torchbox.module.dsp package

Submodules

torchbox.module.dsp.convolution module

class torchbox.module.dsp.convolution.Conv1(axis, in_channels, out_channels, kernel_size=3, stride=1, padding=0, dilation=1, groups=1, bias=True, padding_mode='zeros')

Bases: torch.nn.modules.module.Module

forward(X)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class torchbox.module.dsp.convolution.Conv2(in_channels, out_channels, kernel_size=3, stride=1, padding=0, dilation=1, groups=1, bias=True, padding_mode='zeros')

Bases: torch.nn.modules.module.Module

forward(X)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class torchbox.module.dsp.convolution.FFTConv1(nh, h=None, shape='same', nfft=None, train=True, **kwargs)

Bases: torch.nn.modules.module.Module

forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class torchbox.module.dsp.convolution.MaxPool1(axis, kernel_size, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False)

Bases: torch.nn.modules.module.Module

forward(X)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

class torchbox.module.dsp.convolution.MaxPool2(kernel_size, stride=None, padding=0, dilation=1, return_indices=False, ceil_mode=False)

Bases: torch.nn.modules.module.Module

forward(X)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

torchbox.module.dsp.interpolation module

class torchbox.module.dsp.interpolation.Interp1(*args, **kwargs)

Bases: torch.autograd.function.Function

static backward(ctx, grad_out)

Defines a formula for differentiating the operation with backward mode automatic differentiation (alias to the vjp function).

This function is to be overridden by all subclasses.

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computated w.r.t. the output.

forward(x, y, xnew, out=None)

Linear 1D interpolation on the GPU for Pytorch. This function returns interpolated values of a set of 1-D functions at the desired query points xnew. This function is working similarly to Matlab™ or scipy functions with the linear interpolation mode on, except that it parallelises over any number of desired interpolation problems. The code will run on GPU if all the tensors provided are on a cuda device. :param x: A 1-D or 2-D tensor of real values. :type x: (N, ) or (D, N) Pytorch Tensor :param y: A 1-D or 2-D tensor of real values. The length of y along its

last dimension must be the same as that of x

Parameters

  • xnew ((P,) or (D, P) Pytorch Tensor) – A 1-D or 2-D tensor of real values. xnew can only be 1-D if _both_ x and y are 1-D. Otherwise, its length along the first dimension must be the same as that of whichever x and y is 2-D.

  • out (Pytorch Tensor, same shape as xnew) – Tensor for the output. If None: allocated automatically.

torchbox.module.dsp.polynomialfit module

class torchbox.module.dsp.polynomialfit.PolyFit(w=None, deg=1, trainable=True)

Bases: torch.nn.modules.module.Module

Polynominal fitting

We fit the data using a polynomial function of the form

\[y(x, {\mathbf w})=w_{0}+w_{1} x+w_{2} x^{2}+, \cdots,+w_{M} x^{M}=\sum_{j=0}^{M} w_{j} x^{j} \]
Parameters

  • w (Tensor, optional) – initial coefficient, by default None (generate randomly)

  • deg (int, optional) – degree of the Polynominal, by default 1

  • trainable (bool, optional) – is self.w trainable, by default True

Examples

th.manual_seed(2020)
Ns, k, b = 100, 1.2, 3.0
x = th.linspace(0, 1, Ns)
t = x * k + b + th.randn(Ns)

deg = (0, 1)

polyfit = PolyFit(deg=deg)

lossfunc = th.nn.MSELoss('mean')
optimizer = th.optim.Adam(filter(lambda p: p.requires_grad, polyfit.parameters()), lr=1e-1)

for n in range(100):
    y = polyfit(x)

    loss = lossfunc(y, t)

    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

    print("---Loss %.4f, %.4f, %.4f" % (loss.item(), polyfit.w[0], polyfit.w[1]))

# output
---Loss 16.7143, -0.2315, -0.1427
---Loss 15.5265, -0.1316, -0.0429
---Loss 14.3867, -0.0319, 0.0568
---Loss 13.2957, 0.0675, 0.1561
---Loss 12.2543, 0.1664, 0.2551
                ...
---Loss 0.9669, 2.4470, 1.9995
---Loss 0.9664, 2.4515, 1.9967
---Loss 0.9659, 2.4560, 1.9938
forward(x)

Defines the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Module contents