def test(self):
t = np.linspace(-10, 10, num=100)
target = np.vectorize(self.func1)(t)
wn = Wavenet(Morlet, 30, t, target)
wn.smart_train(t, target, maxiter=500)
plb.plot(t, target, label='Сигнал')
plb.plot(t, wn.sim(t), label='Аппроксимация')
plb.show()
def test(self):
t = np.linspace(-10, 10, num=100)
target = np.vectorize(self.func1)(t)
wn = Wavenet(Morlet, 30, t, target)
wn.smart_train(t, target, maxiter=500)
plb.plot(t, target, label='Сигнал')
plb.plot(t, wn.sim(t), label='Аппроксимация')
plb.show()
class WaveVAE(nn.Module):
def __init__(self):
super(WaveVAE, self).__init__()
self.encoder = Wavenet(out_channels=2,
num_blocks=2,
num_layers=10,
residual_channels=128,
gate_channels=256,
skip_channels=128,
kernel_size=2,
cin_channels=80,
upsample_scales=[16, 16])
self.decoder = Wavenet_Student(num_blocks_student=[1, 1, 1, 1, 1, 1],
num_layers=10)
self.log_eps = nn.Parameter(torch.zeros(1))
def forward(self, x, c):
# Encode
mu_logs = self.encoder(x, c)
mu = mu_logs[:, 0:1, :-1]
logs = mu_logs[:, 1:, :-1]
q_0 = Normal(mu.new_zeros(mu.size()), mu.new_ones(mu.size()))
mean_q = (x[:, :, 1:] - mu) * torch.exp(-logs)
# Reparameterization, Sampling from prior
z = q_0.sample() * torch.exp(self.log_eps) + mean_q
z_prior = q_0.sample()
z = F.pad(z, pad=(1, 0), mode='constant', value=0)
z_prior = F.pad(z_prior, pad=(1, 0), mode='constant', value=0)
c_up = self.encoder.upsample(c)
# Decode
# x_rec : [B, 1, T] (first time step zero-padded)
# mu_tot, logs_tot : [B, 1, T-1]
x_rec, mu_p, log_p = self.decoder(z, c_up)
x_prior = self.decoder.generate(z_prior, c_up)
loss_recon = -0.5 * (-log(2.0 * pi) - 2. * log_p -
torch.pow(x[:, :, 1:] - mu_p, 2) * torch.exp(
(-2.0 * log_p)))
loss_kl = 0.5 * (mean_q**2 + torch.exp(self.log_eps)**2 -
1) - self.log_eps
return x_rec, x_prior, loss_recon.mean(), loss_kl.mean()
def generate(self, z, c):
c_up = self.encoder.upsample(c)
x_sample = self.decoder.generate(z, c_up)
return x_sample
def build_model():
model = Wavenet(out_channels=2,
num_blocks=args.num_blocks,
num_layers=args.num_layers,
residual_channels=args.residual_channels,
gate_channels=args.gate_channels,
skip_channels=args.skip_channels,
kernel_size=args.kernel_size,
cin_channels=args.cin_channels,
upsample_scales=[16, 16])
print(model)
n_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print('number of teacher parameters:', n_params)
return model
def __init__(self):
super(WaveVAE, self).__init__()
self.encoder = Wavenet(out_channels=2,
num_blocks=2,
num_layers=10,
residual_channels=128,
gate_channels=256,
skip_channels=128,
kernel_size=2,
cin_channels=80,
upsample_scales=[16, 16])
self.decoder = Wavenet_Student(num_blocks_student=[1, 1, 1, 1, 1, 1],
num_layers=10)
self.log_eps = nn.Parameter(torch.zeros(1))
def __init__(self, q_channels, dil_width, res_width, skip_width,
kernel_width, n_dilations, n_blocks, learning_rate, log):
'''
Pre-processes audio, calls Wavenet to create the networks, creates the targets,
prepares the loss calculation, optimizer and weights saver
Args:
q_channels: number of quantization levels
dil_width: width of dilation convolutional filters
res_width: width of residual convolutional filters
skip_width: width of skip channels for the softmax output
kernel_width: filter with, first dimension of of kernel and gate dilation component
n_dilations: Number of dilations to be used in every block
n_blocks: Number of blocks in the network
learning_rate: Learning rate to be used
log: callback to log method
'''
self.q_channels = q_channels
self.dilations = n_blocks * [2**x for x in range(n_dilations)]
self.receptive_field = (kernel_width - 1) * sum(
self.dilations) + kernel_width
net = Wavenet(dilations=self.dilations,
kernel_width=kernel_width,
dilation_width=dil_width,
residual_width=res_width,
skip_width=skip_width,
q_channels=self.q_channels,
receptive_field=self.receptive_field,
log=log)
input_batch = tf.placeholder(dtype=tf.float32, shape=None)
one_hot = one_hot_mu_law_encode(input_batch, self.q_channels)
log('Constructed wavenet.')
raw_output = net.construct_network(one_hot)
target_output = construct_target_output(one_hot, self.q_channels,
self.receptive_field)
log('Constructed targets.')
loss = tf.nn.softmax_cross_entropy_with_logits(logits=tf.reshape(
raw_output, [-1, self.q_channels]),
labels=target_output)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
log('Created optimizer.')
trainable = tf.trainable_variables()
self.loss = tf.reduce_mean(loss)
self.optim = optimizer.minimize(self.loss, var_list=trainable)
self.saver = tf.train.Saver(var_list=trainable)
self.input_batch = input_batch
self.net = net
def build_model():
model_t = Wavenet(out_channels=2,
num_blocks=args.num_blocks_t,
num_layers=args.num_layers_t,
residual_channels=args.residual_channels,
gate_channels=args.gate_channels,
skip_channels=args.skip_channels,
kernel_size=args.kernel_size,
cin_channels=args.cin_channels,
upsample_scales=[16, 16])
return model_t
def build_model(): #build wavenet object with arguments set in parser
model = Wavenet(out_channels=2,
num_blocks=args.num_blocks,
num_layers=args.num_layers,
residual_channels=args.residual_channels,
gate_channels=args.gate_channels,
skip_channels=args.skip_channels,
kernel_size=args.kernel_size,
cin_channels=args.cin_channels,
upsample_scales=[16, 16])
return model
def init_models(self):
"""
instantiate the requested models and sent them to the device
"""
# traditionals
if self.nn_arch == 'conv-trad':
self.models = {'cnn': ConvNetTrad(self.n_classes, self.data_size)}
elif self.nn_arch == 'conv-fstride':
self.models = {
'cnn': ConvNetFstride4(self.n_classes, self.data_size)
}
elif self.nn_arch == 'conv-experimental':
self.models = {
'cnn': ConvNetExperimental(self.n_classes, self.data_size)
}
# adversarials
elif self.nn_arch == 'adv-experimental':
self.models = {
'g': G_experimental(self.n_classes, self.data_size),
'd': D_experimental(self.n_classes, self.data_size)
}
elif self.nn_arch == 'adv-experimental3':
self.models = {
'g': G_experimental(self.n_classes, self.data_size),
'd': D_experimental(self.n_classes, self.data_size, out_dim=1)
}
elif self.nn_arch == 'adv-collected-encoder':
self.models = {
'g':
G_experimental(self.n_classes,
self.data_size,
net_class='label-collect-encoder'),
'd':
D_experimental(self.n_classes,
self.data_size,
net_class='label-collect-encoder')
}
# cnns
elif self.nn_arch == 'conv-encoder':
self.models = {
'cnn':
ConvEncoderClassifierNet(self.n_classes,
self.data_size,
net_class='label-collect-encoder',
fc_layer_type='fc1')
}
elif self.nn_arch == 'conv-encoder-stacked':
self.models = {
'cnn':
ConvStackedEncodersNet(self.n_classes, self.data_size,
self.encoder_model)
}
elif self.nn_arch == 'conv-latent':
self.models = {
'cnn': ConvLatentClassifier(self.n_classes, self.data_size)
}
# selected fc
elif self.nn_arch == 'conv-encoder-fc1':
self.models = {
'cnn':
ConvEncoderClassifierNet(self.n_classes,
self.data_size,
net_class='lim-encoder-6',
fc_layer_type='fc1')
}
elif self.nn_arch == 'conv-encoder-fc3':
self.models = {
'cnn':
ConvEncoderClassifierNet(self.n_classes,
self.data_size,
net_class='lim-encoder-6',
fc_layer_type='fc3')
}
# limited encoders adv
elif self.nn_arch == 'adv-lim-encoder':
self.models = {
'g':
G_experimental(self.n_classes,
self.data_size,
net_class='lim-encoder'),
'd':
D_experimental(self.n_classes,
self.data_size,
net_class='lim-encoder')
}
elif self.nn_arch == 'adv-lim-encoder-6':
self.models = {
'g':
G_experimental(self.n_classes,
self.data_size,
net_class='lim-encoder-6'),
'd':
D_experimental(self.n_classes,
self.data_size,
net_class='lim-encoder-6')
}
# limited encoder conv
elif self.nn_arch == 'conv-lim-encoder':
self.models = {
'cnn':
ConvEncoderClassifierNet(self.n_classes,
self.data_size,
net_class='lim-encoder')
}
# wavenet
elif self.nn_arch == 'wavenet':
self.models = {'wav': Wavenet(self.n_classes)}
# not found
else:
print("***Network Architecture not found!")
# send models to device
self.models = dict(
(k, model.to(self.device)) for k, model in self.models.items())
transform=transform)
validation_set = Testset(np.array([2]),
'ccmixter3/',
pad=field,
dilations1=dilations1,
device=device)
loadtr = data.DataLoader(training_set,
batch_size=1,
shuffle=True,
num_workers=0,
worker_init_fn=np.random.seed)
loadval = data.DataLoader(validation_set, batch_size=1, num_workers=0)
# In[6]:
#model = Unet(skipDim, quantization_channels, residualDim,device)
model = Wavenet(field, skipDim, residualDim, dilations0, dilations1)
#model = nn.DataParallel(model)
model = model.cuda()
criterion = nn.CrossEntropyLoss()
# in wavenet paper, they said crossentropyloss is far better than MSELoss
optimizer = optim.Adam(model.parameters(), lr=1e-3, weight_decay=1e-5)
# use adam to train
maxloss = np.zeros(50) + 100
# In[7]:
start_epoch = 0
if continueTrain: # if continueTrain, the program will find the checkpoints
if os.path.isfile(resumefile):
print("=> loading checkpoint '{}'".format(resumefile))
checkpoint = torch.load(resumefile)
batch_loss = 0
print('--Training finished')
if __name__ == '__main__':
"""
main
"""
# input [num x batch x channel x time]
x_train = torch.randn(1, 1, 1, 16000)
y_train = torch.randint(0, 2, (1, 1, 16000))
class_dict = {'a':0, 'b':1}
# norm
x_train = x_train / torch.max(torch.abs(x_train))
# wavenet
wavenet = Wavenet()
print("wavenet: ", wavenet)
# wavenet training
wavenet_training(wavenet, x_train, y_train, class_dict, num_epochs=10, lr=0.001)
# count params
layer_params = wavenet.count_params()
print("layer: {} sum: {}".format(layer_params, sum(layer_params)))
kwargs = {'num_workers': 1, 'pin_memory': True} if use_cuda else {}
# In[5]:
params = {'batch_size': 1, 'shuffle': True, 'num_workers': 1}
training_set = Dataset(np.arange(0, songnum), np.arange(0, songnum),
'ccmixter2/x/', 'ccmixter2/y/')
validation_set = Dataset(np.arange(0, songnum), np.arange(0, songnum),
'ccmixter2/x/', 'ccmixter2/y/')
loadtr = data.DataLoader(training_set,
**params) # pytorch dataloader, more faster than mine
loadval = data.DataLoader(validation_set, **params)
# In[6]:
model = Wavenet(pad, skipDim, quantization_channels, residualDim,
dilations).cuda()
criterion = nn.CrossEntropyLoss()
# in wavenet paper, they said crossentropyloss is far better than MSELoss
# optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum)
optimizer = optim.Adam(model.parameters(), lr=1e-3, weight_decay=1e-5)
# use adam to train
# optimizer = optim.SGD(model.parameters(), lr = 0.1, momentum=0.9, weight_decay=1e-5)
# scheduler = StepLR(optimizer, step_size=30, gamma=0.1)
# scheduler = MultiStepLR(optimizer, milestones=[20,40], gamma=0.1)
# In[7]:
if continueTrain: # if continueTrain, the program will find the checkpoints
if os.path.isfile(resumefile):
print("=> loading checkpoint '{}'".format(resumefile))
checkpoint = torch.load(resumefile)