def test_sigmoid_grad(N=None):
from activations import Sigmoid
N = np.inf if N is None else N
mine = Sigmoid()
gold = torch_gradient_generator(torch.sigmoid)
i = 0
while i < N:
n_ex = np.random.randint(1, 100)
n_dims = np.random.randint(1, 100)
z = random_tensor((n_ex, n_dims))
assert_almost_equal(mine.grad(z), gold(z))
print("PASSED")
i += 1
class WavenetResidualModule(ModuleBase):
def __init__(
self,
ch_residual,
ch_dilation,
dilation,
kernel_width,
optimizer=None,
init="glorot_uniform",
):
"""
A WaveNet-like residual block with causal dilated convolutions.
*Skip path in* >-------------------------------------------> + --------> *Skip path out*
Causal |--> Tanh --| |
*Main |--> Dilated Conv1D -| * --> 1x1 Conv1D --|
path >--| |--> Sigm --| |
in* |-------------------------------------------------> + --------> *Main path out*
*Residual path*
On the final block, the output of the skip path is further processed to
produce the network predictions.
See van den Oord et al. (2016) at https://arxiv.org/pdf/1609.03499.pdf
for further details.
Parameters
----------
ch_residual : int
The number of output channels for the 1x1 Conv1D layer in the main
path
ch_dilation : int
The number of output channels for the causal dilated Conv1D layer
in the main path
dilation : int
The dilation rate for the causal dilated Conv1D layer in the main
path
kernel_width : int
The width of the causal dilated Conv1D kernel in the main path
init : str (default: 'glorot_uniform')
The weight initialization strategy. Valid entries are
{'glorot_normal', 'glorot_uniform', 'he_normal', 'he_uniform'}
optimizer : str or `OptimizerBase` instance (default: None)
The optimization strategy to use when performing gradient updates
within the `update` method. If `None`, use the `SGD` optimizer with
default parameters.
"""
super().__init__()
self.init = init
self.dilation = dilation
self.optimizer = optimizer
self.ch_residual = ch_residual
self.ch_dilation = ch_dilation
self.kernel_width = kernel_width
self._init_params()
def _init_params(self):
self._dv = {}
self.conv_dilation = Conv1D(
stride=1,
pad="causal",
init=self.init,
kernel_width=2,
dilation=self.dilation,
out_ch=self.ch_dilation,
optimizer=self.optimizer,
act_fn=Affine(slope=1, intercept=0),
)
self.tanh = Tanh()
self.sigm = Sigmoid()
self.multiply_gate = Multiply(act_fn=Affine(slope=1, intercept=0))
self.conv_1x1 = Conv1D(
stride=1,
pad="same",
dilation=0,
init=self.init,
kernel_width=1,
out_ch=self.ch_residual,
optimizer=self.optimizer,
act_fn=Affine(slope=1, intercept=0),
)
self.add_residual = Add(act_fn=Affine(slope=1, intercept=0))
self.add_skip = Add(act_fn=Affine(slope=1, intercept=0))
@property
def parameters(self):
return {
"components": {
"conv_1x1": self.conv_1x1.parameters,
"add_skip": self.add_skip.parameters,
"add_residual": self.add_residual.parameters,
"conv_dilation": self.conv_dilation.parameters,
"multiply_gate": self.multiply_gate.parameters,
}
}
@property
def hyperparameters(self):
return {
"layer":
"WavenetResidualModule",
"init":
self.init,
"dilation":
self.dilation,
"optimizer":
self.optimizer,
"ch_residual":
self.ch_residual,
"ch_dilation":
self.ch_dilation,
"kernel_width":
self.kernel_width,
"component_ids": [
"conv_1x1",
"add_skip",
"add_residual",
"conv_dilation",
"multiply_gate",
],
"components": {
"conv_1x1": self.conv_1x1.hyperparameters,
"add_skip": self.add_skip.hyperparameters,
"add_residual": self.add_residual.hyperparameters,
"conv_dilation": self.conv_dilation.hyperparameters,
"multiply_gate": self.multiply_gate.hyperparameters,
},
}
@property
def derived_variables(self):
dv = {
"conv_1x1_out": None,
"conv_dilation_out": None,
"multiply_gate_out": None,
"components": {
"conv_1x1": self.conv_1x1.derived_variables,
"add_skip": self.add_skip.derived_variables,
"add_residual": self.add_residual.derived_variables,
"conv_dilation": self.conv_dilation.derived_variables,
"multiply_gate": self.multiply_gate.derived_variables,
},
}
dv.update(self._dv)
return dv
@property
def gradients(self):
return {
"components": {
"conv_1x1": self.conv_1x1.gradients,
"add_skip": self.add_skip.gradients,
"add_residual": self.add_residual.gradients,
"conv_dilation": self.conv_dilation.gradients,
"multiply_gate": self.multiply_gate.gradients,
}
}
def forward(self, X_main, X_skip=None):
self.X_main, self.X_skip = X_main, X_skip
conv_dilation_out = self.conv_dilation.forward(X_main)
tanh_gate = self.tanh.fn(conv_dilation_out)
sigm_gate = self.sigm.fn(conv_dilation_out)
multiply_gate_out = self.multiply_gate.forward([tanh_gate, sigm_gate])
conv_1x1_out = self.conv_1x1.forward(multiply_gate_out)
# if this is the first wavenet block, initialize the "previous" skip
# connection sum to 0
self.X_skip = np.zeros_like(conv_1x1_out) if X_skip is None else X_skip
Y_skip = self.add_skip.forward([X_skip, conv_1x1_out])
Y_main = self.add_residual.forward([X_main, conv_1x1_out])
self._dv["tanh_out"] = tanh_gate
self._dv["sigm_out"] = sigm_gate
self._dv["conv_dilation_out"] = conv_dilation_out
self._dv["multiply_gate_out"] = multiply_gate_out
self._dv["conv_1x1_out"] = conv_1x1_out
return Y_main, Y_skip
def backward(self, dY_skip, dY_main=None):
dX_skip, dConv_1x1_out = self.add_skip.backward(dY_skip)
# if this is the last wavenet block, dY_main will be None. if not,
# calculate the error contribution from dY_main and add it to the
# contribution from the skip path
dX_main = np.zeros_like(self.X_main)
if dY_main is not None:
dX_main, dConv_1x1_main = self.add_residual.backward(dY_main)
dConv_1x1_out += dConv_1x1_main
dMultiply_out = self.conv_1x1.backward(dConv_1x1_out)
dTanh_out, dSigm_out = self.multiply_gate.backward(dMultiply_out)
conv_dilation_out = self.derived_variables["conv_dilation_out"]
dTanh_in = dTanh_out * self.tanh.grad(conv_dilation_out)
dSigm_in = dSigm_out * self.sigm.grad(conv_dilation_out)
dDilation_out = dTanh_in + dSigm_in
conv_back = self.conv_dilation.backward(dDilation_out)
dX_main += conv_back
self._dv["dLdTanh"] = dTanh_out
self._dv["dLdSigmoid"] = dSigm_out
self._dv["dLdConv_1x1"] = dConv_1x1_out
self._dv["dLdMultiply"] = dMultiply_out
self._dv["dLdConv_dilation"] = dDilation_out
return dX_main, dX_skip
