unary¶

The Ops that have a single input and output are called Unary Ops. Usually this means that an Operation is applied elementwise.

This file contains all of the unary operations in Tricycle.

class Batch[source]¶

Bases: Op

A class for batching a tensor.

back_fn(grad)[source]¶

Compute the gradient for the batch operation.

Parameters:: grad (Tensor) – The gradient tensor.
Returns:: The computed gradient.
Return type:: Tensor

forward(tensor)[source]¶

Mark a tensor as batched.

Parameters:: tensor (Tensor) – The input tensor.
Returns:: A new Tensor marked as batched.
Return type:: Tensor

is_batched: bool¶

class UnaryAdd[source]¶

Bases: Op

A class for adding a constant to a tensor elementwise.

forward(tensor, constant)[source]¶

Add a constant, elementwise, to a tensor. The constant is not differentiable.

Parameters:

tensor (Tensor) – The input tensor.
constant (float) – The constant to add.

Returns:

A new Tensor with the constant added elementwise.

Raises:

AssertionError – If tensor is not a Tensor or constant is not a number.

Return type:

Tensor

class UnaryCos[source]¶

Bases: Op

A class for computing the cosine of a tensor elementwise.

back_fn(grad)[source]¶

Compute the gradient for the cosine operation.

Parameters:: grad (Tensor) – The gradient tensor.
Returns:: The computed gradient.
Return type:: Tensor

forward(tensor)[source]¶

Applies the cosine function, elementwise, to a tensor.

Parameters:: tensor (Tensor) – The input tensor.
Returns:: A new Tensor with the cosine of each element.
Return type:: Tensor

class UnaryDivide[source]¶

Bases: Op

A class for dividing a constant by a tensor elementwise.

forward(constant, tensor)[source]¶

Divide a constant by a tensor, elementwise. The constant is not differentiable.

Parameters:

constant (float | int) – The constant numerator.
tensor (Tensor) – The tensor denominator.

Returns:

A new Tensor with the constant divided by each element of the tensor.

Return type:

Tensor

class UnaryExp[source]¶

Bases: Op

A class for computing the exponential of a tensor elementwise.

back_fn(grad)[source]¶

Compute the gradient for the exponential operation.

Parameters:: grad (Tensor) – The gradient tensor.
Returns:: The computed gradient.
Return type:: Tensor

forward(tensor)[source]¶

Raise every element of a tensor to the power of e.

Parameters:: tensor (Tensor) – The input tensor.
Returns:: A new Tensor with the exponential of each element.
Return type:: Tensor

class UnaryLog[source]¶

Bases: Op

A class for computing the natural logarithm of a tensor elementwise.

REALLY_SMALL_NUMBER = 1e-06¶

back_fn(grad)[source]¶

Compute the gradient for the logarithm operation.

Parameters:: grad (Tensor) – The gradient tensor.
Returns:: The computed gradient.
Return type:: Tensor

forward(tensor)[source]¶

Compute the natural logarithm of each element in the tensor.

Parameters:: tensor (Tensor) – The input tensor.
Returns:: A new Tensor with the natural logarithm of each element.
Return type:: Tensor

class UnaryMask[source]¶

Bases: Op

A class for applying a binary mask to a tensor.

back_fn(grad)[source]¶

Compute the gradient for the mask operation.

Parameters:: grad (Tensor) – The gradient tensor.
Returns:: The computed gradient.
Return type:: Tensor

forward(tensor, mask)[source]¶

Apply a binary mask to a numpy array, setting values to 0 where the mask is True.

Parameters:

tensor (Tensor) – The input tensor.
mask (Tensor) – The binary mask tensor.

Returns:

A new Tensor with the mask applied.

Raises:

AssertionError – If tensor and mask shapes don’t match or if mask requires grad.

Return type:

Tensor

class UnaryMax[source]¶

Bases: Op

A class for computing the elementwise maximum of a tensor and a constant.

back_fn(grad)[source]¶

Compute the gradient for the maximum operation.

Parameters:: grad (Tensor) – The gradient tensor.
Returns:: The computed gradient.
Return type:: Tensor

forward(tensor, constant)[source]¶

Return the max of the tensor and the constant, elementwise. The constant is not differentiable.

Parameters:

tensor (Tensor) – The input tensor.
constant (float) – The constant to compare against.

Returns:

A new Tensor with the elementwise maximum of the input and the constant.

Raises:

AssertionError – If tensor is not a Tensor or constant is not a scalar.

Return type:

Tensor

class UnaryMin[source]¶

Bases: Op

A class for computing the elementwise minimum of a tensor and a constant.

back_fn(grad)[source]¶

Compute the gradient for the minimum operation.

Parameters:: grad (Tensor) – The gradient tensor.
Returns:: The computed gradient.
Return type:: Tensor

forward(tensor, constant)[source]¶

Return the min of the tensor and the constant, elementwise. The constant is not differentiable.

Parameters:

tensor (Tensor) – The input tensor.
constant (float) – The constant to compare against.

Returns:

A new Tensor with the elementwise minimum of the input and the constant.

Raises:

AssertionError – If tensor is not a Tensor or constant is not a scalar.

Return type:

Tensor

class UnaryMultiply[source]¶

Bases: Op

A class for multiplying a tensor by a constant elementwise.

back_fn(grad)[source]¶

Compute the gradient for the multiplication operation.

Parameters:: grad (Tensor) – The gradient tensor.
Returns:: The computed gradient.
Return type:: Tensor

forward(tensor, constant)[source]¶

Multiply a constant, elementwise, to a tensor. The constant is not differentiable.

Parameters:

tensor (Tensor) – The input tensor.
constant (float) – The constant to multiply.

Returns:

A new Tensor with the constant multiplied elementwise.

Raises:

AssertionError – If tensor is not a Tensor or constant is not a scalar.

Return type:

Tensor

class UnaryPower[source]¶

Bases: Op

A class for raising a tensor to a constant power elementwise.

back_fn(grad)[source]¶

Compute the gradient for the power operation.

Parameters:: grad (Tensor) – The gradient tensor.
Returns:: The computed gradient.
Return type:: Tensor

forward(tensor, constant)[source]¶

Raise a tensor to a constant, elementwise. The constant is not differentiable.

Parameters:

tensor (Tensor) – The input tensor.
constant (float) – The constant power.

Returns:

A new Tensor with each element raised to the constant power.

Raises:

AssertionError – If tensor is not a Tensor or constant is not a scalar.

Return type:

Tensor

class UnarySin[source]¶

Bases: Op

A class for computing the sine of a tensor elementwise.

back_fn(grad)[source]¶

Compute the gradient for the sine operation.

Parameters:: grad (Tensor) – The gradient tensor.
Returns:: The computed gradient.
Return type:: Tensor

forward(tensor)[source]¶

Applies the sine function, elementwise, to a tensor.

Parameters:: tensor (Tensor) – The input tensor.
Returns:: A new Tensor with the sine of each element.
Return type:: Tensor

class UnarySquareRoot[source]¶

Bases: Op

A class for computing the square root of a tensor elementwise.

forward(tensor)[source]¶

Apply the square root function.

Parameters:: tensor (Tensor) – The input tensor.
Returns:: A new Tensor with the square root of each element.

class UnarySubtract[source]¶

Bases: Op

A class for subtracting a constant from a tensor elementwise.

forward(tensor, constant)[source]¶

Subtract a constant, elementwise, from a tensor. The constant is not differentiable.

Parameters:

tensor (Tensor) – The input tensor.
constant (float) – The constant to subtract.

Returns:

A new Tensor with the constant subtracted elementwise.

Return type:

Tensor

class UnarySum[source]¶

Bases: Op

A class for summing all values in a tensor.

back_fn(grad)[source]¶

Compute the gradient for the sum operation.

Parameters:: grad (Tensor) – The gradient tensor.
Returns:: The computed gradient.
Raises:: AssertionError – If grad is not a scalar or an empty tensor.
Return type:: Tensor

forward(tensor)[source]¶

Sums all the values in a tensor. Note, this function always produces a non-batched output.

Parameters:: tensor (Tensor) – The input tensor.
Returns:: A new Tensor with the sum of all elements.
Raises:: AssertionError – If the sum operation fails.
Return type:: Tensor

class Unbatch[source]¶

Bases: Op

A class for unbatching a tensor.

back_fn(grad)[source]¶

Compute the gradient for the unbatch operation.

Parameters:: grad (Tensor) – The gradient tensor.
Returns:: The computed gradient.
Return type:: Tensor

forward(tensor)[source]¶

Mark a tensor as unbatched.

Parameters:: tensor (Tensor) – The input tensor.
Returns:: A new Tensor marked as unbatched.
Return type:: Tensor

is_batched: bool¶

batch(tensor)[source]¶

Tell Tricycle to treat this tensor as a batch of tensors.

Parameters:: tensor (Tensor) – The input tensor.
Returns:: A new Tensor marked as batched.
Return type:: Tensor

nothing(tensor)[source]¶

Do nothing to a tensor.

Operations like addition have a back_fn of multiplying by 1 which is equivalent to doing nothing. Instead of adding a bunch of checks to see whether a back_fn exists or not, the logic is much simpler by adding this function that does nothing.

Parameters:: tensor – The input tensor.
Returns:: The input tensor unchanged.

unbatch(tensor)[source]¶

Tell Tricycle to treat this tensor as a single tensor (not a batch of tensors).

Parameters:: tensor (Tensor) – The input tensor.
Returns:: A new Tensor marked as unbatched.
Return type:: Tensor