ORI-Muchim commited on
Commit
521134f
1 Parent(s): 6fd5ca2

Upload 11 files

Browse files
Files changed (11) hide show
  1. README.md +13 -0
  2. app.py +170 -0
  3. attentions.py +300 -0
  4. commons.py +172 -0
  5. export_model.py +13 -0
  6. mel_processing.py +101 -0
  7. models.py +540 -0
  8. modules.py +390 -0
  9. requirements.txt +25 -0
  10. transforms.py +193 -0
  11. utils.py +226 -0
README.md ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ title: ONFIRETTS
3
+ emoji: 🔥
4
+ colorFrom: red
5
+ colorTo: orange
6
+ sdk: gradio
7
+ sdk_version: 3.12.0
8
+ app_file: app.py
9
+ pinned: false
10
+ license: mit
11
+ ---
12
+
13
+ Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
app.py ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import json
2
+ import os
3
+ import re
4
+
5
+ import librosa
6
+ import numpy as np
7
+ import torch
8
+ from torch import no_grad, LongTensor
9
+ import commons
10
+ import utils
11
+ import gradio as gr
12
+ from models import SynthesizerTrn
13
+ from text import text_to_sequence, _clean_text
14
+ from mel_processing import spectrogram_torch
15
+
16
+ limitation = os.getenv("SYSTEM") == "spaces" # limit text and audio length in huggingface spaces
17
+
18
+
19
+ def get_text(text, hps, is_phoneme):
20
+ text_norm = text_to_sequence(text, hps.symbols, [] if is_phoneme else hps.data.text_cleaners)
21
+ if hps.data.add_blank:
22
+ text_norm = commons.intersperse(text_norm, 0)
23
+ text_norm = LongTensor(text_norm)
24
+ return text_norm
25
+
26
+
27
+ def create_tts_fn(model, hps, speaker_ids):
28
+ def tts_fn(text, speaker, speed, is_phoneme):
29
+ if limitation:
30
+ text_len = len(text)
31
+ max_len = 500
32
+ if is_phoneme:
33
+ max_len *= 3
34
+ else:
35
+ if len(hps.data.text_cleaners) > 0 and hps.data.text_cleaners[0] == "zh_ja_mixture_cleaners":
36
+ text_len = len(re.sub("(\[ZH\]|\[JA\])", "", text))
37
+ if text_len > max_len:
38
+ return "Error: Text is too long", None
39
+
40
+ speaker_id = speaker_ids[speaker]
41
+ stn_tst = get_text(text, hps, is_phoneme)
42
+ with no_grad():
43
+ x_tst = stn_tst.unsqueeze(0)
44
+ x_tst_lengths = LongTensor([stn_tst.size(0)])
45
+ sid = LongTensor([speaker_id])
46
+ audio = model.infer(x_tst, x_tst_lengths, sid=sid, noise_scale=.667, noise_scale_w=0.8,
47
+ length_scale=1.0 / speed)[0][0, 0].data.cpu().float().numpy()
48
+ del stn_tst, x_tst, x_tst_lengths, sid
49
+ return "Success", (hps.data.sampling_rate, audio)
50
+
51
+ return tts_fn
52
+
53
+
54
+
55
+
56
+
57
+ def create_to_phoneme_fn(hps):
58
+ def to_phoneme_fn(text):
59
+ return _clean_text(text, hps.data.text_cleaners) if text != "" else ""
60
+
61
+ return to_phoneme_fn
62
+
63
+
64
+ css = """
65
+ #advanced-btn {
66
+ color: white;
67
+ border-color: black;
68
+ background: black;
69
+ font-size: .7rem !important;
70
+ line-height: 19px;
71
+ margin-top: 24px;
72
+ margin-bottom: 12px;
73
+ padding: 2px 8px;
74
+ border-radius: 14px !important;
75
+ }
76
+ #advanced-options {
77
+ display: none;
78
+ margin-bottom: 20px;
79
+ }
80
+ """
81
+
82
+ if __name__ == '__main__':
83
+ models_tts = []
84
+ models_vc = []
85
+ models_soft_vc = []
86
+ # {"title": "ハミダシクリエイティブ", "lang": "日本語 (Japanese)", "example": "こんにちは。", "type": "vits"}
87
+ name = 'ONFIRETTS'
88
+ lang = '한국어 (Korean)'
89
+ example = '야 이 미친새끼야!'
90
+ config_path = f"saved_model/config.json"
91
+ model_path = f"saved_model/model.pth"
92
+ cover_path = f"saved_model/cover.png"
93
+ hps = utils.get_hparams_from_file(config_path)
94
+ model = SynthesizerTrn(
95
+ len(hps.symbols),
96
+ hps.data.filter_length // 2 + 1,
97
+ hps.train.segment_size // hps.data.hop_length,
98
+ n_speakers=hps.data.n_speakers,
99
+ **hps.model)
100
+ utils.load_checkpoint(model_path, model, None)
101
+ model.eval()
102
+ speaker_ids = [sid for sid, name in enumerate(hps.speakers) if name != "None"]
103
+ speakers = [name for sid, name in enumerate(hps.speakers) if name != "None"]
104
+
105
+ t = 'vits'
106
+ models_tts.append((name, cover_path, speakers, lang, example,
107
+ hps.symbols, create_tts_fn(model, hps, speaker_ids),
108
+ create_to_phoneme_fn(hps)))
109
+
110
+
111
+ app = gr.Blocks(css=css)
112
+
113
+ with app:
114
+ gr.Markdown("# ONFIRETTS Using VITS Model\n\n"
115
+ "![visitor badge](https://visitor-badge.glitch.me/badge?page_id=ORI-Muchim.ONFIRETTS)\n\n")
116
+ with gr.Tabs():
117
+ with gr.TabItem("TTS"):
118
+ with gr.Tabs():
119
+ for i, (name, cover_path, speakers, lang, example, symbols, tts_fn,
120
+ to_phoneme_fn) in enumerate(models_tts):
121
+ with gr.TabItem(f"ONFIRE):
122
+ with gr.Column():
123
+ gr.Markdown(f"## {name}\n\n"
124
+ f"![cover](file/{cover_path})\n\n"
125
+ f"lang: {lang}")
126
+ tts_input1 = gr.TextArea(label="Text (500 words limitation)", value=example,
127
+ elem_id=f"tts-input{i}")
128
+ tts_input2 = gr.Dropdown(label="Speaker", choices=speakers,
129
+ type="index", value=speakers[0])
130
+ tts_input3 = gr.Slider(label="Speed", value=1, minimum=0.1, maximum=2, step=0.1)
131
+ with gr.Accordion(label="Advanced Options", open=False):
132
+ phoneme_input = gr.Checkbox(value=False, label="Phoneme input")
133
+ to_phoneme_btn = gr.Button("Covert text to phoneme")
134
+ phoneme_list = gr.Dataset(label="Phoneme list", components=[tts_input1],
135
+ samples=[[x] for x in symbols],
136
+ elem_id=f"phoneme-list{i}")
137
+ phoneme_list_json = gr.Json(value=symbols, visible=False)
138
+ tts_submit = gr.Button("Generate", variant="primary")
139
+ tts_output1 = gr.Textbox(label="Output Message")
140
+ tts_output2 = gr.Audio(label="Output Audio")
141
+ tts_submit.click(tts_fn, [tts_input1, tts_input2, tts_input3, phoneme_input],
142
+ [tts_output1, tts_output2])
143
+ to_phoneme_btn.click(to_phoneme_fn, [tts_input1], [tts_input1])
144
+ phoneme_list.click(None, [phoneme_list, phoneme_list_json], [],
145
+ _js=f"""
146
+ (i,phonemes) => {{
147
+ let root = document.querySelector("body > gradio-app");
148
+ if (root.shadowRoot != null)
149
+ root = root.shadowRoot;
150
+ let text_input = root.querySelector("#tts-input{i}").querySelector("textarea");
151
+ let startPos = text_input.selectionStart;
152
+ let endPos = text_input.selectionEnd;
153
+ let oldTxt = text_input.value;
154
+ let result = oldTxt.substring(0, startPos) + phonemes[i] + oldTxt.substring(endPos);
155
+ text_input.value = result;
156
+ let x = window.scrollX, y = window.scrollY;
157
+ text_input.focus();
158
+ text_input.selectionStart = startPos + phonemes[i].length;
159
+ text_input.selectionEnd = startPos + phonemes[i].length;
160
+ text_input.blur();
161
+ window.scrollTo(x, y);
162
+ return [];
163
+ }}""")
164
+
165
+
166
+ gr.Markdown(
167
+ "Reference \n\n"
168
+ "- [https://huggingface.co/spaces/kdrkdrkdr/ProsekaTTS](https://huggingface.co/spaces/kdrkdrkdr/ProsekaTTS)\n\n"
169
+ )
170
+ app.queue(concurrency_count=3).launch(show_api=False)
attentions.py ADDED
@@ -0,0 +1,300 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import torch
3
+ from torch import nn
4
+ from torch.nn import functional as F
5
+
6
+ import commons
7
+ from modules import LayerNorm
8
+
9
+
10
+ class Encoder(nn.Module):
11
+ def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., window_size=4, **kwargs):
12
+ super().__init__()
13
+ self.hidden_channels = hidden_channels
14
+ self.filter_channels = filter_channels
15
+ self.n_heads = n_heads
16
+ self.n_layers = n_layers
17
+ self.kernel_size = kernel_size
18
+ self.p_dropout = p_dropout
19
+ self.window_size = window_size
20
+
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 i 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
+
40
+ y = self.ffn_layers[i](x, x_mask)
41
+ y = self.drop(y)
42
+ x = self.norm_layers_2[i](x + y)
43
+ x = x * x_mask
44
+ return x
45
+
46
+
47
+ class Decoder(nn.Module):
48
+ def __init__(self, hidden_channels, filter_channels, n_heads, n_layers, kernel_size=1, p_dropout=0., proximal_bias=False, proximal_init=True, **kwargs):
49
+ super().__init__()
50
+ self.hidden_channels = hidden_channels
51
+ self.filter_channels = filter_channels
52
+ self.n_heads = n_heads
53
+ self.n_layers = n_layers
54
+ self.kernel_size = kernel_size
55
+ self.p_dropout = p_dropout
56
+ self.proximal_bias = proximal_bias
57
+ self.proximal_init = proximal_init
58
+
59
+ self.drop = nn.Dropout(p_dropout)
60
+ self.self_attn_layers = nn.ModuleList()
61
+ self.norm_layers_0 = nn.ModuleList()
62
+ self.encdec_attn_layers = nn.ModuleList()
63
+ self.norm_layers_1 = nn.ModuleList()
64
+ self.ffn_layers = nn.ModuleList()
65
+ self.norm_layers_2 = nn.ModuleList()
66
+ for i in range(self.n_layers):
67
+ self.self_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout, proximal_bias=proximal_bias, proximal_init=proximal_init))
68
+ self.norm_layers_0.append(LayerNorm(hidden_channels))
69
+ self.encdec_attn_layers.append(MultiHeadAttention(hidden_channels, hidden_channels, n_heads, p_dropout=p_dropout))
70
+ self.norm_layers_1.append(LayerNorm(hidden_channels))
71
+ self.ffn_layers.append(FFN(hidden_channels, hidden_channels, filter_channels, kernel_size, p_dropout=p_dropout, causal=True))
72
+ self.norm_layers_2.append(LayerNorm(hidden_channels))
73
+
74
+ def forward(self, x, x_mask, h, h_mask):
75
+ """
76
+ x: decoder input
77
+ h: encoder output
78
+ """
79
+ self_attn_mask = commons.subsequent_mask(x_mask.size(2)).to(device=x.device, dtype=x.dtype)
80
+ encdec_attn_mask = h_mask.unsqueeze(2) * x_mask.unsqueeze(-1)
81
+ x = x * x_mask
82
+ for i in range(self.n_layers):
83
+ y = self.self_attn_layers[i](x, x, self_attn_mask)
84
+ y = self.drop(y)
85
+ x = self.norm_layers_0[i](x + y)
86
+
87
+ y = self.encdec_attn_layers[i](x, h, encdec_attn_mask)
88
+ y = self.drop(y)
89
+ x = self.norm_layers_1[i](x + y)
90
+
91
+ y = self.ffn_layers[i](x, x_mask)
92
+ y = self.drop(y)
93
+ x = self.norm_layers_2[i](x + y)
94
+ x = x * x_mask
95
+ return x
96
+
97
+
98
+ class MultiHeadAttention(nn.Module):
99
+ def __init__(self, channels, out_channels, n_heads, p_dropout=0., window_size=None, heads_share=True, block_length=None, proximal_bias=False, proximal_init=False):
100
+ super().__init__()
101
+ assert channels % n_heads == 0
102
+
103
+ self.channels = channels
104
+ self.out_channels = out_channels
105
+ self.n_heads = n_heads
106
+ self.p_dropout = p_dropout
107
+ self.window_size = window_size
108
+ self.heads_share = heads_share
109
+ self.block_length = block_length
110
+ self.proximal_bias = proximal_bias
111
+ self.proximal_init = proximal_init
112
+ self.attn = None
113
+
114
+ self.k_channels = channels // n_heads
115
+ self.conv_q = nn.Conv1d(channels, channels, 1)
116
+ self.conv_k = nn.Conv1d(channels, channels, 1)
117
+ self.conv_v = nn.Conv1d(channels, channels, 1)
118
+ self.conv_o = nn.Conv1d(channels, out_channels, 1)
119
+ self.drop = nn.Dropout(p_dropout)
120
+
121
+ if window_size is not None:
122
+ n_heads_rel = 1 if heads_share else n_heads
123
+ rel_stddev = self.k_channels**-0.5
124
+ self.emb_rel_k = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
125
+ self.emb_rel_v = nn.Parameter(torch.randn(n_heads_rel, window_size * 2 + 1, self.k_channels) * rel_stddev)
126
+
127
+ nn.init.xavier_uniform_(self.conv_q.weight)
128
+ nn.init.xavier_uniform_(self.conv_k.weight)
129
+ nn.init.xavier_uniform_(self.conv_v.weight)
130
+ if proximal_init:
131
+ with torch.no_grad():
132
+ self.conv_k.weight.copy_(self.conv_q.weight)
133
+ self.conv_k.bias.copy_(self.conv_q.bias)
134
+
135
+ def forward(self, x, c, attn_mask=None):
136
+ q = self.conv_q(x)
137
+ k = self.conv_k(c)
138
+ v = self.conv_v(c)
139
+
140
+ x, self.attn = self.attention(q, k, v, mask=attn_mask)
141
+
142
+ x = self.conv_o(x)
143
+ return x
144
+
145
+ def attention(self, query, key, value, mask=None):
146
+ # reshape [b, d, t] -> [b, n_h, t, d_k]
147
+ b, d, t_s, t_t = (*key.size(), query.size(2))
148
+ query = query.view(b, self.n_heads, self.k_channels, t_t).transpose(2, 3)
149
+ key = key.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
150
+ value = value.view(b, self.n_heads, self.k_channels, t_s).transpose(2, 3)
151
+
152
+ scores = torch.matmul(query / math.sqrt(self.k_channels), key.transpose(-2, -1))
153
+ if self.window_size is not None:
154
+ assert t_s == t_t, "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 = torch.ones_like(scores).triu(-self.block_length).tril(self.block_length)
167
+ scores = scores.masked_fill(block_mask == 0, -1e4)
168
+ p_attn = F.softmax(scores, dim=-1) # [b, n_h, t_t, t_s]
169
+ p_attn = self.drop(p_attn)
170
+ output = torch.matmul(p_attn, value)
171
+ if self.window_size is not None:
172
+ relative_weights = self._absolute_position_to_relative_position(p_attn)
173
+ value_relative_embeddings = self._get_relative_embeddings(self.emb_rel_v, t_s)
174
+ output = output + self._matmul_with_relative_values(relative_weights, value_relative_embeddings)
175
+ output = output.transpose(2, 3).contiguous().view(b, d, t_t) # [b, n_h, t_t, d_k] -> [b, d, t_t]
176
+ return output, p_attn
177
+
178
+ def _matmul_with_relative_values(self, x, y):
179
+ """
180
+ x: [b, h, l, m]
181
+ y: [h or 1, m, d]
182
+ ret: [b, h, l, d]
183
+ """
184
+ ret = torch.matmul(x, y.unsqueeze(0))
185
+ return ret
186
+
187
+ def _matmul_with_relative_keys(self, x, y):
188
+ """
189
+ x: [b, h, l, d]
190
+ y: [h or 1, m, d]
191
+ ret: [b, h, l, m]
192
+ """
193
+ ret = torch.matmul(x, y.unsqueeze(0).transpose(-2, -1))
194
+ return ret
195
+
196
+ def _get_relative_embeddings(self, relative_embeddings, length):
197
+ max_relative_position = 2 * self.window_size + 1
198
+ # Pad first before slice to avoid using cond ops.
199
+ pad_length = max(length - (self.window_size + 1), 0)
200
+ slice_start_position = max((self.window_size + 1) - length, 0)
201
+ slice_end_position = slice_start_position + 2 * length - 1
202
+ if pad_length > 0:
203
+ padded_relative_embeddings = F.pad(
204
+ relative_embeddings,
205
+ commons.convert_pad_shape([[0, 0], [pad_length, pad_length], [0, 0]]))
206
+ else:
207
+ padded_relative_embeddings = relative_embeddings
208
+ used_relative_embeddings = padded_relative_embeddings[:,slice_start_position:slice_end_position]
209
+ return used_relative_embeddings
210
+
211
+ def _relative_position_to_absolute_position(self, x):
212
+ """
213
+ x: [b, h, l, 2*l-1]
214
+ ret: [b, h, l, l]
215
+ """
216
+ batch, heads, length, _ = x.size()
217
+ # Concat columns of pad to shift from relative to absolute indexing.
218
+ x = F.pad(x, commons.convert_pad_shape([[0,0],[0,0],[0,0],[0,1]]))
219
+
220
+ # Concat extra elements so to add up to shape (len+1, 2*len-1).
221
+ x_flat = x.view([batch, heads, length * 2 * length])
222
+ x_flat = F.pad(x_flat, commons.convert_pad_shape([[0,0],[0,0],[0,length-1]]))
223
+
224
+ # Reshape and slice out the padded elements.
225
+ x_final = x_flat.view([batch, heads, length+1, 2*length-1])[:, :, :length, length-1:]
226
+ return x_final
227
+
228
+ def _absolute_position_to_relative_position(self, x):
229
+ """
230
+ x: [b, h, l, l]
231
+ ret: [b, h, l, 2*l-1]
232
+ """
233
+ batch, heads, length, _ = x.size()
234
+ # padd along column
235
+ x = F.pad(x, commons.convert_pad_shape([[0, 0], [0, 0], [0, 0], [0, length-1]]))
236
+ x_flat = x.view([batch, heads, length**2 + length*(length -1)])
237
+ # add 0's in the beginning that will skew the elements after reshape
238
+ x_flat = F.pad(x_flat, commons.convert_pad_shape([[0, 0], [0, 0], [length, 0]]))
239
+ x_final = x_flat.view([batch, heads, length, 2*length])[:,:,:,1:]
240
+ return x_final
241
+
242
+ def _attention_bias_proximal(self, length):
243
+ """Bias for self-attention to encourage attention to close positions.
244
+ Args:
245
+ length: an integer scalar.
246
+ Returns:
247
+ a Tensor with shape [1, 1, length, length]
248
+ """
249
+ r = torch.arange(length, dtype=torch.float32)
250
+ diff = torch.unsqueeze(r, 0) - torch.unsqueeze(r, 1)
251
+ return torch.unsqueeze(torch.unsqueeze(-torch.log1p(torch.abs(diff)), 0), 0)
252
+
253
+
254
+ class FFN(nn.Module):
255
+ def __init__(self, in_channels, out_channels, filter_channels, kernel_size, p_dropout=0., activation=None, causal=False):
256
+ super().__init__()
257
+ self.in_channels = in_channels
258
+ self.out_channels = out_channels
259
+ self.filter_channels = filter_channels
260
+ self.kernel_size = kernel_size
261
+ self.p_dropout = p_dropout
262
+ self.activation = activation
263
+ self.causal = causal
264
+
265
+ if causal:
266
+ self.padding = self._causal_padding
267
+ else:
268
+ self.padding = self._same_padding
269
+
270
+ self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size)
271
+ self.conv_2 = nn.Conv1d(filter_channels, out_channels, kernel_size)
272
+ self.drop = nn.Dropout(p_dropout)
273
+
274
+ def forward(self, x, x_mask):
275
+ x = self.conv_1(self.padding(x * x_mask))
276
+ if self.activation == "gelu":
277
+ x = x * torch.sigmoid(1.702 * x)
278
+ else:
279
+ x = torch.relu(x)
280
+ x = self.drop(x)
281
+ x = self.conv_2(self.padding(x * x_mask))
282
+ return x * x_mask
283
+
284
+ def _causal_padding(self, x):
285
+ if self.kernel_size == 1:
286
+ return x
287
+ pad_l = self.kernel_size - 1
288
+ pad_r = 0
289
+ padding = [[0, 0], [0, 0], [pad_l, pad_r]]
290
+ x = F.pad(x, commons.convert_pad_shape(padding))
291
+ return x
292
+
293
+ def _same_padding(self, x):
294
+ if self.kernel_size == 1:
295
+ return x
296
+ pad_l = (self.kernel_size - 1) // 2
297
+ pad_r = self.kernel_size // 2
298
+ padding = [[0, 0], [0, 0], [pad_l, pad_r]]
299
+ x = F.pad(x, commons.convert_pad_shape(padding))
300
+ return x
commons.py ADDED
@@ -0,0 +1,172 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import torch
3
+ from torch.nn import functional as F
4
+ import torch.jit
5
+
6
+
7
+ def script_method(fn, _rcb=None):
8
+ return fn
9
+
10
+
11
+ def script(obj, optimize=True, _frames_up=0, _rcb=None):
12
+ return obj
13
+
14
+
15
+ torch.jit.script_method = script_method
16
+ torch.jit.script = script
17
+
18
+
19
+ def init_weights(m, mean=0.0, std=0.01):
20
+ classname = m.__class__.__name__
21
+ if classname.find("Conv") != -1:
22
+ m.weight.data.normal_(mean, std)
23
+
24
+
25
+ def get_padding(kernel_size, dilation=1):
26
+ return int((kernel_size*dilation - dilation)/2)
27
+
28
+
29
+ def convert_pad_shape(pad_shape):
30
+ l = pad_shape[::-1]
31
+ pad_shape = [item for sublist in l for item in sublist]
32
+ return pad_shape
33
+
34
+
35
+ def intersperse(lst, item):
36
+ result = [item] * (len(lst) * 2 + 1)
37
+ result[1::2] = lst
38
+ return result
39
+
40
+
41
+ def kl_divergence(m_p, logs_p, m_q, logs_q):
42
+ """KL(P||Q)"""
43
+ kl = (logs_q - logs_p) - 0.5
44
+ kl += 0.5 * (torch.exp(2. * logs_p) + ((m_p - m_q)**2)) * torch.exp(-2. * logs_q)
45
+ return kl
46
+
47
+
48
+ def rand_gumbel(shape):
49
+ """Sample from the Gumbel distribution, protect from overflows."""
50
+ uniform_samples = torch.rand(shape) * 0.99998 + 0.00001
51
+ return -torch.log(-torch.log(uniform_samples))
52
+
53
+
54
+ def rand_gumbel_like(x):
55
+ g = rand_gumbel(x.size()).to(dtype=x.dtype, device=x.device)
56
+ return g
57
+
58
+
59
+ def slice_segments(x, ids_str, segment_size=4):
60
+ ret = torch.zeros_like(x[:, :, :segment_size])
61
+ for i in range(x.size(0)):
62
+ idx_str = ids_str[i]
63
+ idx_end = idx_str + segment_size
64
+ ret[i] = x[i, :, idx_str:idx_end]
65
+ return ret
66
+
67
+
68
+ def rand_slice_segments(x, x_lengths=None, segment_size=4):
69
+ b, d, t = x.size()
70
+ if x_lengths is None:
71
+ x_lengths = t
72
+ ids_str_max = x_lengths - segment_size + 1
73
+ ids_str = (torch.rand([b]).to(device=x.device) * ids_str_max).to(dtype=torch.long)
74
+ ret = slice_segments(x, ids_str, segment_size)
75
+ return ret, ids_str
76
+
77
+
78
+ def get_timing_signal_1d(
79
+ length, channels, min_timescale=1.0, max_timescale=1.0e4):
80
+ position = torch.arange(length, dtype=torch.float)
81
+ num_timescales = channels // 2
82
+ log_timescale_increment = (
83
+ math.log(float(max_timescale) / float(min_timescale)) /
84
+ (num_timescales - 1))
85
+ inv_timescales = min_timescale * torch.exp(
86
+ torch.arange(num_timescales, dtype=torch.float) * -log_timescale_increment)
87
+ scaled_time = position.unsqueeze(0) * inv_timescales.unsqueeze(1)
88
+ signal = torch.cat([torch.sin(scaled_time), torch.cos(scaled_time)], 0)
89
+ signal = F.pad(signal, [0, 0, 0, channels % 2])
90
+ signal = signal.view(1, channels, length)
91
+ return signal
92
+
93
+
94
+ def add_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4):
95
+ b, channels, length = x.size()
96
+ signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
97
+ return x + signal.to(dtype=x.dtype, device=x.device)
98
+
99
+
100
+ def cat_timing_signal_1d(x, min_timescale=1.0, max_timescale=1.0e4, axis=1):
101
+ b, channels, length = x.size()
102
+ signal = get_timing_signal_1d(length, channels, min_timescale, max_timescale)
103
+ return torch.cat([x, signal.to(dtype=x.dtype, device=x.device)], axis)
104
+
105
+
106
+ def subsequent_mask(length):
107
+ mask = torch.tril(torch.ones(length, length)).unsqueeze(0).unsqueeze(0)
108
+ return mask
109
+
110
+
111
+ @torch.jit.script
112
+ def fused_add_tanh_sigmoid_multiply(input_a, input_b, n_channels):
113
+ n_channels_int = n_channels[0]
114
+ in_act = input_a + input_b
115
+ t_act = torch.tanh(in_act[:, :n_channels_int, :])
116
+ s_act = torch.sigmoid(in_act[:, n_channels_int:, :])
117
+ acts = t_act * s_act
118
+ return acts
119
+
120
+
121
+ def convert_pad_shape(pad_shape):
122
+ l = pad_shape[::-1]
123
+ pad_shape = [item for sublist in l for item in sublist]
124
+ return pad_shape
125
+
126
+
127
+ def shift_1d(x):
128
+ x = F.pad(x, convert_pad_shape([[0, 0], [0, 0], [1, 0]]))[:, :, :-1]
129
+ return x
130
+
131
+
132
+ def sequence_mask(length, max_length=None):
133
+ if max_length is None:
134
+ max_length = length.max()
135
+ x = torch.arange(max_length, dtype=length.dtype, device=length.device)
136
+ return x.unsqueeze(0) < length.unsqueeze(1)
137
+
138
+
139
+ def generate_path(duration, mask):
140
+ """
141
+ duration: [b, 1, t_x]
142
+ mask: [b, 1, t_y, t_x]
143
+ """
144
+ device = duration.device
145
+
146
+ b, _, t_y, t_x = mask.shape
147
+ cum_duration = torch.cumsum(duration, -1)
148
+
149
+ cum_duration_flat = cum_duration.view(b * t_x)
150
+ path = sequence_mask(cum_duration_flat, t_y).to(mask.dtype)
151
+ path = path.view(b, t_x, t_y)
152
+ path = path - F.pad(path, convert_pad_shape([[0, 0], [1, 0], [0, 0]]))[:, :-1]
153
+ path = path.unsqueeze(1).transpose(2,3) * mask
154
+ return path
155
+
156
+
157
+ def clip_grad_value_(parameters, clip_value, norm_type=2):
158
+ if isinstance(parameters, torch.Tensor):
159
+ parameters = [parameters]
160
+ parameters = list(filter(lambda p: p.grad is not None, parameters))
161
+ norm_type = float(norm_type)
162
+ if clip_value is not None:
163
+ clip_value = float(clip_value)
164
+
165
+ total_norm = 0
166
+ for p in parameters:
167
+ param_norm = p.grad.data.norm(norm_type)
168
+ total_norm += param_norm.item() ** norm_type
169
+ if clip_value is not None:
170
+ p.grad.data.clamp_(min=-clip_value, max=clip_value)
171
+ total_norm = total_norm ** (1. / norm_type)
172
+ return total_norm
export_model.py ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+
3
+ if __name__ == '__main__':
4
+ model_path = "saved_model/11/model.pth"
5
+ output_path = "saved_model/11/model1.pth"
6
+ checkpoint_dict = torch.load(model_path, map_location='cpu')
7
+ checkpoint_dict_new = {}
8
+ for k, v in checkpoint_dict.items():
9
+ if k == "optimizer":
10
+ print("remove optimizer")
11
+ continue
12
+ checkpoint_dict_new[k] = v
13
+ torch.save(checkpoint_dict_new, output_path)
mel_processing.py ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.utils.data
3
+ from librosa.filters import mel as librosa_mel_fn
4
+
5
+ MAX_WAV_VALUE = 32768.0
6
+
7
+
8
+ def dynamic_range_compression_torch(x, C=1, clip_val=1e-5):
9
+ """
10
+ PARAMS
11
+ ------
12
+ C: compression factor
13
+ """
14
+ return torch.log(torch.clamp(x, min=clip_val) * C)
15
+
16
+
17
+ def dynamic_range_decompression_torch(x, C=1):
18
+ """
19
+ PARAMS
20
+ ------
21
+ C: compression factor used to compress
22
+ """
23
+ return torch.exp(x) / C
24
+
25
+
26
+ def spectral_normalize_torch(magnitudes):
27
+ output = dynamic_range_compression_torch(magnitudes)
28
+ return output
29
+
30
+
31
+ def spectral_de_normalize_torch(magnitudes):
32
+ output = dynamic_range_decompression_torch(magnitudes)
33
+ return output
34
+
35
+
36
+ mel_basis = {}
37
+ hann_window = {}
38
+
39
+
40
+ def spectrogram_torch(y, n_fft, sampling_rate, hop_size, win_size, center=False):
41
+ if torch.min(y) < -1.:
42
+ print('min value is ', torch.min(y))
43
+ if torch.max(y) > 1.:
44
+ print('max value is ', torch.max(y))
45
+
46
+ global hann_window
47
+ dtype_device = str(y.dtype) + '_' + str(y.device)
48
+ wnsize_dtype_device = str(win_size) + '_' + dtype_device
49
+ if wnsize_dtype_device not in hann_window:
50
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
51
+
52
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
53
+ y = y.squeeze(1)
54
+
55
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
56
+ center=center, pad_mode='reflect', normalized=False, onesided=True, return_complex=False)
57
+
58
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
59
+ return spec
60
+
61
+
62
+ def spec_to_mel_torch(spec, n_fft, num_mels, sampling_rate, fmin, fmax):
63
+ global mel_basis
64
+ dtype_device = str(spec.dtype) + '_' + str(spec.device)
65
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
66
+ if fmax_dtype_device not in mel_basis:
67
+ mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
68
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=spec.dtype, device=spec.device)
69
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
70
+ spec = spectral_normalize_torch(spec)
71
+ return spec
72
+
73
+
74
+ def mel_spectrogram_torch(y, n_fft, num_mels, sampling_rate, hop_size, win_size, fmin, fmax, center=False):
75
+ if torch.min(y) < -1.:
76
+ print('min value is ', torch.min(y))
77
+ if torch.max(y) > 1.:
78
+ print('max value is ', torch.max(y))
79
+
80
+ global mel_basis, hann_window
81
+ dtype_device = str(y.dtype) + '_' + str(y.device)
82
+ fmax_dtype_device = str(fmax) + '_' + dtype_device
83
+ wnsize_dtype_device = str(win_size) + '_' + dtype_device
84
+ if fmax_dtype_device not in mel_basis:
85
+ mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
86
+ mel_basis[fmax_dtype_device] = torch.from_numpy(mel).to(dtype=y.dtype, device=y.device)
87
+ if wnsize_dtype_device not in hann_window:
88
+ hann_window[wnsize_dtype_device] = torch.hann_window(win_size).to(dtype=y.dtype, device=y.device)
89
+
90
+ y = torch.nn.functional.pad(y.unsqueeze(1), (int((n_fft-hop_size)/2), int((n_fft-hop_size)/2)), mode='reflect')
91
+ y = y.squeeze(1)
92
+
93
+ spec = torch.stft(y, n_fft, hop_length=hop_size, win_length=win_size, window=hann_window[wnsize_dtype_device],
94
+ center=center, pad_mode='reflect', normalized=False, onesided=True)
95
+
96
+ spec = torch.sqrt(spec.pow(2).sum(-1) + 1e-6)
97
+
98
+ spec = torch.matmul(mel_basis[fmax_dtype_device], spec)
99
+ spec = spectral_normalize_torch(spec)
100
+
101
+ return spec
models.py ADDED
@@ -0,0 +1,540 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import math
2
+ import torch
3
+ from torch import nn
4
+ from torch.nn import functional as F
5
+
6
+ import commons
7
+ import modules
8
+ import attentions
9
+ import monotonic_align
10
+
11
+ from torch.nn import Conv1d, ConvTranspose1d, Conv2d
12
+ from torch.nn.utils import weight_norm, remove_weight_norm, spectral_norm
13
+ from commons import init_weights, get_padding
14
+
15
+
16
+ class StochasticDurationPredictor(nn.Module):
17
+ def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, n_flows=4, gin_channels=0):
18
+ super().__init__()
19
+ filter_channels = in_channels # it needs to be removed from future version.
20
+ self.in_channels = in_channels
21
+ self.filter_channels = filter_channels
22
+ self.kernel_size = kernel_size
23
+ self.p_dropout = p_dropout
24
+ self.n_flows = n_flows
25
+ self.gin_channels = gin_channels
26
+
27
+ self.log_flow = modules.Log()
28
+ self.flows = nn.ModuleList()
29
+ self.flows.append(modules.ElementwiseAffine(2))
30
+ for i in range(n_flows):
31
+ self.flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
32
+ self.flows.append(modules.Flip())
33
+
34
+ self.post_pre = nn.Conv1d(1, filter_channels, 1)
35
+ self.post_proj = nn.Conv1d(filter_channels, filter_channels, 1)
36
+ self.post_convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
37
+ self.post_flows = nn.ModuleList()
38
+ self.post_flows.append(modules.ElementwiseAffine(2))
39
+ for i in range(4):
40
+ self.post_flows.append(modules.ConvFlow(2, filter_channels, kernel_size, n_layers=3))
41
+ self.post_flows.append(modules.Flip())
42
+
43
+ self.pre = nn.Conv1d(in_channels, filter_channels, 1)
44
+ self.proj = nn.Conv1d(filter_channels, filter_channels, 1)
45
+ self.convs = modules.DDSConv(filter_channels, kernel_size, n_layers=3, p_dropout=p_dropout)
46
+ if gin_channels != 0:
47
+ self.cond = nn.Conv1d(gin_channels, filter_channels, 1)
48
+
49
+ def forward(self, x, x_mask, w=None, g=None, reverse=False, noise_scale=1.0):
50
+ x = torch.detach(x)
51
+ x = self.pre(x)
52
+ if g is not None:
53
+ g = torch.detach(g)
54
+ x = x + self.cond(g)
55
+ x = self.convs(x, x_mask)
56
+ x = self.proj(x) * x_mask
57
+
58
+ if not reverse:
59
+ flows = self.flows
60
+ assert w is not None
61
+
62
+ logdet_tot_q = 0
63
+ h_w = self.post_pre(w)
64
+ h_w = self.post_convs(h_w, x_mask)
65
+ h_w = self.post_proj(h_w) * x_mask
66
+ e_q = torch.randn(w.size(0), 2, w.size(2)).to(device=x.device, dtype=x.dtype) * x_mask
67
+ z_q = e_q
68
+ for flow in self.post_flows:
69
+ z_q, logdet_q = flow(z_q, x_mask, g=(x + h_w))
70
+ logdet_tot_q += logdet_q
71
+ z_u, z1 = torch.split(z_q, [1, 1], 1)
72
+ u = torch.sigmoid(z_u) * x_mask
73
+ z0 = (w - u) * x_mask
74
+ logdet_tot_q += torch.sum((F.logsigmoid(z_u) + F.logsigmoid(-z_u)) * x_mask, [1, 2])
75
+ logq = torch.sum(-0.5 * (math.log(2 * math.pi) + (e_q ** 2)) * x_mask, [1, 2]) - logdet_tot_q
76
+
77
+ logdet_tot = 0
78
+ z0, logdet = self.log_flow(z0, x_mask)
79
+ logdet_tot += logdet
80
+ z = torch.cat([z0, z1], 1)
81
+ for flow in flows:
82
+ z, logdet = flow(z, x_mask, g=x, reverse=reverse)
83
+ logdet_tot = logdet_tot + logdet
84
+ nll = torch.sum(0.5 * (math.log(2 * math.pi) + (z ** 2)) * x_mask, [1, 2]) - logdet_tot
85
+ return nll + logq # [b]
86
+ else:
87
+ flows = list(reversed(self.flows))
88
+ flows = flows[:-2] + [flows[-1]] # remove a useless vflow
89
+ z = torch.randn(x.size(0), 2, x.size(2)).to(device=x.device, dtype=x.dtype) * noise_scale
90
+ for flow in flows:
91
+ z = flow(z, x_mask, g=x, reverse=reverse)
92
+ z0, z1 = torch.split(z, [1, 1], 1)
93
+ logw = z0
94
+ return logw
95
+
96
+
97
+ class DurationPredictor(nn.Module):
98
+ def __init__(self, in_channels, filter_channels, kernel_size, p_dropout, gin_channels=0):
99
+ super().__init__()
100
+
101
+ self.in_channels = in_channels
102
+ self.filter_channels = filter_channels
103
+ self.kernel_size = kernel_size
104
+ self.p_dropout = p_dropout
105
+ self.gin_channels = gin_channels
106
+
107
+ self.drop = nn.Dropout(p_dropout)
108
+ self.conv_1 = nn.Conv1d(in_channels, filter_channels, kernel_size, padding=kernel_size // 2)
109
+ self.norm_1 = modules.LayerNorm(filter_channels)
110
+ self.conv_2 = nn.Conv1d(filter_channels, filter_channels, kernel_size, padding=kernel_size // 2)
111
+ self.norm_2 = modules.LayerNorm(filter_channels)
112
+ self.proj = nn.Conv1d(filter_channels, 1, 1)
113
+
114
+ if gin_channels != 0:
115
+ self.cond = nn.Conv1d(gin_channels, in_channels, 1)
116
+
117
+ def forward(self, x, x_mask, g=None):
118
+ x = torch.detach(x)
119
+ if g is not None:
120
+ g = torch.detach(g)
121
+ x = x + self.cond(g)
122
+ x = self.conv_1(x * x_mask)
123
+ x = torch.relu(x)
124
+ x = self.norm_1(x)
125
+ x = self.drop(x)
126
+ x = self.conv_2(x * x_mask)
127
+ x = torch.relu(x)
128
+ x = self.norm_2(x)
129
+ x = self.drop(x)
130
+ x = self.proj(x * x_mask)
131
+ return x * x_mask
132
+
133
+
134
+ class TextEncoder(nn.Module):
135
+ def __init__(self,
136
+ n_vocab,
137
+ out_channels,
138
+ hidden_channels,
139
+ filter_channels,
140
+ n_heads,
141
+ n_layers,
142
+ kernel_size,
143
+ p_dropout):
144
+ super().__init__()
145
+ self.n_vocab = n_vocab
146
+ self.out_channels = out_channels
147
+ self.hidden_channels = hidden_channels
148
+ self.filter_channels = filter_channels
149
+ self.n_heads = n_heads
150
+ self.n_layers = n_layers
151
+ self.kernel_size = kernel_size
152
+ self.p_dropout = p_dropout
153
+
154
+ if self.n_vocab != 0:
155
+ self.emb = nn.Embedding(n_vocab, hidden_channels)
156
+ nn.init.normal_(self.emb.weight, 0.0, hidden_channels ** -0.5)
157
+
158
+ self.encoder = attentions.Encoder(
159
+ hidden_channels,
160
+ filter_channels,
161
+ n_heads,
162
+ n_layers,
163
+ kernel_size,
164
+ p_dropout)
165
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
166
+
167
+ def forward(self, x, x_lengths):
168
+ if self.n_vocab != 0:
169
+ x = self.emb(x) * math.sqrt(self.hidden_channels) # [b, t, h]
170
+ x = torch.transpose(x, 1, -1) # [b, h, t]
171
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
172
+
173
+ x = self.encoder(x * x_mask, x_mask)
174
+ stats = self.proj(x) * x_mask
175
+
176
+ m, logs = torch.split(stats, self.out_channels, dim=1)
177
+ return x, m, logs, x_mask
178
+
179
+
180
+ class ResidualCouplingBlock(nn.Module):
181
+ def __init__(self,
182
+ channels,
183
+ hidden_channels,
184
+ kernel_size,
185
+ dilation_rate,
186
+ n_layers,
187
+ n_flows=4,
188
+ gin_channels=0):
189
+ super().__init__()
190
+ self.channels = channels
191
+ self.hidden_channels = hidden_channels
192
+ self.kernel_size = kernel_size
193
+ self.dilation_rate = dilation_rate
194
+ self.n_layers = n_layers
195
+ self.n_flows = n_flows
196
+ self.gin_channels = gin_channels
197
+
198
+ self.flows = nn.ModuleList()
199
+ for i in range(n_flows):
200
+ self.flows.append(
201
+ modules.ResidualCouplingLayer(channels, hidden_channels, kernel_size, dilation_rate, n_layers,
202
+ gin_channels=gin_channels, mean_only=True))
203
+ self.flows.append(modules.Flip())
204
+
205
+ def forward(self, x, x_mask, g=None, reverse=False):
206
+ if not reverse:
207
+ for flow in self.flows:
208
+ x, _ = flow(x, x_mask, g=g, reverse=reverse)
209
+ else:
210
+ for flow in reversed(self.flows):
211
+ x = flow(x, x_mask, g=g, reverse=reverse)
212
+ return x
213
+
214
+
215
+ class PosteriorEncoder(nn.Module):
216
+ def __init__(self,
217
+ in_channels,
218
+ out_channels,
219
+ hidden_channels,
220
+ kernel_size,
221
+ dilation_rate,
222
+ n_layers,
223
+ gin_channels=0):
224
+ super().__init__()
225
+ self.in_channels = in_channels
226
+ self.out_channels = out_channels
227
+ self.hidden_channels = hidden_channels
228
+ self.kernel_size = kernel_size
229
+ self.dilation_rate = dilation_rate
230
+ self.n_layers = n_layers
231
+ self.gin_channels = gin_channels
232
+
233
+ self.pre = nn.Conv1d(in_channels, hidden_channels, 1)
234
+ self.enc = modules.WN(hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=gin_channels)
235
+ self.proj = nn.Conv1d(hidden_channels, out_channels * 2, 1)
236
+
237
+ def forward(self, x, x_lengths, g=None):
238
+ x_mask = torch.unsqueeze(commons.sequence_mask(x_lengths, x.size(2)), 1).to(x.dtype)
239
+ x = self.pre(x) * x_mask
240
+ x = self.enc(x, x_mask, g=g)
241
+ stats = self.proj(x) * x_mask
242
+ m, logs = torch.split(stats, self.out_channels, dim=1)
243
+ z = (m + torch.randn_like(m) * torch.exp(logs)) * x_mask
244
+ return z, m, logs, x_mask
245
+
246
+
247
+ class Generator(torch.nn.Module):
248
+ def __init__(self, initial_channel, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates,
249
+ upsample_initial_channel, upsample_kernel_sizes, gin_channels=0):
250
+ super(Generator, self).__init__()
251
+ self.num_kernels = len(resblock_kernel_sizes)
252
+ self.num_upsamples = len(upsample_rates)
253
+ self.conv_pre = Conv1d(initial_channel, upsample_initial_channel, 7, 1, padding=3)
254
+ resblock = modules.ResBlock1 if resblock == '1' else modules.ResBlock2
255
+
256
+ self.ups = nn.ModuleList()
257
+ for i, (u, k) in enumerate(zip(upsample_rates, upsample_kernel_sizes)):
258
+ self.ups.append(weight_norm(
259
+ ConvTranspose1d(upsample_initial_channel // (2 ** i), upsample_initial_channel // (2 ** (i + 1)),
260
+ k, u, padding=(k - u) // 2)))
261
+
262
+ self.resblocks = nn.ModuleList()
263
+ for i in range(len(self.ups)):
264
+ ch = upsample_initial_channel // (2 ** (i + 1))
265
+ for j, (k, d) in enumerate(zip(resblock_kernel_sizes, resblock_dilation_sizes)):
266
+ self.resblocks.append(resblock(ch, k, d))
267
+
268
+ self.conv_post = Conv1d(ch, 1, 7, 1, padding=3, bias=False)
269
+ self.ups.apply(init_weights)
270
+
271
+ if gin_channels != 0:
272
+ self.cond = nn.Conv1d(gin_channels, upsample_initial_channel, 1)
273
+
274
+ def forward(self, x, g=None):
275
+ x = self.conv_pre(x)
276
+ if g is not None:
277
+ x = x + self.cond(g)
278
+
279
+ for i in range(self.num_upsamples):
280
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
281
+ x = self.ups[i](x)
282
+ xs = None
283
+ for j in range(self.num_kernels):
284
+ if xs is None:
285
+ xs = self.resblocks[i * self.num_kernels + j](x)
286
+ else:
287
+ xs += self.resblocks[i * self.num_kernels + j](x)
288
+ x = xs / self.num_kernels
289
+ x = F.leaky_relu(x)
290
+ x = self.conv_post(x)
291
+ x = torch.tanh(x)
292
+
293
+ return x
294
+
295
+ def remove_weight_norm(self):
296
+ print('Removing weight norm...')
297
+ for l in self.ups:
298
+ remove_weight_norm(l)
299
+ for l in self.resblocks:
300
+ l.remove_weight_norm()
301
+
302
+
303
+ class DiscriminatorP(torch.nn.Module):
304
+ def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
305
+ super(DiscriminatorP, self).__init__()
306
+ self.period = period
307
+ self.use_spectral_norm = use_spectral_norm
308
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
309
+ self.convs = nn.ModuleList([
310
+ norm_f(Conv2d(1, 32, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
311
+ norm_f(Conv2d(32, 128, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
312
+ norm_f(Conv2d(128, 512, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
313
+ norm_f(Conv2d(512, 1024, (kernel_size, 1), (stride, 1), padding=(get_padding(kernel_size, 1), 0))),
314
+ norm_f(Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(get_padding(kernel_size, 1), 0))),
315
+ ])
316
+ self.conv_post = norm_f(Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
317
+
318
+ def forward(self, x):
319
+ fmap = []
320
+
321
+ # 1d to 2d
322
+ b, c, t = x.shape
323
+ if t % self.period != 0: # pad first
324
+ n_pad = self.period - (t % self.period)
325
+ x = F.pad(x, (0, n_pad), "reflect")
326
+ t = t + n_pad
327
+ x = x.view(b, c, t // self.period, self.period)
328
+
329
+ for l in self.convs:
330
+ x = l(x)
331
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
332
+ fmap.append(x)
333
+ x = self.conv_post(x)
334
+ fmap.append(x)
335
+ x = torch.flatten(x, 1, -1)
336
+
337
+ return x, fmap
338
+
339
+
340
+ class DiscriminatorS(torch.nn.Module):
341
+ def __init__(self, use_spectral_norm=False):
342
+ super(DiscriminatorS, self).__init__()
343
+ norm_f = weight_norm if use_spectral_norm == False else spectral_norm
344
+ self.convs = nn.ModuleList([
345
+ norm_f(Conv1d(1, 16, 15, 1, padding=7)),
346
+ norm_f(Conv1d(16, 64, 41, 4, groups=4, padding=20)),
347
+ norm_f(Conv1d(64, 256, 41, 4, groups=16, padding=20)),
348
+ norm_f(Conv1d(256, 1024, 41, 4, groups=64, padding=20)),
349
+ norm_f(Conv1d(1024, 1024, 41, 4, groups=256, padding=20)),
350
+ norm_f(Conv1d(1024, 1024, 5, 1, padding=2)),
351
+ ])
352
+ self.conv_post = norm_f(Conv1d(1024, 1, 3, 1, padding=1))
353
+
354
+ def forward(self, x):
355
+ fmap = []
356
+
357
+ for l in self.convs:
358
+ x = l(x)
359
+ x = F.leaky_relu(x, modules.LRELU_SLOPE)
360
+ fmap.append(x)
361
+ x = self.conv_post(x)
362
+ fmap.append(x)
363
+ x = torch.flatten(x, 1, -1)
364
+
365
+ return x, fmap
366
+
367
+
368
+ class MultiPeriodDiscriminator(torch.nn.Module):
369
+ def __init__(self, use_spectral_norm=False):
370
+ super(MultiPeriodDiscriminator, self).__init__()
371
+ periods = [2, 3, 5, 7, 11]
372
+
373
+ discs = [DiscriminatorS(use_spectral_norm=use_spectral_norm)]
374
+ discs = discs + [DiscriminatorP(i, use_spectral_norm=use_spectral_norm) for i in periods]
375
+ self.discriminators = nn.ModuleList(discs)
376
+
377
+ def forward(self, y, y_hat):
378
+ y_d_rs = []
379
+ y_d_gs = []
380
+ fmap_rs = []
381
+ fmap_gs = []
382
+ for i, d in enumerate(self.discriminators):
383
+ y_d_r, fmap_r = d(y)
384
+ y_d_g, fmap_g = d(y_hat)
385
+ y_d_rs.append(y_d_r)
386
+ y_d_gs.append(y_d_g)
387
+ fmap_rs.append(fmap_r)
388
+ fmap_gs.append(fmap_g)
389
+
390
+ return y_d_rs, y_d_gs, fmap_rs, fmap_gs
391
+
392
+
393
+ class SynthesizerTrn(nn.Module):
394
+ """
395
+ Synthesizer for Training
396
+ """
397
+
398
+ def __init__(self,
399
+ n_vocab,
400
+ spec_channels,
401
+ segment_size,
402
+ inter_channels,
403
+ hidden_channels,
404
+ filter_channels,
405
+ n_heads,
406
+ n_layers,
407
+ kernel_size,
408
+ p_dropout,
409
+ resblock,
410
+ resblock_kernel_sizes,
411
+ resblock_dilation_sizes,
412
+ upsample_rates,
413
+ upsample_initial_channel,
414
+ upsample_kernel_sizes,
415
+ n_speakers=0,
416
+ gin_channels=0,
417
+ use_sdp=True,
418
+ **kwargs):
419
+
420
+ super().__init__()
421
+ self.n_vocab = n_vocab
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.n_speakers = n_speakers
438
+ self.gin_channels = gin_channels
439
+
440
+ self.use_sdp = use_sdp
441
+
442
+ self.enc_p = TextEncoder(n_vocab,
443
+ inter_channels,
444
+ hidden_channels,
445
+ filter_channels,
446
+ n_heads,
447
+ n_layers,
448
+ kernel_size,
449
+ p_dropout)
450
+ self.dec = Generator(inter_channels, resblock, resblock_kernel_sizes, resblock_dilation_sizes, upsample_rates,
451
+ upsample_initial_channel, upsample_kernel_sizes, gin_channels=gin_channels)
452
+ self.enc_q = PosteriorEncoder(spec_channels, inter_channels, hidden_channels, 5, 1, 16,
453
+ gin_channels=gin_channels)
454
+ self.flow = ResidualCouplingBlock(inter_channels, hidden_channels, 5, 1, 4, gin_channels=gin_channels)
455
+
456
+ if use_sdp:
457
+ self.dp = StochasticDurationPredictor(hidden_channels, 192, 3, 0.5, 4, gin_channels=gin_channels)
458
+ else:
459
+ self.dp = DurationPredictor(hidden_channels, 256, 3, 0.5, gin_channels=gin_channels)
460
+
461
+ if n_speakers > 1:
462
+ self.emb_g = nn.Embedding(n_speakers, gin_channels)
463
+
464
+ def forward(self, x, x_lengths, y, y_lengths, sid=None):
465
+
466
+ x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
467
+ if self.n_speakers > 1:
468
+ g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
469
+ else:
470
+ g = None
471
+
472
+ z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g)
473
+ z_p = self.flow(z, y_mask, g=g)
474
+
475
+ with torch.no_grad():
476
+ # negative cross-entropy
477
+ s_p_sq_r = torch.exp(-2 * logs_p) # [b, d, t]
478
+ neg_cent1 = torch.sum(-0.5 * math.log(2 * math.pi) - logs_p, [1], keepdim=True) # [b, 1, t_s]
479
+ neg_cent2 = torch.matmul(-0.5 * (z_p ** 2).transpose(1, 2),
480
+ s_p_sq_r) # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]
481
+ neg_cent3 = torch.matmul(z_p.transpose(1, 2), (m_p * s_p_sq_r)) # [b, t_t, d] x [b, d, t_s] = [b, t_t, t_s]
482
+ neg_cent4 = torch.sum(-0.5 * (m_p ** 2) * s_p_sq_r, [1], keepdim=True) # [b, 1, t_s]
483
+ neg_cent = neg_cent1 + neg_cent2 + neg_cent3 + neg_cent4
484
+
485
+ attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
486
+ attn = monotonic_align.maximum_path(neg_cent, attn_mask.squeeze(1)).unsqueeze(1).detach()
487
+
488
+ w = attn.sum(2)
489
+ if self.use_sdp:
490
+ l_length = self.dp(x, x_mask, w, g=g)
491
+ l_length = l_length / torch.sum(x_mask)
492
+ else:
493
+ logw_ = torch.log(w + 1e-6) * x_mask
494
+ logw = self.dp(x, x_mask, g=g)
495
+ l_length = torch.sum((logw - logw_) ** 2, [1, 2]) / torch.sum(x_mask) # for averaging
496
+
497
+ # expand prior
498
+ m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2)
499
+ logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1, 2)
500
+
501
+ z_slice, ids_slice = commons.rand_slice_segments(z, y_lengths, self.segment_size)
502
+ o = self.dec(z_slice, g=g)
503
+ return o, l_length, attn, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
504
+
505
+ def infer(self, x, x_lengths, sid=None, noise_scale=1, length_scale=1, noise_scale_w=1., max_len=None):
506
+ x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
507
+ if self.n_speakers > 1:
508
+ g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
509
+ else:
510
+ g = None
511
+
512
+ if self.use_sdp:
513
+ logw = self.dp(x, x_mask, g=g, reverse=True, noise_scale=noise_scale_w)
514
+ else:
515
+ logw = self.dp(x, x_mask, g=g)
516
+ w = torch.exp(logw) * x_mask * length_scale
517
+ w_ceil = torch.ceil(w)
518
+ y_lengths = torch.clamp_min(torch.sum(w_ceil, [1, 2]), 1).long()
519
+ y_mask = torch.unsqueeze(commons.sequence_mask(y_lengths, None), 1).to(x_mask.dtype)
520
+ attn_mask = torch.unsqueeze(x_mask, 2) * torch.unsqueeze(y_mask, -1)
521
+ attn = commons.generate_path(w_ceil, attn_mask)
522
+
523
+ m_p = torch.matmul(attn.squeeze(1), m_p.transpose(1, 2)).transpose(1, 2) # [b, t', t], [b, t, d] -> [b, d, t']
524
+ logs_p = torch.matmul(attn.squeeze(1), logs_p.transpose(1, 2)).transpose(1,
525
+ 2) # [b, t', t], [b, t, d] -> [b, d, t']
526
+
527
+ z_p = m_p + torch.randn_like(m_p) * torch.exp(logs_p) * noise_scale
528
+ z = self.flow(z_p, y_mask, g=g, reverse=True)
529
+ o = self.dec((z * y_mask)[:, :, :max_len], g=g)
530
+ return o, attn, y_mask, (z, z_p, m_p, logs_p)
531
+
532
+ def voice_conversion(self, y, y_lengths, sid_src, sid_tgt):
533
+ assert self.n_speakers > 1, "n_speakers have to be larger than 1."
534
+ g_src = self.emb_g(sid_src).unsqueeze(-1)
535
+ g_tgt = self.emb_g(sid_tgt).unsqueeze(-1)
536
+ z, m_q, logs_q, y_mask = self.enc_q(y, y_lengths, g=g_src)
537
+ z_p = self.flow(z, y_mask, g=g_src)
538
+ z_hat = self.flow(z_p, y_mask, g=g_tgt, reverse=True)
539
+ o_hat = self.dec(z_hat * y_mask, g=g_tgt)
540
+ return o_hat, y_mask, (z, z_p, z_hat)
modules.py ADDED
@@ -0,0 +1,390 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+
9
+ from torch.nn import Conv1d, ConvTranspose1d, AvgPool1d, Conv2d
10
+ from torch.nn.utils import weight_norm, remove_weight_norm
11
+
12
+ import commons
13
+ from commons import init_weights, get_padding
14
+ from transforms import piecewise_rational_quadratic_transform
15
+
16
+
17
+ LRELU_SLOPE = 0.1
18
+
19
+
20
+ class LayerNorm(nn.Module):
21
+ def __init__(self, channels, eps=1e-5):
22
+ super().__init__()
23
+ self.channels = channels
24
+ self.eps = eps
25
+
26
+ self.gamma = nn.Parameter(torch.ones(channels))
27
+ self.beta = nn.Parameter(torch.zeros(channels))
28
+
29
+ def forward(self, x):
30
+ x = x.transpose(1, -1)
31
+ x = F.layer_norm(x, (self.channels,), self.gamma, self.beta, self.eps)
32
+ return x.transpose(1, -1)
33
+
34
+
35
+ class ConvReluNorm(nn.Module):
36
+ def __init__(self, in_channels, hidden_channels, out_channels, kernel_size, n_layers, p_dropout):
37
+ super().__init__()
38
+ self.in_channels = in_channels
39
+ self.hidden_channels = hidden_channels
40
+ self.out_channels = out_channels
41
+ self.kernel_size = kernel_size
42
+ self.n_layers = n_layers
43
+ self.p_dropout = p_dropout
44
+ assert n_layers > 1, "Number of layers should be larger than 0."
45
+
46
+ self.conv_layers = nn.ModuleList()
47
+ self.norm_layers = nn.ModuleList()
48
+ self.conv_layers.append(nn.Conv1d(in_channels, hidden_channels, kernel_size, padding=kernel_size//2))
49
+ self.norm_layers.append(LayerNorm(hidden_channels))
50
+ self.relu_drop = nn.Sequential(
51
+ nn.ReLU(),
52
+ nn.Dropout(p_dropout))
53
+ for _ in range(n_layers-1):
54
+ self.conv_layers.append(nn.Conv1d(hidden_channels, hidden_channels, kernel_size, padding=kernel_size//2))
55
+ self.norm_layers.append(LayerNorm(hidden_channels))
56
+ self.proj = nn.Conv1d(hidden_channels, out_channels, 1)
57
+ self.proj.weight.data.zero_()
58
+ self.proj.bias.data.zero_()
59
+
60
+ def forward(self, x, x_mask):
61
+ x_org = x
62
+ for i in range(self.n_layers):
63
+ x = self.conv_layers[i](x * x_mask)
64
+ x = self.norm_layers[i](x)
65
+ x = self.relu_drop(x)
66
+ x = x_org + self.proj(x)
67
+ return x * x_mask
68
+
69
+
70
+ class DDSConv(nn.Module):
71
+ """
72
+ Dialted and Depth-Separable Convolution
73
+ """
74
+ def __init__(self, channels, kernel_size, n_layers, p_dropout=0.):
75
+ super().__init__()
76
+ self.channels = channels
77
+ self.kernel_size = kernel_size
78
+ self.n_layers = n_layers
79
+ self.p_dropout = p_dropout
80
+
81
+ self.drop = nn.Dropout(p_dropout)
82
+ self.convs_sep = nn.ModuleList()
83
+ self.convs_1x1 = nn.ModuleList()
84
+ self.norms_1 = nn.ModuleList()
85
+ self.norms_2 = nn.ModuleList()
86
+ for i in range(n_layers):
87
+ dilation = kernel_size ** i
88
+ padding = (kernel_size * dilation - dilation) // 2
89
+ self.convs_sep.append(nn.Conv1d(channels, channels, kernel_size,
90
+ groups=channels, dilation=dilation, padding=padding
91
+ ))
92
+ self.convs_1x1.append(nn.Conv1d(channels, channels, 1))
93
+ self.norms_1.append(LayerNorm(channels))
94
+ self.norms_2.append(LayerNorm(channels))
95
+
96
+ def forward(self, x, x_mask, g=None):
97
+ if g is not None:
98
+ x = x + g
99
+ for i in range(self.n_layers):
100
+ y = self.convs_sep[i](x * x_mask)
101
+ y = self.norms_1[i](y)
102
+ y = F.gelu(y)
103
+ y = self.convs_1x1[i](y)
104
+ y = self.norms_2[i](y)
105
+ y = F.gelu(y)
106
+ y = self.drop(y)
107
+ x = x + y
108
+ return x * x_mask
109
+
110
+
111
+ class WN(torch.nn.Module):
112
+ def __init__(self, hidden_channels, kernel_size, dilation_rate, n_layers, gin_channels=0, p_dropout=0):
113
+ super(WN, self).__init__()
114
+ assert(kernel_size % 2 == 1)
115
+ self.hidden_channels =hidden_channels
116
+ self.kernel_size = kernel_size,
117
+ self.dilation_rate = dilation_rate
118
+ self.n_layers = n_layers
119
+ self.gin_channels = gin_channels
120
+ self.p_dropout = p_dropout
121
+
122
+ self.in_layers = torch.nn.ModuleList()
123
+ self.res_skip_layers = torch.nn.ModuleList()
124
+ self.drop = nn.Dropout(p_dropout)
125
+
126
+ if gin_channels != 0:
127
+ cond_layer = torch.nn.Conv1d(gin_channels, 2*hidden_channels*n_layers, 1)
128
+ self.cond_layer = torch.nn.utils.weight_norm(cond_layer, name='weight')
129
+
130
+ for i in range(n_layers):
131
+ dilation = dilation_rate ** i
132
+ padding = int((kernel_size * dilation - dilation) / 2)
133
+ in_layer = torch.nn.Conv1d(hidden_channels, 2*hidden_channels, kernel_size,
134
+ dilation=dilation, padding=padding)
135
+ in_layer = torch.nn.utils.weight_norm(in_layer, name='weight')
136
+ self.in_layers.append(in_layer)
137
+
138
+ # last one is not necessary
139
+ if i < n_layers - 1:
140
+ res_skip_channels = 2 * hidden_channels
141
+ else:
142
+ res_skip_channels = hidden_channels
143
+
144
+ res_skip_layer = torch.nn.Conv1d(hidden_channels, res_skip_channels, 1)
145
+ res_skip_layer = torch.nn.utils.weight_norm(res_skip_layer, name='weight')
146
+ self.res_skip_layers.append(res_skip_layer)
147
+
148
+ def forward(self, x, x_mask, g=None, **kwargs):
149
+ output = torch.zeros_like(x)
150
+ n_channels_tensor = torch.IntTensor([self.hidden_channels])
151
+
152
+ if g is not None:
153
+ g = self.cond_layer(g)
154
+
155
+ for i in range(self.n_layers):
156
+ x_in = self.in_layers[i](x)
157
+ if g is not None:
158
+ cond_offset = i * 2 * self.hidden_channels
159
+ g_l = g[:,cond_offset:cond_offset+2*self.hidden_channels,:]
160
+ else:
161
+ g_l = torch.zeros_like(x_in)
162
+
163
+ acts = commons.fused_add_tanh_sigmoid_multiply(
164
+ x_in,
165
+ g_l,
166
+ n_channels_tensor)
167
+ acts = self.drop(acts)
168
+
169
+ res_skip_acts = self.res_skip_layers[i](acts)
170
+ if i < self.n_layers - 1:
171
+ res_acts = res_skip_acts[:,:self.hidden_channels,:]
172
+ x = (x + res_acts) * x_mask
173
+ output = output + res_skip_acts[:,self.hidden_channels:,:]
174
+ else:
175
+ output = output + res_skip_acts
176
+ return output * x_mask
177
+
178
+ def remove_weight_norm(self):
179
+ if self.gin_channels != 0:
180
+ torch.nn.utils.remove_weight_norm(self.cond_layer)
181
+ for l in self.in_layers:
182
+ torch.nn.utils.remove_weight_norm(l)
183
+ for l in self.res_skip_layers:
184
+ torch.nn.utils.remove_weight_norm(l)
185
+
186
+
187
+ class ResBlock1(torch.nn.Module):
188
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
189
+ super(ResBlock1, self).__init__()
190
+ self.convs1 = nn.ModuleList([
191
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
192
+ padding=get_padding(kernel_size, dilation[0]))),
193
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
194
+ padding=get_padding(kernel_size, dilation[1]))),
195
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[2],
196
+ padding=get_padding(kernel_size, dilation[2])))
197
+ ])
198
+ self.convs1.apply(init_weights)
199
+
200
+ self.convs2 = nn.ModuleList([
201
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
202
+ padding=get_padding(kernel_size, 1))),
203
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
204
+ padding=get_padding(kernel_size, 1))),
205
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=1,
206
+ padding=get_padding(kernel_size, 1)))
207
+ ])
208
+ self.convs2.apply(init_weights)
209
+
210
+ def forward(self, x, x_mask=None):
211
+ for c1, c2 in zip(self.convs1, self.convs2):
212
+ xt = F.leaky_relu(x, LRELU_SLOPE)
213
+ if x_mask is not None:
214
+ xt = xt * x_mask
215
+ xt = c1(xt)
216
+ xt = F.leaky_relu(xt, LRELU_SLOPE)
217
+ if x_mask is not None:
218
+ xt = xt * x_mask
219
+ xt = c2(xt)
220
+ x = xt + x
221
+ if x_mask is not None:
222
+ x = x * x_mask
223
+ return x
224
+
225
+ def remove_weight_norm(self):
226
+ for l in self.convs1:
227
+ remove_weight_norm(l)
228
+ for l in self.convs2:
229
+ remove_weight_norm(l)
230
+
231
+
232
+ class ResBlock2(torch.nn.Module):
233
+ def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
234
+ super(ResBlock2, self).__init__()
235
+ self.convs = nn.ModuleList([
236
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[0],
237
+ padding=get_padding(kernel_size, dilation[0]))),
238
+ weight_norm(Conv1d(channels, channels, kernel_size, 1, dilation=dilation[1],
239
+ padding=get_padding(kernel_size, dilation[1])))
240
+ ])
241
+ self.convs.apply(init_weights)
242
+
243
+ def forward(self, x, x_mask=None):
244
+ for c in self.convs:
245
+ xt = F.leaky_relu(x, LRELU_SLOPE)
246
+ if x_mask is not None:
247
+ xt = xt * x_mask
248
+ xt = c(xt)
249
+ x = xt + x
250
+ if x_mask is not None:
251
+ x = x * x_mask
252
+ return x
253
+
254
+ def remove_weight_norm(self):
255
+ for l in self.convs:
256
+ remove_weight_norm(l)
257
+
258
+
259
+ class Log(nn.Module):
260
+ def forward(self, x, x_mask, reverse=False, **kwargs):
261
+ if not reverse:
262
+ y = torch.log(torch.clamp_min(x, 1e-5)) * x_mask
263
+ logdet = torch.sum(-y, [1, 2])
264
+ return y, logdet
265
+ else:
266
+ x = torch.exp(x) * x_mask
267
+ return x
268
+
269
+
270
+ class Flip(nn.Module):
271
+ def forward(self, x, *args, reverse=False, **kwargs):
272
+ x = torch.flip(x, [1])
273
+ if not reverse:
274
+ logdet = torch.zeros(x.size(0)).to(dtype=x.dtype, device=x.device)
275
+ return x, logdet
276
+ else:
277
+ return x
278
+
279
+
280
+ class ElementwiseAffine(nn.Module):
281
+ def __init__(self, channels):
282
+ super().__init__()
283
+ self.channels = channels
284
+ self.m = nn.Parameter(torch.zeros(channels,1))
285
+ self.logs = nn.Parameter(torch.zeros(channels,1))
286
+
287
+ def forward(self, x, x_mask, reverse=False, **kwargs):
288
+ if not reverse:
289
+ y = self.m + torch.exp(self.logs) * x
290
+ y = y * x_mask
291
+ logdet = torch.sum(self.logs * x_mask, [1,2])
292
+ return y, logdet
293
+ else:
294
+ x = (x - self.m) * torch.exp(-self.logs) * x_mask
295
+ return x
296
+
297
+
298
+ class ResidualCouplingLayer(nn.Module):
299
+ def __init__(self,
300
+ channels,
301
+ hidden_channels,
302
+ kernel_size,
303
+ dilation_rate,
304
+ n_layers,
305
+ p_dropout=0,
306
+ gin_channels=0,
307
+ mean_only=False):
308
+ assert channels % 2 == 0, "channels should be divisible by 2"
309
+ super().__init__()
310
+ self.channels = channels
311
+ self.hidden_channels = hidden_channels
312
+ self.kernel_size = kernel_size
313
+ self.dilation_rate = dilation_rate
314
+ self.n_layers = n_layers
315
+ self.half_channels = channels // 2
316
+ self.mean_only = mean_only
317
+
318
+ self.pre = nn.Conv1d(self.half_channels, hidden_channels, 1)
319
+ self.enc = WN(hidden_channels, kernel_size, dilation_rate, n_layers, p_dropout=p_dropout, gin_channels=gin_channels)
320
+ self.post = nn.Conv1d(hidden_channels, self.half_channels * (2 - mean_only), 1)
321
+ self.post.weight.data.zero_()
322
+ self.post.bias.data.zero_()
323
+
324
+ def forward(self, x, x_mask, g=None, reverse=False):
325
+ x0, x1 = torch.split(x, [self.half_channels]*2, 1)
326
+ h = self.pre(x0) * x_mask
327
+ h = self.enc(h, x_mask, g=g)
328
+ stats = self.post(h) * x_mask
329
+ if not self.mean_only:
330
+ m, logs = torch.split(stats, [self.half_channels]*2, 1)
331
+ else:
332
+ m = stats
333
+ logs = torch.zeros_like(m)
334
+
335
+ if not reverse:
336
+ x1 = m + x1 * torch.exp(logs) * x_mask
337
+ x = torch.cat([x0, x1], 1)
338
+ logdet = torch.sum(logs, [1,2])
339
+ return x, logdet
340
+ else:
341
+ x1 = (x1 - m) * torch.exp(-logs) * x_mask
342
+ x = torch.cat([x0, x1], 1)
343
+ return x
344
+
345
+
346
+ class ConvFlow(nn.Module):
347
+ def __init__(self, in_channels, filter_channels, kernel_size, n_layers, num_bins=10, tail_bound=5.0):
348
+ super().__init__()
349
+ self.in_channels = in_channels
350
+ self.filter_channels = filter_channels
351
+ self.kernel_size = kernel_size
352
+ self.n_layers = n_layers
353
+ self.num_bins = num_bins
354
+ self.tail_bound = tail_bound
355
+ self.half_channels = in_channels // 2
356
+
357
+ self.pre = nn.Conv1d(self.half_channels, filter_channels, 1)
358
+ self.convs = DDSConv(filter_channels, kernel_size, n_layers, p_dropout=0.)
359
+ self.proj = nn.Conv1d(filter_channels, self.half_channels * (num_bins * 3 - 1), 1)
360
+ self.proj.weight.data.zero_()
361
+ self.proj.bias.data.zero_()
362
+
363
+ def forward(self, x, x_mask, g=None, reverse=False):
364
+ x0, x1 = torch.split(x, [self.half_channels]*2, 1)
365
+ h = self.pre(x0)
366
+ h = self.convs(h, x_mask, g=g)
367
+ h = self.proj(h) * x_mask
368
+
369
+ b, c, t = x0.shape
370
+ h = h.reshape(b, c, -1, t).permute(0, 1, 3, 2) # [b, cx?, t] -> [b, c, t, ?]
371
+
372
+ unnormalized_widths = h[..., :self.num_bins] / math.sqrt(self.filter_channels)
373
+ unnormalized_heights = h[..., self.num_bins:2*self.num_bins] / math.sqrt(self.filter_channels)
374
+ unnormalized_derivatives = h[..., 2 * self.num_bins:]
375
+
376
+ x1, logabsdet = piecewise_rational_quadratic_transform(x1,
377
+ unnormalized_widths,
378
+ unnormalized_heights,
379
+ unnormalized_derivatives,
380
+ inverse=reverse,
381
+ tails='linear',
382
+ tail_bound=self.tail_bound
383
+ )
384
+
385
+ x = torch.cat([x0, x1], 1) * x_mask
386
+ logdet = torch.sum(logabsdet * x_mask, [1,2])
387
+ if not reverse:
388
+ return x, logdet
389
+ else:
390
+ return x
requirements.txt ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ numba
2
+ librosa
3
+ matplotlib
4
+ numpy
5
+ phonemizer
6
+ scipy
7
+ tensorboard
8
+ torch
9
+ torchvision
10
+ torchaudio
11
+ unidecode
12
+ pyopenjtalk>=0.3.0
13
+ jamo
14
+ pypinyin
15
+ ko_pron
16
+ jieba
17
+ cn2an
18
+ protobuf
19
+ inflect
20
+ eng_to_ipa
21
+ ko_pron
22
+ indic_transliteration
23
+ num_thai
24
+ opencc
25
+ gradio
transforms.py ADDED
@@ -0,0 +1,193 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from torch.nn import functional as F
3
+
4
+ import numpy as np
5
+
6
+
7
+ DEFAULT_MIN_BIN_WIDTH = 1e-3
8
+ DEFAULT_MIN_BIN_HEIGHT = 1e-3
9
+ DEFAULT_MIN_DERIVATIVE = 1e-3
10
+
11
+
12
+ def piecewise_rational_quadratic_transform(inputs,
13
+ unnormalized_widths,
14
+ unnormalized_heights,
15
+ unnormalized_derivatives,
16
+ inverse=False,
17
+ tails=None,
18
+ tail_bound=1.,
19
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
20
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
21
+ min_derivative=DEFAULT_MIN_DERIVATIVE):
22
+
23
+ if tails is None:
24
+ spline_fn = rational_quadratic_spline
25
+ spline_kwargs = {}
26
+ else:
27
+ spline_fn = unconstrained_rational_quadratic_spline
28
+ spline_kwargs = {
29
+ 'tails': tails,
30
+ 'tail_bound': tail_bound
31
+ }
32
+
33
+ outputs, logabsdet = spline_fn(
34
+ inputs=inputs,
35
+ unnormalized_widths=unnormalized_widths,
36
+ unnormalized_heights=unnormalized_heights,
37
+ unnormalized_derivatives=unnormalized_derivatives,
38
+ inverse=inverse,
39
+ min_bin_width=min_bin_width,
40
+ min_bin_height=min_bin_height,
41
+ min_derivative=min_derivative,
42
+ **spline_kwargs
43
+ )
44
+ return outputs, logabsdet
45
+
46
+
47
+ def searchsorted(bin_locations, inputs, eps=1e-6):
48
+ bin_locations[..., -1] += eps
49
+ return torch.sum(
50
+ inputs[..., None] >= bin_locations,
51
+ dim=-1
52
+ ) - 1
53
+
54
+
55
+ def unconstrained_rational_quadratic_spline(inputs,
56
+ unnormalized_widths,
57
+ unnormalized_heights,
58
+ unnormalized_derivatives,
59
+ inverse=False,
60
+ tails='linear',
61
+ tail_bound=1.,
62
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
63
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
64
+ min_derivative=DEFAULT_MIN_DERIVATIVE):
65
+ inside_interval_mask = (inputs >= -tail_bound) & (inputs <= tail_bound)
66
+ outside_interval_mask = ~inside_interval_mask
67
+
68
+ outputs = torch.zeros_like(inputs)
69
+ logabsdet = torch.zeros_like(inputs)
70
+
71
+ if tails == 'linear':
72
+ unnormalized_derivatives = F.pad(unnormalized_derivatives, pad=(1, 1))
73
+ constant = np.log(np.exp(1 - min_derivative) - 1)
74
+ unnormalized_derivatives[..., 0] = constant
75
+ unnormalized_derivatives[..., -1] = constant
76
+
77
+ outputs[outside_interval_mask] = inputs[outside_interval_mask]
78
+ logabsdet[outside_interval_mask] = 0
79
+ else:
80
+ raise RuntimeError('{} tails are not implemented.'.format(tails))
81
+
82
+ outputs[inside_interval_mask], logabsdet[inside_interval_mask] = rational_quadratic_spline(
83
+ inputs=inputs[inside_interval_mask],
84
+ unnormalized_widths=unnormalized_widths[inside_interval_mask, :],
85
+ unnormalized_heights=unnormalized_heights[inside_interval_mask, :],
86
+ unnormalized_derivatives=unnormalized_derivatives[inside_interval_mask, :],
87
+ inverse=inverse,
88
+ left=-tail_bound, right=tail_bound, bottom=-tail_bound, top=tail_bound,
89
+ min_bin_width=min_bin_width,
90
+ min_bin_height=min_bin_height,
91
+ min_derivative=min_derivative
92
+ )
93
+
94
+ return outputs, logabsdet
95
+
96
+ def rational_quadratic_spline(inputs,
97
+ unnormalized_widths,
98
+ unnormalized_heights,
99
+ unnormalized_derivatives,
100
+ inverse=False,
101
+ left=0., right=1., bottom=0., top=1.,
102
+ min_bin_width=DEFAULT_MIN_BIN_WIDTH,
103
+ min_bin_height=DEFAULT_MIN_BIN_HEIGHT,
104
+ min_derivative=DEFAULT_MIN_DERIVATIVE):
105
+ if torch.min(inputs) < left or torch.max(inputs) > right:
106
+ raise ValueError('Input to a transform is not within its domain')
107
+
108
+ num_bins = unnormalized_widths.shape[-1]
109
+
110
+ if min_bin_width * num_bins > 1.0:
111
+ raise ValueError('Minimal bin width too large for the number of bins')
112
+ if min_bin_height * num_bins > 1.0:
113
+ raise ValueError('Minimal bin height too large for the number of bins')
114
+
115
+ widths = F.softmax(unnormalized_widths, dim=-1)
116
+ widths = min_bin_width + (1 - min_bin_width * num_bins) * widths
117
+ cumwidths = torch.cumsum(widths, dim=-1)
118
+ cumwidths = F.pad(cumwidths, pad=(1, 0), mode='constant', value=0.0)
119
+ cumwidths = (right - left) * cumwidths + left
120
+ cumwidths[..., 0] = left
121
+ cumwidths[..., -1] = right
122
+ widths = cumwidths[..., 1:] - cumwidths[..., :-1]
123
+
124
+ derivatives = min_derivative + F.softplus(unnormalized_derivatives)
125
+
126
+ heights = F.softmax(unnormalized_heights, dim=-1)
127
+ heights = min_bin_height + (1 - min_bin_height * num_bins) * heights
128
+ cumheights = torch.cumsum(heights, dim=-1)
129
+ cumheights = F.pad(cumheights, pad=(1, 0), mode='constant', value=0.0)
130
+ cumheights = (top - bottom) * cumheights + bottom
131
+ cumheights[..., 0] = bottom
132
+ cumheights[..., -1] = top
133
+ heights = cumheights[..., 1:] - cumheights[..., :-1]
134
+
135
+ if inverse:
136
+ bin_idx = searchsorted(cumheights, inputs)[..., None]
137
+ else:
138
+ bin_idx = searchsorted(cumwidths, inputs)[..., None]
139
+
140
+ input_cumwidths = cumwidths.gather(-1, bin_idx)[..., 0]
141
+ input_bin_widths = widths.gather(-1, bin_idx)[..., 0]
142
+
143
+ input_cumheights = cumheights.gather(-1, bin_idx)[..., 0]
144
+ delta = heights / widths
145
+ input_delta = delta.gather(-1, bin_idx)[..., 0]
146
+
147
+ input_derivatives = derivatives.gather(-1, bin_idx)[..., 0]
148
+ input_derivatives_plus_one = derivatives[..., 1:].gather(-1, bin_idx)[..., 0]
149
+
150
+ input_heights = heights.gather(-1, bin_idx)[..., 0]
151
+
152
+ if inverse:
153
+ a = (((inputs - input_cumheights) * (input_derivatives
154
+ + input_derivatives_plus_one
155
+ - 2 * input_delta)
156
+ + input_heights * (input_delta - input_derivatives)))
157
+ b = (input_heights * input_derivatives
158
+ - (inputs - input_cumheights) * (input_derivatives
159
+ + input_derivatives_plus_one
160
+ - 2 * input_delta))
161
+ c = - input_delta * (inputs - input_cumheights)
162
+
163
+ discriminant = b.pow(2) - 4 * a * c
164
+ assert (discriminant >= 0).all()
165
+
166
+ root = (2 * c) / (-b - torch.sqrt(discriminant))
167
+ outputs = root * input_bin_widths + input_cumwidths
168
+
169
+ theta_one_minus_theta = root * (1 - root)
170
+ denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta)
171
+ * theta_one_minus_theta)
172
+ derivative_numerator = input_delta.pow(2) * (input_derivatives_plus_one * root.pow(2)
173
+ + 2 * input_delta * theta_one_minus_theta
174
+ + input_derivatives * (1 - root).pow(2))
175
+ logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
176
+
177
+ return outputs, -logabsdet
178
+ else:
179
+ theta = (inputs - input_cumwidths) / input_bin_widths
180
+ theta_one_minus_theta = theta * (1 - theta)
181
+
182
+ numerator = input_heights * (input_delta * theta.pow(2)
183
+ + input_derivatives * theta_one_minus_theta)
184
+ denominator = input_delta + ((input_derivatives + input_derivatives_plus_one - 2 * input_delta)
185
+ * theta_one_minus_theta)
186
+ outputs = input_cumheights + numerator / denominator
187
+
188
+ derivative_numerator = input_delta.pow(2) * (input_derivatives_plus_one * theta.pow(2)
189
+ + 2 * input_delta * theta_one_minus_theta
190
+ + input_derivatives * (1 - theta).pow(2))
191
+ logabsdet = torch.log(derivative_numerator) - 2 * torch.log(denominator)
192
+
193
+ return outputs, logabsdet
utils.py ADDED
@@ -0,0 +1,226 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import glob
3
+ import sys
4
+ import argparse
5
+ import logging
6
+ import json
7
+ import subprocess
8
+ import numpy as np
9
+ from scipy.io.wavfile import read
10
+ import torch
11
+
12
+ MATPLOTLIB_FLAG = False
13
+
14
+ logging.basicConfig(stream=sys.stdout, level=logging.ERROR)
15
+ logger = logging
16
+
17
+
18
+ def load_checkpoint(checkpoint_path, model, optimizer=None):
19
+ assert os.path.isfile(checkpoint_path)
20
+ checkpoint_dict = torch.load(checkpoint_path, map_location='cpu')
21
+ iteration = checkpoint_dict['iteration']
22
+ learning_rate = checkpoint_dict['learning_rate']
23
+ if optimizer is not None:
24
+ optimizer.load_state_dict(checkpoint_dict['optimizer'])
25
+ saved_state_dict = checkpoint_dict['model']
26
+ if hasattr(model, 'module'):
27
+ state_dict = model.module.state_dict()
28
+ else:
29
+ state_dict = model.state_dict()
30
+ new_state_dict = {}
31
+ for k, v in state_dict.items():
32
+ try:
33
+ new_state_dict[k] = saved_state_dict[k]
34
+ except:
35
+ logger.info("%s is not in the checkpoint" % k)
36
+ new_state_dict[k] = v
37
+ if hasattr(model, 'module'):
38
+ model.module.load_state_dict(new_state_dict)
39
+ else:
40
+ model.load_state_dict(new_state_dict)
41
+ logger.info("Loaded checkpoint '{}' (iteration {})".format(
42
+ checkpoint_path, iteration))
43
+ return model, optimizer, learning_rate, iteration
44
+
45
+
46
+ def plot_spectrogram_to_numpy(spectrogram):
47
+ global MATPLOTLIB_FLAG
48
+ if not MATPLOTLIB_FLAG:
49
+ import matplotlib
50
+ matplotlib.use("Agg")
51
+ MATPLOTLIB_FLAG = True
52
+ mpl_logger = logging.getLogger('matplotlib')
53
+ mpl_logger.setLevel(logging.WARNING)
54
+ import matplotlib.pylab as plt
55
+ import numpy as np
56
+
57
+ fig, ax = plt.subplots(figsize=(10, 2))
58
+ im = ax.imshow(spectrogram, aspect="auto", origin="lower",
59
+ interpolation='none')
60
+ plt.colorbar(im, ax=ax)
61
+ plt.xlabel("Frames")
62
+ plt.ylabel("Channels")
63
+ plt.tight_layout()
64
+
65
+ fig.canvas.draw()
66
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
67
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
68
+ plt.close()
69
+ return data
70
+
71
+
72
+ def plot_alignment_to_numpy(alignment, info=None):
73
+ global MATPLOTLIB_FLAG
74
+ if not MATPLOTLIB_FLAG:
75
+ import matplotlib
76
+ matplotlib.use("Agg")
77
+ MATPLOTLIB_FLAG = True
78
+ mpl_logger = logging.getLogger('matplotlib')
79
+ mpl_logger.setLevel(logging.WARNING)
80
+ import matplotlib.pylab as plt
81
+ import numpy as np
82
+
83
+ fig, ax = plt.subplots(figsize=(6, 4))
84
+ im = ax.imshow(alignment.transpose(), aspect='auto', origin='lower',
85
+ interpolation='none')
86
+ fig.colorbar(im, ax=ax)
87
+ xlabel = 'Decoder timestep'
88
+ if info is not None:
89
+ xlabel += '\n\n' + info
90
+ plt.xlabel(xlabel)
91
+ plt.ylabel('Encoder timestep')
92
+ plt.tight_layout()
93
+
94
+ fig.canvas.draw()
95
+ data = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')
96
+ data = data.reshape(fig.canvas.get_width_height()[::-1] + (3,))
97
+ plt.close()
98
+ return data
99
+
100
+
101
+ def load_wav_to_torch(full_path):
102
+ sampling_rate, data = read(full_path)
103
+ return torch.FloatTensor(data.astype(np.float32)), sampling_rate
104
+
105
+
106
+ def load_filepaths_and_text(filename, split="|"):
107
+ with open(filename, encoding='utf-8') as f:
108
+ filepaths_and_text = [line.strip().split(split) for line in f]
109
+ return filepaths_and_text
110
+
111
+
112
+ def get_hparams(init=True):
113
+ parser = argparse.ArgumentParser()
114
+ parser.add_argument('-c', '--config', type=str, default="./configs/base.json",
115
+ help='JSON file for configuration')
116
+ parser.add_argument('-m', '--model', type=str, required=True,
117
+ help='Model name')
118
+
119
+ args = parser.parse_args()
120
+ model_dir = os.path.join("./logs", args.model)
121
+
122
+ if not os.path.exists(model_dir):
123
+ os.makedirs(model_dir)
124
+
125
+ config_path = args.config
126
+ config_save_path = os.path.join(model_dir, "config.json")
127
+ if init:
128
+ with open(config_path, "r") as f:
129
+ data = f.read()
130
+ with open(config_save_path, "w") as f:
131
+ f.write(data)
132
+ else:
133
+ with open(config_save_path, "r") as f:
134
+ data = f.read()
135
+ config = json.loads(data)
136
+
137
+ hparams = HParams(**config)
138
+ hparams.model_dir = model_dir
139
+ return hparams
140
+
141
+
142
+ def get_hparams_from_dir(model_dir):
143
+ config_save_path = os.path.join(model_dir, "config.json")
144
+ with open(config_save_path, "r") as f:
145
+ data = f.read()
146
+ config = json.loads(data)
147
+
148
+ hparams = HParams(**config)
149
+ hparams.model_dir = model_dir
150
+ return hparams
151
+
152
+
153
+ def get_hparams_from_file(config_path):
154
+ with open(config_path, "r", encoding="utf-8") as f:
155
+ data = f.read()
156
+ config = json.loads(data)
157
+
158
+ hparams = HParams(**config)
159
+ return hparams
160
+
161
+
162
+ def check_git_hash(model_dir):
163
+ source_dir = os.path.dirname(os.path.realpath(__file__))
164
+ if not os.path.exists(os.path.join(source_dir, ".git")):
165
+ logger.warn("{} is not a git repository, therefore hash value comparison will be ignored.".format(
166
+ source_dir
167
+ ))
168
+ return
169
+
170
+ cur_hash = subprocess.getoutput("git rev-parse HEAD")
171
+
172
+ path = os.path.join(model_dir, "githash")
173
+ if os.path.exists(path):
174
+ saved_hash = open(path).read()
175
+ if saved_hash != cur_hash:
176
+ logger.warn("git hash values are different. {}(saved) != {}(current)".format(
177
+ saved_hash[:8], cur_hash[:8]))
178
+ else:
179
+ open(path, "w").write(cur_hash)
180
+
181
+
182
+ def get_logger(model_dir, filename="train.log"):
183
+ global logger
184
+ logger = logging.getLogger(os.path.basename(model_dir))
185
+ logger.setLevel(logging.DEBUG)
186
+
187
+ formatter = logging.Formatter("%(asctime)s\t%(name)s\t%(levelname)s\t%(message)s")
188
+ if not os.path.exists(model_dir):
189
+ os.makedirs(model_dir)
190
+ h = logging.FileHandler(os.path.join(model_dir, filename))
191
+ h.setLevel(logging.DEBUG)
192
+ h.setFormatter(formatter)
193
+ logger.addHandler(h)
194
+ return logger
195
+
196
+
197
+ class HParams():
198
+ def __init__(self, **kwargs):
199
+ for k, v in kwargs.items():
200
+ if type(v) == dict:
201
+ v = HParams(**v)
202
+ self[k] = v
203
+
204
+ def keys(self):
205
+ return self.__dict__.keys()
206
+
207
+ def items(self):
208
+ return self.__dict__.items()
209
+
210
+ def values(self):
211
+ return self.__dict__.values()
212
+
213
+ def __len__(self):
214
+ return len(self.__dict__)
215
+
216
+ def __getitem__(self, key):
217
+ return getattr(self, key)
218
+
219
+ def __setitem__(self, key, value):
220
+ return setattr(self, key, value)
221
+
222
+ def __contains__(self, key):
223
+ return key in self.__dict__
224
+
225
+ def __repr__(self):
226
+ return self.__dict__.__repr__()