
|
147 |
+
else: output = output + res_skip_acts
|
148 |
+
return output * x_mask
|
149 |
+
|
150 |
+
def remove_weight_norm(self):
|
151 |
+
if self.gin_channels != 0: torch.nn.utils.remove_weight_norm(self.cond_layer)
|
152 |
+
for l in self.in_layers: torch.nn.utils.remove_weight_norm(l)
|
153 |
+
for l in self.res_skip_layers: torch.nn.utils.remove_weight_norm(l)
|
154 |
+
|
155 |
+
class ResBlock1(torch.nn.Module):
|
156 |
+
def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
|
157 |
+
super(ResBlock1, self).__init__()
|
158 |
+
self.convs1 = nn.ModuleList(
|
159 |
+
[
|
160 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0], padding=get_padding(kernel_size, dilation[0]))),
|
161 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1], padding=get_padding(kernel_size, dilation[1]))),
|
162 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2], padding=get_padding(kernel_size, dilation[2])))])
|
163 |
+
self.convs1.apply(init_weights)
|
164 |
+
|
165 |
+
self.convs2 = nn.ModuleList(
|
166 |
+
[
|
167 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1, padding=get_padding(kernel_size, 1))),
|
168 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1, padding=get_padding(kernel_size, 1))),
|
169 |
+
weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1, padding=get_padding(kernel_size, 1)))])
|
170 |
+
self.convs2.apply(init_weights)
|
171 |
+
|
172 |
+
def forward(self, x, x_mask=None):
|
173 |
+
for c1, c2 in zip(self.convs1, self.convs2):
|
174 |
+
xt = F.leaky_relu(x, LRELU_SLOPE)
|
175 |
+
if x_mask is not None: xt = xt * x_mask
|
176 |
+
xt = c1(xt)
|
177 |
+
xt = F.leaky_relu(xt, LRELU_SLOPE)
|
178 |
+
if x_mask is not None: xt = xt * x_mask
|
179 |
+
xt = c2(xt)
|
180 |
+
x = xt + x
|
181 |
+
if x_mask is not None:
|
182 |
+
x = x * x_mask
|
183 |
+
return x
|
184 |
+
|
185 |
+
def remove_weight_norm(self):
|
186 |
+
for l in self.convs1: remove_weight_norm(l)
|
187 |
+
for l in self.convs2: remove_weight_norm(l)
|
188 |
+
|
189 |
+
class ResBlock2(torch.nn.Module):
|
190 |
+
def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
|
191 |
+
super(ResBlock2, self).__init__()
|
192 |
+
self.convs = nn.ModuleList([weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0], padding=get_padding(kernel_size, dilation[0]))),weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1], padding=get_padding(kernel_size, dilation[1])))])
|
193 |
+
self.convs.apply(init_weights)
|
194 |
+
|
195 |
+
def forward(self, x, x_mask=None):
|
196 |
+
for c in self.convs:
|
197 |
+
xt = F.leaky_relu(x, LRELU_SLOPE)
|
198 |
+
if x_mask is not None: xt = xt * x_mask
|
199 |
+
xt = c(xt)
|
200 |
+
x = xt + x
|
201 |
+
if x_mask is not None: x = x * x_mask
|
202 |
+
return x
|
203 |
+
|
204 |
+
def remove_weight_norm(self):
|
205 |
+
for l in self.convs: remove_weight_norm(l)
|
206 |
+
|
207 |
+
class Log(nn.Module):
|
208 |
+
def forward(self, x, x_mask, reverse=False, **kwargs):
|
209 |
+
if not reverse:
|
210 |
+
y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
|
211 |
+
logdet = torch.sum(-y, [1, 2])
|
212 |
+
return y, logdet
|
213 |
+
else:
|
214 |
+
x = torch.exp(x) * x_mask
|
215 |
+
return x
|
216 |
+
|
217 |
+
class Flip(nn.Module):
|
218 |
+
def forward(self, x, *args, reverse=False, **kwargs):
|
219 |
+
x = torch.flip(x, [1])
|
220 |
+
if reverse: return x
|
221 |
+
logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
|
222 |
+
return x, logdet
|
223 |
+
|
224 |
+
class ElementwiseAffine(nn.Module):
|
225 |
+
def __init__(self, channels):
|
226 |
+
super().__init__()
|
227 |
+
self.channels = channels
|
228 |
+
self.m = nn.Parameter(torch.zeros(channels, 1))
|
229 |
+
self.logs = nn.Parameter(torch.zeros(channels, 1))
|
230 |
+
|
231 |
+
def forward(self, x, x_mask, reverse=False, **kwargs):
|
232 |
+
if not reverse:
|
233 |
+
y = self.m + torch.exp(self.logs) * x
|
234 |
+
y = y * x_mask
|
235 |
+
logdet = torch.sum(self.logs * x_mask, [1, 2])
|
236 |
+
return y, logdet
|
237 |
+
else:
|
238 |
+
x = (x - self.m) * torch.exp(-self.logs) * x_mask
|
239 |
+
return x
|
240 |
+
|
241 |
+
class ResidualCouplingLayer(nn.Module):
|
242 |
+
def __init__(self, channels, hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=0, gin_channels=0, mean_only=False):
|
243 |
+
assert channels % 2 == 0, "channels should be divisible by 2"
|
244 |
+
super().__init__()
|
245 |
+
self.channels = channels
|
246 |
+
self.hidden_channels = hidden_channels
|
247 |
+
self.kernel_size = kernel_size
|
248 |
+
self.dilation_rate = dilation_rate
|
249 |
+
self.n_layers = n_layers
|
250 |
+
self.half_channels = channels // 2
|
251 |
+
self.mean_only = mean_only
|
252 |
+
|
253 |
+
self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
|
254 |
+
self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
|
255 |
+
self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
|
256 |
+
self.post.weight.data.zero_()
|
257 |
+
self.post.bias.data.zero_()
|
258 |
+
|
259 |
+
def forward(self, x, x_mask, g=None, reverse=False):
|
260 |
+
x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
|
261 |
+
h = self.pre(x0) * x_mask
|
262 |
+
h = self.enc(h, x_mask, g=g)
|
263 |
+
stats = self.post(h) * x_mask
|
264 |
+
if not self.mean_only: m, logs = torch.split(stats, [self.half_channels] * 2, 1)
|
265 |
+
else:
|
266 |
+
m = stats
|
267 |
+
logs = torch.zeros_like(m)
|
268 |
+
|
269 |
+
if not reverse:
|
270 |
+
x1 = m + x1 * torch.exp(logs) * x_mask
|
271 |
+
x = torch.cat([x0, x1], 1)
|
272 |
+
logdet = torch.sum(logs, [1, 2])
|
273 |
+
return x, logdet
|
274 |
+
else:
|
275 |
+
x1 = (x1 - m) * torch.exp(-logs) * x_mask
|
276 |
+
x = torch.cat([x0, x1], 1)
|
277 |
+
return x
|
278 |
+
|
279 |
+
def remove_weight_norm(self): self.enc.remove_weight_norm()
|
280 |
+
|
281 |
+
class ConvFlow(nn.Module):
|
282 |
+
def __init__(self, in_channels, filter_channels, kernel_size, n_layers, num_bins=10, tail_bound=5.0):
|
283 |
+
super().__init__()
|
284 |
+
self.in_channels = in_channels
|
285 |
+
self.filter_channels = filter_channels
|
286 |
+
self.kernel_size = kernel_size
|
287 |
+
self.n_layers = n_layers
|
288 |
+
self.num_bins = num_bins
|
289 |
+
self.tail_bound = tail_bound
|
290 |
+
self.half_channels = in_channels // 2
|
291 |
+
|
292 |
+
self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
|
293 |
+
self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.0)
|
294 |
+
self.proj = nn.Conv1d(filter_channels, self.half_channels * (num_bins * 3 - 1), 1)
|
295 |
+
self.proj.weight.data.zero_()
|
296 |
+
self.proj.bias.data.zero_()
|
297 |
+
|
298 |
+
def forward(self, x, x_mask, g=None, reverse=False):
|
299 |
+
x0, x1 = torch.split(x, [self.half_channels] * 2, 1)
|
300 |
+
h = self.pre(x0)
|
301 |
+
h = self.convs(h, x_mask, g=g)
|
302 |
+
h = self.proj(h) * x_mask
|
303 |
+
|
304 |
+
b, c, t = x0.shape
|
305 |
+
h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2)
|
306 |
+
|
307 |
+
unnormalized_widths = h[..., : self.num_bins] / math.sqrt(self.filter_channels)
|
308 |
+
unnormalized_heights = h[..., self.num_bins : 2 * self.num_bins] / math.sqrt(self.filter_channels)
|
309 |
+
unnormalized_derivatives = h[..., 2 * self.num_bins :]
|
310 |
+
|
311 |
+
x1, logabsdet = piecewise_rational_quadratic_transform(x1, unnormalized_widths, unnormalized_heights, unnormalized_derivatives, inverse=reverse, tails="linear", tail_bound=self.tail_bound)
|
312 |
+
|
313 |
+
x = torch.cat([x0, x1], 1) * x_mask
|
314 |
+
logdet = torch.sum(logabsdet * x_mask, [1, 2])
|
315 |
+
return (x, logdet) if not reverse else x
|
lib/infer_pack/onnx_inference.py
ADDED
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import onnxruntime
|
2 |
+
import librosa
|
3 |
+
import numpy as np
|
4 |
+
from lib.infer_pack.modules.F0Predictor.PMF0Predictor import PMF0Predictor
|
5 |
+
|
6 |
+
VEC_PATH = "pretrained/vec-768-layer-12.onnx"
|
7 |
+
MAX_LENGTH = 50.0
|
8 |
+
F0_MIN = 50
|
9 |
+
F0_MAX = 1100
|
10 |
+
F0_MEL_MIN = 1127 * np.log(1 + F0_MIN / 700)
|
11 |
+
F0_MEL_MAX = 1127 * np.log(1 + F0_MAX / 700)
|
12 |
+
RESAMPLING_RATE = 16000
|
13 |
+
|
14 |
+
class ContentVectorModel:
|
15 |
+
def __init__(self, vector_path=VEC_PATH):
|
16 |
+
providers = ["CPUExecutionProvider"]
|
17 |
+
self.model = onnxruntime.InferenceSession(vector_path, providers=providers)
|
18 |
+
|
19 |
+
def __call__(self, audio_wave): return self.process_audio(audio_wave)
|
20 |
+
|
21 |
+
def process_audio(self, audio_wave):
|
22 |
+
features = audio_wave.mean(-1) if audio_wave.ndim == 2 else audio_wave
|
23 |
+
features = np.expand_dims(np.expand_dims(features, 0), 0)
|
24 |
+
onnx_input = {self.model.get_inputs()[0].name: features}
|
25 |
+
logits = self.model.run(None, onnx_input)[0]
|
26 |
+
return logits.transpose(0, 2, 1)
|
27 |
+
|
28 |
+
|
29 |
+
class OnnxRVC:
|
30 |
+
def __init__(self, model_path, sampling_rate=40000, hop_size=512, vector_path=VEC_PATH):
|
31 |
+
self.vec_model = ContentVectorModel(f"{vector_path}.onnx")
|
32 |
+
providers = ["CPUExecutionProvider"]
|
33 |
+
self.model = onnxruntime.InferenceSession(model_path, providers=providers)
|
34 |
+
self.sampling_rate = sampling_rate
|
35 |
+
self.hop_size = hop_size
|
36 |
+
|
37 |
+
def forward(self, hubert, hubert_length, pitch, pitchf, ds, rnd):
|
38 |
+
onnx_input = {self.model.get_inputs()[0].name: hubert, self.model.get_inputs()[1].name: hubert_length, self.model.get_inputs()[2].name: pitch, self.model.get_inputs()[3].name: pitchf, self.model.get_inputs()[4].name: ds, self.model.get_inputs()[5].name: rnd}
|
39 |
+
return (self.model.run(None, onnx_input)[0] * 32767).astype(np.int16)
|
40 |
+
|
41 |
+
def inference(self, raw_path, sid, f0_method="pm", f0_up_key=0, pad_time=0.5, cr_threshold=0.02):
|
42 |
+
f0_predictor = PMF0Predictor(hop_length=self.hop_size, sampling_rate=self.sampling_rate, threshold=cr_threshold)
|
43 |
+
wav, sr = librosa.load(raw_path, sr=self.sampling_rate)
|
44 |
+
org_length = len(wav)
|
45 |
+
if org_length / sr > MAX_LENGTH: raise RuntimeError("Reached Max Length")
|
46 |
+
|
47 |
+
wav16k = librosa.resample(wav, orig_sr=sr, target_sr=RESAMPLING_RATE)
|
48 |
+
hubert = self.vec_model(wav16k)
|
49 |
+
hubert = np.repeat(hubert, 2, axis=2).transpose(0, 2, 1).astype(np.float32)
|
50 |
+
hubert_length = hubert.shape[1]
|
51 |
+
|
52 |
+
pitchf = f0_predictor.compute_f0(wav, hubert_length)
|
53 |
+
pitchf *= 2 ** (f0_up_key / 12)
|
54 |
+
pitch = pitchf.copy()
|
55 |
+
f0_mel = 1127 * np.log(1 + pitch / 700)
|
56 |
+
f0_mel = np.clip(f0_mel - F0_MEL_MIN, 0, None) * 254 / (F0_MEL_MAX - F0_MEL_MIN) + 1
|
57 |
+
pitch = np.rint(f0_mel).astype(np.int64)
|
58 |
+
|
59 |
+
pitchf = pitchf.reshape(1, -1).astype(np.float32)
|
60 |
+
pitch = pitch.reshape(1, -1)
|
61 |
+
ds = np.array([sid]).astype(np.int64)
|
62 |
+
rnd = np.random.randn(1, 192, hubert_length).astype(np.float32)
|
63 |
+
hubert_length = np.array([hubert_length]).astype(np.int64)
|
64 |
+
|
65 |
+
out_wav = self.forward(hubert, hubert_length, pitch, pitchf, ds, rnd).squeeze()
|
66 |
+
out_wav = np.pad(out_wav, (0, 2 * self.hop_size), "constant")
|
67 |
+
return out_wav[:org_length]
|
lib/infer_pack/transforms.py
ADDED
@@ -0,0 +1,115 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from torch.nn import functional as F
|
3 |
+
import numpy as np
|
4 |
+
|
5 |
+
DEFAULT_MIN_BIN_WIDTH = 1e-3
|
6 |
+
DEFAULT_MIN_BIN_HEIGHT = 1e-3
|
7 |
+
DEFAULT_MIN_DERIVATIVE = 1e-3
|
8 |
+
|
9 |
+
def piecewise_rational_quadratic_transform(inputs, unnormalized_widths, unnormalized_heights, unnormalized_derivatives, inverse=False, tails=None, tail_bound=1.0, min_bin_width=DEFAULT_MIN_BIN_WIDTH, min_bin_height=DEFAULT_MIN_BIN_HEIGHT, min_derivative=DEFAULT_MIN_DERIVATIVE):
|
10 |
+
if tails is None:
|
11 |
+
spline_fn = rational_quadratic_spline
|
12 |
+
spline_kwargs = {}
|
13 |
+
else:
|
14 |
+
spline_fn = unconstrained_rational_quadratic_spline
|
15 |
+
spline_kwargs = {"tails": tails, "tail_bound": tail_bound}
|
16 |
+
|
17 |
+
return spline_fn(inputs=inputs, unnormalized_widths=unnormalized_widths, unnormalized_heights=unnormalized_heights, unnormalized_derivatives=unnormalized_derivatives, inverse=inverse, min_bin_width=min_bin_width, min_bin_height=min_bin_height, min_derivative=min_derivative, **spline_kwargs)
|
18 |
+
|
19 |
+
def searchsorted(bin_locations, inputs, eps=1e-6):
|
20 |
+
bin_locations[..., -1] += eps
|
21 |
+
return torch.sum(inputs[..., None] >= bin_locations, dim=-1) - 1
|
22 |
+
|
23 |
+
def unconstrained_rational_quadratic_spline(inputs, unnormalized_widths, unnormalized_heights, unnormalized_derivatives, inverse=False, tails="linear", tail_bound=1.0, min_bin_width=DEFAULT_MIN_BIN_WIDTH, min_bin_height=DEFAULT_MIN_BIN_HEIGHT, min_derivative=DEFAULT_MIN_DERIVATIVE):
|
24 |
+
if tails != "linear": raise RuntimeError(f"{tails} tails are not implemented.")
|
25 |
+
|
26 |
+
unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
|
27 |
+
constant = np.log(np.exp(1 - min_derivative) - 1)
|
28 |
+
unnormalized_derivatives[..., 0] = constant
|
29 |
+
unnormalized_derivatives[..., -1] = constant
|
30 |
+
|
31 |
+
inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
|
32 |
+
outside_interval_mask = ~inside_interval_mask
|
33 |
+
outputs = torch.where(outside_interval_mask, inputs, torch.zeros_like(inputs))
|
34 |
+
logabsdet = torch.zeros_like(inputs)
|
35 |
+
|
36 |
+
inside_outputs, inside_logabsdet = rational_quadratic_spline(
|
37 |
+
inputs=inputs[inside_interval_mask],
|
38 |
+
unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
|
39 |
+
unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
|
40 |
+
unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
|
41 |
+
inverse=inverse, left=-tail_bound, right=tail_bound, bottom=-tail_bound, top=tail_bound,
|
42 |
+
min_bin_width=min_bin_width, min_bin_height=min_bin_height, min_derivative=min_derivative)
|
43 |
+
|
44 |
+
outputs[inside_interval_mask] = inside_outputs
|
45 |
+
logabsdet[inside_interval_mask] = inside_logabsdet
|
46 |
+
|
47 |
+
return outputs, logabsdet
|
48 |
+
|
49 |
+
def rational_quadratic_spline(inputs, unnormalized_widths, unnormalized_heights, unnormalized_derivatives, inverse=False, left=0.0, right=1.0, bottom=0.0, top=1.0, min_bin_width=DEFAULT_MIN_BIN_WIDTH, min_bin_height=DEFAULT_MIN_BIN_HEIGHT, min_derivative=DEFAULT_MIN_DERIVATIVE):
|
50 |
+
num_bins = unnormalized_widths.shape[-1]
|
51 |
+
|
52 |
+
if min_bin_width * num_bins > 1.0: raise ValueError("Minimal bin width too large for the number of bins")
|
53 |
+
if min_bin_height * num_bins > 1.0: raise ValueError("Minimal bin height too large for the number of bins")
|
54 |
+
|
55 |
+
widths, heights = compute_widths_and_heights(unnormalized_widths, unnormalized_heights, min_bin_width, min_bin_height, num_bins, left, right, bottom, top)
|
56 |
+
cumwidths, cumheights = widths.cumsum(dim=-1), heights.cumsum(dim=-1)
|
57 |
+
cumwidths[..., 0] = left
|
58 |
+
cumwidths[..., -1] = right
|
59 |
+
cumheights[..., 0] = bottom
|
60 |
+
cumheights[..., -1] = top
|
61 |
+
widths, heights = cumwidths[..., 1:] - cumwidths[..., :-1], cumheights[..., 1:] - cumheights[..., :-1]
|
62 |
+
|
63 |
+
derivatives = min_derivative + F.softplus(unnormalized_derivatives)
|
64 |
+
|
65 |
+
if inverse: bin_idx = searchsorted(cumheights, inputs)[..., None]
|
66 |
+
else: bin_idx = searchsorted(cumwidths, inputs)[..., None]
|
67 |
+
|
68 |
+
gather_args = (-1, bin_idx)
|
69 |
+
input_cumwidths, input_bin_widths, input_cumheights, input_delta, input_derivatives, input_derivatives_plus_one, input_heights = map(
|
70 |
+
lambda tensor: tensor.gather(*gather_args)[..., 0],
|
71 |
+
(cumwidths, widths, cumheights, heights / widths, derivatives, derivatives[..., 1:], heights))
|
72 |
+
|
73 |
+
if inverse: outputs, logabsdet = inverse_rational_quadratic_spline(inputs, input_cumheights, input_heights, input_derivatives, input_derivatives_plus_one, input_delta, input_bin_widths, input_cumwidths)
|
74 |
+
else: outputs, logabsdet = direct_rational_quadratic_spline(inputs, input_cumwidths, input_bin_widths, input_cumheights, input_heights, input_derivatives, input_derivatives_plus_one, input_delta)
|
75 |
+
|
76 |
+
return outputs, logabsdet
|
77 |
+
|
78 |
+
def compute_widths_and_heights(unnormalized_widths, unnormalized_heights, min_bin_width, min_bin_height, num_bins, left, right, bottom, top):
|
79 |
+
widths = F.softmax(unnormalized_widths, dim=-1)
|
80 |
+
widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
|
81 |
+
widths = (right - left) * widths + left
|
82 |
+
|
83 |
+
heights = F.softmax(unnormalized_heights, dim=-1)
|
84 |
+
heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
|
85 |
+
heights = (top - bottom) * heights + bottom
|
86 |
+
|
87 |
+
return widths, heights
|
88 |
+
|
89 |
+
def inverse_rational_quadratic_spline(inputs, input_cumheights, input_heights, input_derivatives, input_derivatives_plus_one, input_delta, input_bin_widths, input_cumwidths):
|
90 |
+
a = (inputs - input_cumheights) * (input_derivatives + input_derivatives_plus_one - 2 * input_delta) + input_heights * (input_delta - input_derivatives)
|
91 |
+
b = input_heights * input_derivatives - (inputs - input_cumheights) * (input_derivatives + input_derivatives_plus_one - 2 * input_delta)
|
92 |
+
c = -input_delta * (inputs - input_cumheights)
|
93 |
+
|
94 |
+
discriminant = b.pow(2) - 4 * a * c
|
95 |
+
assert (discriminant >= 0).all()
|
96 |
+
|
97 |
+
root = (2 * c) / (-b - torch.sqrt(discriminant))
|
98 |
+
outputs = root * input_bin_widths + input_cumwidths
|
99 |
+
theta_one_minus_theta = root * (1 - root)
|
100 |
+
denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta)* theta_one_minus_theta)
|
101 |
+
derivative_numerator = input_delta.pow(2) * (input_derivatives_plus_one * root.pow(2)+ 2 * input_delta * theta_one_minus_theta+ input_derivatives * (1 - root).pow(2))
|
102 |
+
logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
|
103 |
+
|
104 |
+
return outputs, -logabsdet
|
105 |
+
|
106 |
+
def direct_rational_quadratic_spline(inputs, input_cumwidths, input_bin_widths, input_cumheights, input_heights, input_derivatives, input_derivatives_plus_one, input_delta):
|
107 |
+
theta = (inputs - input_cumwidths) / input_bin_widths
|
108 |
+
theta_one_minus_theta = theta * (1 - theta)
|
109 |
+
numerator = input_heights * (input_delta * theta.pow(2) + input_derivatives * theta_one_minus_theta)
|
110 |
+
denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta) * theta_one_minus_theta)
|
111 |
+
outputs = input_cumheights + numerator / denominator
|
112 |
+
derivative_numerator = input_delta.pow(2) * (input_derivatives_plus_one * theta.pow(2) + 2 * input_delta * theta_one_minus_theta + input_derivatives * (1 - theta).pow(2))
|
113 |
+
logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
|
114 |
+
|
115 |
+
return outputs, logabsdet
|
models/FernandoAlonso/FernandoAlonso/added_IVF582_Flat_nprobe_1_fernando2_v2.index
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:9c5d70bed6e4bed0b54eaef9c1f2dd180fda94035287baafadb99a456b8cbd52
|
3 |
+
size 71801099
|
models/FernandoAlonso/FernandoAlonso/fernando2.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:3926c2925ebe47d0c7a8d5b947813f0866a08d2970a10cb69d64f1b6e2371bbc
|
3 |
+
size 55224197
|
models/HolasoyGerman/HolasoyGerman/HolasoyGerman.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:db9c812c5ecd6e7afa5b0bee0300741ba87731097c6449b64924d46c10644247
|
3 |
+
size 55226033
|
models/HolasoyGerman/HolasoyGerman/added_IVF3117_Flat_nprobe_1_HolasoyGerman_v2.index
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:a11850d919513a315e2a4ec41586a74c650f0ccaf3fc16d8aebff4e63ce07dc4
|
3 |
+
size 384023779
|
models/HolasoyGerman/model_info.json
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"HolasoyGerman": {
|
3 |
+
"title": "HolasoyGerman",
|
4 |
+
"model_path": "HolasoyGerman.pth",
|
5 |
+
"feature_retrieval_library": "added_IVF3117_Flat_nprobe_1_HolasoyGerman_v2.index",
|
6 |
+
"model_path_checksum": "74633371eb2c2a75292633ee8922fddc",
|
7 |
+
"index_path_checksum": "febffc21f866063b40900cd068d379b9"
|
8 |
+
}
|
9 |
+
}
|
models/Homer/Homer/HomerEsp.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:309244d07dd5f611f4bca2415a933a6bf499d1a803cd8df45e005d8256373b65
|
3 |
+
size 55027130
|
models/Homer/Homer/added_IVF360_Flat_nprobe_1_HomerEsp_v1.index
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:a3b3375a8e01675a1a46439af33ff4941f52eb9592b8784c154dcd55cf9eba03
|
3 |
+
size 14861971
|
models/Homer/model_info.json
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"Homer": {
|
3 |
+
"title": "Homer",
|
4 |
+
"model_path": "HomerEsp.pth",
|
5 |
+
"feature_retrieval_library": "added_IVF360_Flat_nprobe_1_HomerEsp_v1.index",
|
6 |
+
"model_path_checksum": "7df9a9cdf205fe13f5ea6d73d7e9fd41",
|
7 |
+
"index_path_checksum": "5d554ed74ef44a057d4f997dabe8c3ea"
|
8 |
+
}
|
9 |
+
}
|
models/Ibai/Ibai/Ibai.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:aa102e749f66ede51b3bc43a4cceaeeaac7abd26c9f20348a233d042349d0467
|
3 |
+
size 55227410
|
models/Ibai/Ibai/added_IVF4601_Flat_nprobe_1_Ibai_v2.index
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:177232ae8f7428fbc925719e11fcfd73a48b0919c3f94fd64eae6ff9f815c440
|
3 |
+
size 566889539
|
models/Ibai/model_info.json
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"Ibai": {
|
3 |
+
"title": "Ibai",
|
4 |
+
"model_path": "Ibai.pth",
|
5 |
+
"feature_retrieval_library": "added_IVF4601_Flat_nprobe_1_Ibai_v2.index",
|
6 |
+
"model_path_checksum": "098e96b7158edd654d0e269959ac9a20",
|
7 |
+
"index_path_checksum": "b16da2fa68a6341f8fc409a83db68dcf"
|
8 |
+
}
|
9 |
+
}
|
models/IlloJuan/IlloJuan/IlloJuan.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:80c3229a3b5ee888ffb9f8b85385de9705966b156ea4471619f25186318f55b4
|
3 |
+
size 55223738
|
models/IlloJuan/IlloJuan/added_IVF593_Flat_nprobe_1_IlloJuan_v2.index
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:e3bce6d0a1f554f5b5e38bb8f5d72e15a05c2383261f72e4fb5c06c2d714763f
|
3 |
+
size 73162459
|
models/IlloJuan/model_info.json
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"IlloJuan": {
|
3 |
+
"title": "IlloJuan",
|
4 |
+
"model_path": "IlloJuan.pth",
|
5 |
+
"feature_retrieval_library": "added_IVF593_Flat_nprobe_1_IlloJuan_v2.index",
|
6 |
+
"model_path_checksum": "06a20ffd281284950537ebdbd227b8a9",
|
7 |
+
"index_path_checksum": "fcae1daa14cc24d862ca09216a02595c"
|
8 |
+
}
|
9 |
+
}
|
models/Quevedo/Quevedo/added_IVF2301_Flat_nprobe_10.index
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:0fdfabf3c2a559c0c5649f9991e667157bb14440a49860dce02e06ec26a34846
|
3 |
+
size 95005027
|
models/Quevedo/Quevedo/quevedo.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:071f9b5e30ae0c219b47de25a9300b4097f5a7d8025946088ce5d505d9ae6227
|
3 |
+
size 55026095
|
models/Quevedo/model_info.json
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"Quevedo": {
|
3 |
+
"title": "Quevedo",
|
4 |
+
"model_path": "quevedo.pth",
|
5 |
+
"feature_retrieval_library": "added_IVF2301_Flat_nprobe_10.index",
|
6 |
+
"model_path_checksum": "6325e08a29a9cbfd0625dfe34be24a0c",
|
7 |
+
"index_path_checksum": "924db9553a0a70538a53e625fbadde40"
|
8 |
+
}
|
9 |
+
}
|
models/Shadoune666/Shadoune666/Shadoune666.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:1447dd914d38c33b377b913c5ebc5f9bddf5a71195e2bb9b400af9bde6735e91
|
3 |
+
size 55225115
|
models/Shadoune666/Shadoune666/added_IVF716_Flat_nprobe_1_Shadoune666_v2.index
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:bf519d09b556d5052f68a0be8628d10071e294de284291810b9ed88d2f8e5981
|
3 |
+
size 88328379
|
models/Shadoune666/model_info.json
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"Shadoune666": {
|
3 |
+
"title": "Shadoune666",
|
4 |
+
"model_path": "Shadoune666.pth",
|
5 |
+
"feature_retrieval_library": "added_IVF716_Flat_nprobe_1_Shadoune666_v2.index",
|
6 |
+
"model_path_checksum": "7bca067c9055104ebea2bf3f6c7bfde4",
|
7 |
+
"index_path_checksum": "b23d61f502025b4f9d967e4de1935d0b"
|
8 |
+
}
|
9 |
+
}
|
models/Spreen/Spreen/Spreen.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:4be95ea3bb27866c12fafd46ee9997165d025f50e35260b23e2c175c942f6cff
|
3 |
+
size 55216420
|
models/Spreen/Spreen/added_IVF4737_Flat_nprobe_1_Spreen_v2.index
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:669ca397e95bf51ff0e3f1aabe1ca60c83a5a6c85653ba0f1e550089c63b4305
|
3 |
+
size 583660139
|
models/Spreen/model_info.json
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"Spreen": {
|
3 |
+
"title": "Spreen",
|
4 |
+
"model_path": "Spreen.pth",
|
5 |
+
"feature_retrieval_library": "added_IVF4737_Flat_nprobe_1_Spreen_v2.index",
|
6 |
+
"model_path_checksum": "dbdb20d98928308c2d048444efb34b77",
|
7 |
+
"index_path_checksum": "213bbb9c387426fadc97c41e97f34b0a"
|
8 |
+
}
|
9 |
+
}
|
models/Totakeke/Totakeke/added_IVF256_Flat_nprobe_1_full_totakeke_v2.index
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:b88bd6d822fa8391fa2ca6dd6e0c6eac8ac031385519da3fd87d2a6e8f272151
|
3 |
+
size 31588619
|
models/Totakeke/Totakeke/full_totakeke_e150_s44400.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:a3fb774cbbb74ae6005a87e370fc01337d3a24f7c5bb8c0962b335e4bd69488b
|
3 |
+
size 57590901
|
models/Totakeke/model_info.json
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"Totakeke": {
|
3 |
+
"title": "Totakeke",
|
4 |
+
"model_path": "full_totakeke_e150_s44400.pth",
|
5 |
+
"feature_retrieval_library": "added_IVF256_Flat_nprobe_1_full_totakeke_v2.index",
|
6 |
+
"model_path_checksum": "ccb3a07fd141c6e918eca8a969d95c5d",
|
7 |
+
"index_path_checksum": "d4b2d2662efdc0bcb8f3610a686525c9"
|
8 |
+
}
|
9 |
+
}
|
models/Vegetta777/Vegetta777/Vegetta777.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f13483615d05c1b31a4eb0fbee35ac5dca9f9a48a36a0165b947c08303ce75e5
|
3 |
+
size 55218256
|
models/Vegetta777/Vegetta777/added_IVF2021_Flat_nprobe_1_Vegetta777_v2.index
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:dd713395db0957eb626a40d096ee29ec8bebe19fac3d9ce356a3e587920c71d8
|
3 |
+
size 248990419
|
models/Vegetta777/model_info.json
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"Vegetta777": {
|
3 |
+
"title": "Vegetta777",
|
4 |
+
"model_path": "Vegetta777.pth",
|
5 |
+
"feature_retrieval_library": "added_IVF2021_Flat_nprobe_1_Vegetta777_v2.index",
|
6 |
+
"model_path_checksum": "9bc80abfc506e98056e1db7b10000dea",
|
7 |
+
"index_path_checksum": "5a53e6f6a997e892e5fb26a7a6959550"
|
8 |
+
}
|
9 |
+
}
|
models/Villager/Villager/added_IVF81_Flat_nprobe_1_v2.index
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d8b8f5b445b1a1c3ee7155f3a121075399ca5b5a01032c459168b739c7c633e5
|
3 |
+
size 10013219
|
models/Villager/Villager/villager.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:c2d94a1084bc9d65cd60ab7fb229a83476e79f8b75ec89e7343d0b9352ea1739
|
3 |
+
size 55223738
|
models/Villager/model_info.json
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"Villager": {
|
3 |
+
"title": "Villager",
|
4 |
+
"model_path": "villager.pth",
|
5 |
+
"feature_retrieval_library": "added_IVF81_Flat_nprobe_1_v2.index",
|
6 |
+
"model_path_checksum": "587b2be8f6fa60f7533dddf1a007e2fa",
|
7 |
+
"index_path_checksum": "13ef69f0f2181f77d45c4d36d7364f2b"
|
8 |
+
}
|
9 |
+
}
|
models/WalterWhite/WalterWhite/added_IVF458_Flat_nprobe_1_ww2_v2.index
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:c2030573a1685644ec51b816bb93194d813a30aa7e066fe27b849dc1e80f7706
|
3 |
+
size 56527379
|