dtorch.jtensors

class jtensors.JTensors

The tensor class wrapping a numpy array to handle autograd.

grad: JTensors: Way to access the gradients of the tensors after the backward() method was called on a result of an operation on said tensors.

require_grads: bool: If set to True, the gradients of this tensor will be calculated when the backward() method is called.

ndims: int: Number of dimensions this tensor’s shape have.

shape: Tuple[int]: The shape of the tensor.

dtype: numpy.ndtype: The type of the tensor’s elements.

T: JTensors: the transpose of the current tensor.

itemsize: int: Size of each elements in bytes.

size: int: Number of elements in the tensor.

stride: int: The current stride of the tensor (stride * itemsize).

backward(base_tensor, forced=False) → None

Parameters:

base_tensor (JTensors) – The base tensor the gradients will be accumulated on top of.
forced (bool) – Only useful internaly (may be used with caution)

Backpropagate though the network to calculate the gradients of tensor linked. require_grads attribute need to be set to True to backpropagate.

numpy() → numpy.ndarray: Transform the tensor into a numpy() array.

transpose() → JTensors: Transpose the tensor.

reshape(*shape) → JTensors

Parameters:: shape (int) – The shape of the new tensor. Ex : tensor.reshape(1, 2, 3)

max() → JTensors: Return a 1 element tensor containing the highest element of the current tensor.

sum(axis: Tuple[int] | None = None, keepdims: bool = False) → JTensors

Parameters:

axis (Optional[Tuple[int]]) – The axis that will be summed. If not provided, all axis are summed.
keepdims (bool) – Keep the dimension that are summed.

Return a tensor containing the sum over specific axis or over all axis if not provided.

rearrange(pattern: str) → JTensors

Parameters:: pattern (str) – The pattern to rearrange

The method is similar to einops rearrange method.

Ex:

>> import dtorch.jtensors as dt
>> u = dt.tensor([[4, 2, 4], [5, 2, 6]])
>> dt.rearrange(u, 'ab->ba')
jtensor([[4 5] [2 2] [4 6]])

shuffle() → JTensors: Return a shuffled version of the tensor. Does not support backpropagration

norm() → JTensors: Return a one element tensor containing the norm of the tensor.

detach() → JTensors: Return a copy of the tensor that is not linked to the ancient autograd graph. Useful for TBPTT methods

clone() → JTensors: Return a clone of the tensor.

unsqueeze(dim: int) → JTensors

Parameters:: dim (int) – Add a new dimension of size 1 at dimension index given.

squeeze(dim: int) → JTensors

Parameters:: dim (int) – The dimension to squeeze. (It must be of size 1)