def __init__(self,
d_word_vec=77,
n_layers=3,
n_head=1,
d_k=16,
d_v=16,
d_model=77,
d_inner=16,
dropout=0.1,
n_position=200,
seq_len=15,
con_size=3,
days=1,
kernel='linear',
kernel_size_tcn=3,
kernel_size_scn=2):
super().__init__()
self.encoder = Encoder(d_word_vec,
n_layers,
n_head,
d_k,
d_v,
d_model,
d_inner,
dropout,
n_position,
kernel=kernel,
kernel_size_tcn=kernel_size_tcn,
kernel_size_scn=kernel_size_scn)
self.con1 = nn.Conv1d(d_model, days, con_size)
self.ff1 = nn.Linear(seq_len - con_size + 1, d_word_vec)
for p in self.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
def __init__(self,
n_src_vocab,
n_max_seq,
n_layers=2,
n_head=2,
d_word_vec=100,
d_model=100,
d_inner_hid=100,
d_k=100,
d_v=100,
dropout=0.1,
proj_share_weight=True):
super(Decepticon, self).__init__()
self.encoder = Encoder(n_src_vocab,
n_max_seq,
n_layers=n_layers,
n_head=n_head,
d_word_vec=d_word_vec,
d_model=d_model,
d_inner_hid=d_inner_hid,
dropout=dropout)
assert d_model == d_word_vec, 'To facilitate the residual connections' \
'the dimensions of all module output shall be the same.'
def __init__(self, in_channels=13, len_max_seq=100,
d_word_vec=512, d_model=512, d_inner=2048,
n_layers=6, n_head=8, d_k=64, d_v=64,
dropout=0.2, nclasses=6):
super(TransformerEncoder, self).__init__()
self.d_model = 512
self.inlayernorm = nn.LayerNorm(in_channels)
self.convlayernorm = nn.LayerNorm(d_model)
self.outlayernorm = nn.LayerNorm(d_model)
self.inconv = torch.nn.Conv1d(in_channels, d_model, 1)
self.encoder = Encoder(
n_src_vocab=None, len_max_seq=len_max_seq,
d_word_vec=d_word_vec, d_model=d_model, d_inner=d_inner,
n_layers=n_layers, n_head=n_head, d_k=d_k, d_v=d_v,
dropout=dropout)
self.outlinear = nn.Linear(d_model, nclasses, bias=False)
self.tempmaxpool = nn.MaxPool1d(len_max_seq)
self.logsoftmax = nn.LogSoftmax(dim=-1)
def __init__(self, use_postnet=True, n_spkers=1):
super(FastSpeech2, self).__init__()
### Speaker Embedding Table ###
self.use_spk_embed = hp.use_spk_embed
if self.use_spk_embed:
self.n_spkers = n_spkers
self.spk_embed_dim = hp.spk_embed_dim
self.spk_embed_weight_std = hp.spk_embed_weight_std
self.embed_speakers = Embedding(n_spkers, self.spk_embed_dim, padding_idx=None, std=self.spk_embed_weight_std)
self.use_emo_embed = hp.use_emo_embed
if self.use_emo_embed:
self.n_emotes = n_emotes
self.emo_embed_dim = hp.emo_embed_dim
self.emo_embed_weight_std = hp.emo_embed_weight_std
self.embed_emotions = Embedding(n_emotes, self.emo_embed_dim, padding_idx=None, std=self.emo_embed_weight_std)
### Encoder, Speaker Integrator, Variance Adaptor, Deocder, Postnet ###
self.encoder = Encoder()
if self.use_spk_embed:
self.speaker_integrator = SpeakerIntegrator()
self.variance_adaptor = VarianceAdaptor()
self.decoder = Decoder()
self.mel_linear = nn.Linear(hp.decoder_hidden, hp.n_mel_channels)
self.use_postnet = use_postnet
if self.use_postnet:
self.postnet = PostNet()
def __init__(self):
super(FastSpeech, self).__init__()
self.encoder = Encoder()
self.length_regulator = LengthRegulator()
self.decoder = Decoder()
self.mel_linear = Linear(hp.decoder_output_size, hp.num_mels)
self.postnet = PostNet()
def __init__(self):
super(FastSpeech2, self).__init__()
self.encoder = Encoder()
self.variance_adaptor = VarianceAdaptor()
self.decoder = Decoder()
self.mel_linear = Linear(hp.decoder_hidden, hp.n_mel_channels)
self.postnet = PostNet()
def __init__(self):
super(FastSpeech, self).__init__()
self.encoder = Encoder()
self.length_regulator = LengthRegulator()
self.decoder = Decoder()
self.mel_linear = Linear(hp.decoder_dim, hp.num_mels)
self.postnet = CBHG(hp.num_mels, K=8, projections=[256, hp.num_mels])
self.last_linear = Linear(hp.num_mels * 2, hp.num_mels)
def __init__(self):
super(StyleEncoder, self).__init__()
self.text_encoder = Encoder()
self.audio_encoder = AudioEncoder()
self.text_linear_down = nn.Sequential(
nn.Linear(hp.encoder_hidden, hp.va_neck_hidden_t), nn.ReLU())
self.speaker_linear_p = nn.Sequential(
nn.Linear(hp.speaker_embed_dim, hp.va_neck_hidden_p * 2),
nn.ReLU())
self.speaker_linear = nn.Sequential(
nn.Linear(hp.speaker_embed_dim, hp.encoder_hidden), nn.ReLU())
def __init__(self, py_vocab_size,hz_vocab_size=None, use_postnet=True):
super(FastSpeech2, self).__init__()
self.encoder = Encoder(py_vocab_size, hz_vocab_size = hz_vocab_size)
self.variance_adaptor = VarianceAdaptor()
self.decoder = Decoder()
self.mel_linear = nn.Linear(hp.decoder_hidden, hp.n_mel_channels)
self.use_postnet = use_postnet
if self.use_postnet:
self.postnet = UNet(scale=8)
def __init__(self, use_postnet=True):
super(FastSpeech2, self).__init__()
self.encoder = Encoder()
self.variance_adaptor = TacotronDuration()
self.decoder = Decoder()
self.mel_linear = nn.Linear(hp.decoder_hidden, hp.n_mel_channels)
self.use_postnet = use_postnet
if self.use_postnet:
self.postnet = PostNet()
def __init__(self, use_postnet=True):
super(FastSpeech2, self).__init__()
# self.gst = GST()
self.encoder = Encoder()
self.variance_adaptor = VarianceAdaptor()
self.decoder = Decoder()
if hp.vocoder=='WORLD':
# self.f0_decoder= Decoder()
self.ap_linear = nn.Linear(hp.decoder_hidden, hp.n_ap_channels)
self.sp_linear = nn.Linear(hp.decoder_hidden, hp.n_sp_channels)
else:
self.mel_linear = nn.Linear(hp.decoder_hidden, hp.n_mel_channels)
self.use_postnet = use_postnet
if self.use_postnet:
self.postnet = PostNet()
def __init__(self, hparams):
super().__init__()
self.hparams = hparams
hparams = self.hparams # a['key'] (if so) -> a.key
self.enc = Encoder(
n_src_vocab=hparams.vocab_size, len_max_seq=hparams.max_len,
d_word_vec=hparams.d_model, d_model=hparams.d_model,
d_inner=hparams.d_inner_hid, d_k=hparams.d_k, d_v=hparams.d_v,
n_layers=hparams.n_layers, n_head=hparams.n_head,
dropout=hparams.dropout)
self.word = nn.Linear(hparams.d_model, hparams.vocab_size, bias=False)
nn.init.xavier_normal_(self.word.weight)
self.x_logit_scale = 1.
if hparams.share_emb_prj_weight:
self.word.weight = self.enc.src_word_emb.weight
self.x_logit_scale = (hparams.d_model ** -0.5)
self.loc = nn.Linear(hparams.d_model, 1)
def __init__(self,
src_vocab_size,
tgt_vocab_size,
n_layer=6,
d_model=512,
d_ff=2048,
n_head=8,
dropout=0.1):
super(Transformer, self).__init__()
self.src_embed = nn.Sequential(Embeddings(d_model, src_vocab_size),
PositionalEncoding(d_model, dropout))
self.tgt_embed = nn.Sequential(Embeddings(d_model, tgt_vocab_size),
PositionalEncoding(d_model, dropout))
self.encoder = Encoder(n_head, d_model, d_ff, dropout, n_layer)
self.decoder = Decoder(n_head, d_model, d_ff, dropout, n_layer)
self.generator = Generator(d_model, tgt_vocab_size)
# Initialize parameters with Glorot / fan_avg.
for p in self.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
def __init__(self, d_src_vec, len_seq, d_emb_vec, n_layers, n_head, d_k,
d_v, d_inner, dropout):
super(model, self).__init__()
self.d_src_vec = d_src_vec
self.d_emb_vec = d_emb_vec
self.len_seq = len_seq
self.n_layers = n_layers
self.n_head = n_head
self.dropout = dropout
self.d_inner = d_inner
self.ffn = ffn_compressed(d_in=self.d_src_vec,
d_hid=self.d_inner,
d_out=self.d_emb_vec)
self.encoder = Encoder(len_seq=self.len_seq,
d_word_vec=self.d_emb_vec,
n_layers=self.n_layers,
n_head=self.n_head,
d_k=self.d_emb_vec // self.n_head,
d_v=self.d_emb_vec // self.n_head,
d_inner=self.d_inner,
dropout=self.dropout)
#Fully connected. Seems to have a lot of params
# self.FC1 = nn.Linear(self.d_emb_vec * self.len_seq , 64)
# self.FC2 = nn.Linear(64, 8)
# self.FC3 = nn.Linear(8, 2)
# #Average pooling over features
# self.avg_pooling = nn.AvgPool1d(d_emb_vec-1, stride=1) #d_emb_vec-1: To have 2 classes
# self.FC = nn.Linear(len_seq * 2, 2) #2: binary classification
# self.softmax = nn.Softmax(dim=-1)
#Average pooling over sequence
self.avg_pooling = nn.AvgPool1d(
len_seq, stride=1) #self.len_seq: To have 1 averaged token
self.FC = nn.Linear(d_emb_vec, 2) #2: binary classification
self.softmax = nn.Softmax(dim=-1)
import torch.nn.functional as F
import torch.optim as optim
from torchtext.data import Field, Dataset, BucketIterator
from torchtext.datasets import TranslationDataset
import transformer.Constants as Constants
from transformer.Layers import EncoderLayer
from transformer.Models import Transformer, Encoder
from transformer.Optim import ScheduledOptim
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
encoder_stacks = Encoder(d_model=32,
d_inner=64,
n_layers=2,
n_head=4,
d_k=16,
d_v=16,
dropout=0.1)
criterion = torch.nn.MSELoss().to(device)
optimizer = torch.optim.SGD(encoder_stacks.parameters(), lr=1)
src = torch.rand(1, 2, 32, requires_grad=True)
tgt = torch.rand(1, 2, 32)
print(src)
encoder_stacks.train()
for i in range(100):
def __init__(self,
source_dataset,
batch_size,
epochs,
window_size,
device,
plot_file,
train_data,
test_data,
valid_data,
target_column,
target_min,
target_max,
d_inner,
n_layers,
n_head_,
d_k,
d_v,
n_warmup_steps,
criterion,
target_name,
d_model,
model_file=None,
load_data=False,
load_model=False):
self.data_frame = self.read_dataset(source_dataset)
self.batch_size = batch_size
self.epochs = epochs
self.device = device
self.target_column = target_column
self.window = window_size
self.plot_file = plot_file
self.n_layers = n_layers
self.n_head = n_head_
self.d_inner = d_inner
self.warmup_step = n_warmup_steps
self.d_k = d_k
self.d_v = d_v
self.d_model = d_model
self.target_name = target_name
self.input_mask = torch.ones([self.batch_size, 1, self.window],
dtype=torch.int,
device=device)
self.target_max = target_max
self.target_min = target_min
self.model_file = model_file
self.prev_epoch = 0
if load_data:
self.train_df = pd.read_csv(train_data)
self.test_df = pd.read_csv(test_data)
self.valid_df = pd.read_csv(valid_data)
else:
self.train_df, self.valid_df, self.test_df = self.organize_dataset(
train_data, test_data, valid_data)
pad_col = [
'col' + str(i) for i in range(self.train_df.shape[1], self.d_model)
]
for col in pad_col:
self.train_df[col] = 0
self.test_df[col] = 0
self.valid_df[col] = 0
self.columns = self.train_df.shape[1]
self.model = Encoder(n_position=200,
d_word_vec=self.columns,
d_model=self.columns,
d_inner=d_inner,
n_layers=n_layers,
n_head=n_head_,
d_k=d_k,
d_v=d_v,
dropout=0).to(device)
if load_model:
self.model = torch.load(self.model_file)['model']
self.model.eval()
self.model = self.model.to(device)
self.prev_epoch = torch.load(self.model_file)['epoch']
self.criterion = criterion
self.optimizer = ScheduledOptim(
optim.Adam(self.model.parameters(), betas=(0.9, 0.98), eps=1e-09),
2.0,
self.columns,
n_warmup_steps,
n_step=self.prev_epoch * (math.floor(
self.train_df.shape[0] / self.window * self.batch_size)))
self.loss_list = []
self.lr_list = []
return mel_output, mel_output_postnet, duration_predictor_output
else:
length_regulator_output, decoder_pos = self.length_regulator(
encoder_output, encoder_mask, alpha=alpha)
decoder_output = self.decoder(length_regulator_output, decoder_pos)
mel_output = self.mel_linear(decoder_output)
mel_output_postnet = self.postnet(mel_output) + mel_output
return mel_output, mel_output_postnet
if __name__ == "__main__":
# Test
test_encoder = Encoder()
test_decoder = Decoder()
# print(test_encoder)
# print(test_decoder)
test_src = torch.stack([
torch.Tensor([1, 2, 4, 3, 2, 5, 0, 0]),
torch.Tensor([3, 4, 2, 6, 7, 1, 2, 3])
]).long()
test_pos = torch.stack([
torch.Tensor([1, 2, 3, 4, 5, 6, 0, 0]),
torch.Tensor([1, 2, 3, 4, 5, 6, 7, 8])
]).long()
test_target = torch.stack([
torch.Tensor([0, 2, 3, 0, 3, 2, 1, 0]),
torch.Tensor([1, 2, 3, 2, 2, 0, 3, 6])