Tensor

class nvtripy.Tensor(data: Any, dtype: dtype | None = None, device: device | None = None, name: str | None = None, fetch_stack_info: bool = True)[source]

Bases: object

A tensor is a multi-dimensional array that contains elements of a uniform data type.

Parameters:

  • data (Any) – The data with which to initialize the tensor.

  • dtype (dtype | None) – The data type of the tensor.

  • device (device | None) – The device on which to allocate the tensor.

  • name (str | None) – The name of the tensor. If provided, this must be a unique string.

  • fetch_stack_info (bool) – Whether to fetch stack information for the tensor. Stack information allows Tripy to generate much higher quality error messages at the cost of a small overhead when initializing the tensor.

Example
1tensor = tp.Tensor([1.0, 2.0, 3.0], dtype=tp.float32)
Local Variables
>>> tensor
tensor([1.0000, 2.0000, 3.0000], dtype=float32, loc=gpu:0, shape=(3,))
__add__(other: Tensor | Number) → Tensor

Performs an elementwise sum.

Parameters:

  • self (Tensor) – Tensor to be added to other.

  • other (Tensor | Number) – The tensor to add to this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([1, 2])
2b = tp.Tensor([2, 3])
3output = a + b
Local Variables
>>> a
tensor([1, 2], dtype=int32, loc=gpu:0, shape=(2,))

>>> b
tensor([2, 3], dtype=int32, loc=gpu:0, shape=(2,))

>>> output
tensor([3, 5], dtype=int32, loc=gpu:0, shape=(2,))
__eq__(other: Tensor | Number) → Tensor

Performs an ‘equal’ comparison.

Parameters:

  • self (Tensor) – Tensor to be compared with other.

  • other (Tensor | Number) – The tensor to be compared to this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([2, 3])
2b = tp.Tensor([2, 5])
3output = b == a
Local Variables
>>> a
tensor([2, 3], dtype=int32, loc=gpu:0, shape=(2,))

>>> b
tensor([2, 5], dtype=int32, loc=gpu:0, shape=(2,))

>>> output
tensor([True, False], dtype=bool, loc=gpu:0, shape=(2,))
__floordiv__(other: Tensor | Number) → Tensor

Performs an elementwise floor division.

Parameters:

  • self (Tensor) – Tensor to be floor-divided by other.

  • other (Tensor | Number) – The tensor by which to floor-divide this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([4.0, 6.0])
2b = tp.Tensor([3.0, 4.0])
3output = a // b
Local Variables
>>> a
tensor([4.0000, 6.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> b
tensor([3.0000, 4.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> output
tensor([1.0000, 1.0000], dtype=float32, loc=gpu:0, shape=(2,))
__ge__(other: Tensor | Number) → Tensor

Performs a ‘greater than or equal’ comparison.

Parameters:

  • self (Tensor) – Tensor to be compared with other.

  • other (Tensor | Number) – The tensor to be compared to this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([2, 3])
2b = tp.Tensor([2, 1])
3output = b >= a
Local Variables
>>> a
tensor([2, 3], dtype=int32, loc=gpu:0, shape=(2,))

>>> b
tensor([2, 1], dtype=int32, loc=gpu:0, shape=(2,))

>>> output
tensor([True, False], dtype=bool, loc=gpu:0, shape=(2,))
__getitem__(index: slice | int | Tensor | Sequence[slice | int | Tensor]) → Tensor

Returns a tensor containing a slice of this tensor.

Parameters:

  • self (Tensor) – Tensor that will be sliced.

  • index (slice | int | Tensor | Sequence[slice | int | Tensor]) – The index (as an int or Tripy tensor) or slice.

Returns:

A tensor containing the slice of this tensor.

Return type:

Tensor

Example
1input = tp.reshape(tp.arange(6, dtype=tp.float32), (1, 2, 3, 1))
2output = input[:, 1:2, :-1, 0]
Local Variables
>>> input
tensor(
    [[[[0.0000],
       [1.0000],
       [2.0000]],

      [[3.0000],
       [4.0000],
       [5.0000]]]], 
    dtype=float32, loc=gpu:0, shape=(1, 2, 3, 1))

>>> output
tensor(
    [[[3.0000, 4.0000]]], 
    dtype=float32, loc=gpu:0, shape=(1, 1, 2))
Example: Negative step size
1input = tp.arange(10)
2output = input[8:2:-1]
Local Variables
>>> input
tensor([0.0000, 1.0000, 2.0000, 3.0000, 4.0000, 5.0000, 6.0000, 7.0000, 8.0000, 9.0000], dtype=float32, loc=gpu:0, shape=(10,))

>>> output
tensor([8.0000, 7.0000, 6.0000, 5.0000, 4.0000, 3.0000], dtype=float32, loc=gpu:0, shape=(6,))
__gt__(other: Tensor | Number) → Tensor

Performs a ‘greater than’ comparison.

Parameters:

  • self (Tensor) – Tensor to be compared with other.

  • other (Tensor | Number) – The tensor to be compared to this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([2, 3])
2b = tp.Tensor([3, 1])
3output = b > a
Local Variables
>>> a
tensor([2, 3], dtype=int32, loc=gpu:0, shape=(2,))

>>> b
tensor([3, 1], dtype=int32, loc=gpu:0, shape=(2,))

>>> output
tensor([True, False], dtype=bool, loc=gpu:0, shape=(2,))
__le__(other: Tensor | Number) → Tensor

Performs a ‘less than or equal’ comparison.

Parameters:

  • self (Tensor) – Tensor to be compared with other.

  • other (Tensor | Number) – The tensor to be compared to this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([2, 3])
2b = tp.Tensor([2, 5])
3output = b <= a
Local Variables
>>> a
tensor([2, 3], dtype=int32, loc=gpu:0, shape=(2,))

>>> b
tensor([2, 5], dtype=int32, loc=gpu:0, shape=(2,))

>>> output
tensor([True, False], dtype=bool, loc=gpu:0, shape=(2,))
__lt__(other: Tensor | Number) → Tensor

Performs a ‘less than’ comparison.

Parameters:

  • self (Tensor) – Tensor to be compared with other.

  • other (Tensor | Number) – The tensor to be compared to this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([2, 3])
2b = tp.Tensor([1, 5])
3output = b < a
Local Variables
>>> a
tensor([2, 3], dtype=int32, loc=gpu:0, shape=(2,))

>>> b
tensor([1, 5], dtype=int32, loc=gpu:0, shape=(2,))

>>> output
tensor([True, False], dtype=bool, loc=gpu:0, shape=(2,))
__matmul__(other: Tensor) → Tensor

Performs matrix multiplication between two tensors.

  • If both tensors are 1D, a dot product is performed.

  • If both tensors are 2D, matrix multiplication is performed.

  • If either argument, but not both, is 1D, matrix-vector multiplication is performed.

  • If both tensors are 2D or higher dimensional and have differnt ranks, a dimension is inserted

    and batched matrix multiplication is performed with broadcast of relevant dimension.

Parameters:

  • self (Tensor) – Tensor to be multiplied with other.

  • other (Tensor) – The tensor by which to multiply.

Returns:

A new tensor.

Return type:

Tensor

Example
1a = tp.iota((2, 3), dtype=tp.float32)
2b = tp.iota((3, 2), dtype=tp.float32)
3
4output = a @ b
Local Variables
>>> a
tensor(
    [[0.0000, 0.0000, 0.0000],
     [1.0000, 1.0000, 1.0000]], 
    dtype=float32, loc=gpu:0, shape=(2, 3))

>>> b
tensor(
    [[0.0000, 0.0000],
     [1.0000, 1.0000],
     [2.0000, 2.0000]], 
    dtype=float32, loc=gpu:0, shape=(3, 2))

>>> output
tensor(
    [[0.0000, 0.0000],
     [3.0000, 3.0000]], 
    dtype=float32, loc=gpu:0, shape=(2, 2))
__mod__(other: Tensor | Number) → Tensor

Performs a modulo operation.

Parameters:

  • self (Tensor) – The tensor to be divided by other.

  • other (Tensor | Number) – The tensor by which to divide self. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape containing the result of the modulo operation.

Return type:

Tensor

Example
1a = tp.Tensor([4.0, 6.0])
2b = tp.Tensor([3.0, 4.0])
3output = a % b
Local Variables
>>> a
tensor([4.0000, 6.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> b
tensor([3.0000, 4.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> output
tensor([1.0000, 2.0000], dtype=float32, loc=gpu:0, shape=(2,))
__mul__(other: Tensor | Number) → Tensor

Performs an elementwise multiplication.

Parameters:

  • self (Tensor) – Tensor to be multiplied by other.

  • other (Tensor | Number) – The tensor by which to multiply this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([1.0, 2.0])
2b = tp.Tensor([2.0, 3.0])
3output = a * b
Local Variables
>>> a
tensor([1.0000, 2.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> b
tensor([2.0000, 3.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> output
tensor([2.0000, 6.0000], dtype=float32, loc=gpu:0, shape=(2,))
__ne__(other: Tensor | Number) → Tensor

Performs a ‘not equal’ comparison.

Parameters:

  • self (Tensor) – Tensor to be compared with other.

  • other (Tensor | Number) – The tensor to be compared to this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([2, 3])
2b = tp.Tensor([1, 3])
3output = b != a
Local Variables
>>> a
tensor([2, 3], dtype=int32, loc=gpu:0, shape=(2,))

>>> b
tensor([1, 3], dtype=int32, loc=gpu:0, shape=(2,))

>>> output
tensor([True, False], dtype=bool, loc=gpu:0, shape=(2,))
__pow__(other: Tensor | Number) → Tensor

Performs an elementwise exponentiation.

Parameters:

  • self (Tensor) – Tensor to be exponentiated by other.

  • other (Tensor | Number) – The tensor by which to exponentiate this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([1.0, 2.0])
2b = tp.Tensor([2.0, 3.0])
3output = a**b
Local Variables
>>> a
tensor([1.0000, 2.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> b
tensor([2.0000, 3.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> output
tensor([1.0000, 8.0000], dtype=float32, loc=gpu:0, shape=(2,))
__radd__(other: Tensor | Number) → Tensor

Performs an elementwise sum.

Parameters:

  • self (Tensor) – Tensor to be added to other.

  • other (Tensor | Number) – The tensor to add to this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([1, 2])
2b = tp.Tensor([2, 3])
3output = a + b
Local Variables
>>> a
tensor([1, 2], dtype=int32, loc=gpu:0, shape=(2,))

>>> b
tensor([2, 3], dtype=int32, loc=gpu:0, shape=(2,))

>>> output
tensor([3, 5], dtype=int32, loc=gpu:0, shape=(2,))
__rfloordiv__(other: Tensor | Number) → Tensor

Performs an elementwise floor division.

Parameters:

  • self (Tensor) – Tensor to be floor-divided by other.

  • other (Tensor | Number) – The tensor to be floor-divided by this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = 2
2b = tp.Tensor([2.0, 3.0])
3output = a // b
Local Variables
>>> b
tensor([2.0000, 3.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> output
tensor([1.0000, 0.0000], dtype=float32, loc=gpu:0, shape=(2,))
__rmod__(other: Tensor | Number) → Tensor

Performs a modulo operation.

Parameters:

  • self (Tensor) – The tensor to be divided by other.

  • other (Tensor | Number) – The tensor by which to divide self. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape containing the result of the modulo operation.

Return type:

Tensor

Example
1a = tp.Tensor([4.0, 6.0])
2output = 2 % a
Local Variables
>>> a
tensor([4.0000, 6.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> output
tensor([2.0000, 2.0000], dtype=float32, loc=gpu:0, shape=(2,))
__rmul__(other: Tensor | Number) → Tensor

Performs an elementwise multiplication.

Parameters:

  • self (Tensor) – Tensor to be multiplied by other.

  • other (Tensor | Number) – The tensor by which to multiply this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([1.0, 2.0])
2b = tp.Tensor([2.0, 3.0])
3output = a * b
Local Variables
>>> a
tensor([1.0000, 2.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> b
tensor([2.0000, 3.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> output
tensor([2.0000, 6.0000], dtype=float32, loc=gpu:0, shape=(2,))
__rpow__(other: Tensor | Number) → Tensor

Performs an elementwise exponentiation.

Parameters:

  • self (Tensor) – Tensor to be exponentiated by other.

  • other (Tensor | Number) – The tensor to be exponentiated by this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = 2.0
2b = tp.Tensor([2.0, 3.0])
3output = a**b
Local Variables
>>> b
tensor([2.0000, 3.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> output
tensor([4.0000, 8.0000], dtype=float32, loc=gpu:0, shape=(2,))
__rsub__(other: Tensor | Number) → Tensor

Performs an elementwise subtraction.

Parameters:

  • self (Tensor) – Tensor to be subtracted by other.

  • other (Tensor | Number) – The tensor to be subtracted from this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = 1
2b = tp.Tensor([1, 2])
3output = a - b
Local Variables
>>> b
tensor([1, 2], dtype=int32, loc=gpu:0, shape=(2,))

>>> output
tensor([0, -1], dtype=int32, loc=gpu:0, shape=(2,))
__rtruediv__(other: Tensor | Number) → Tensor

Performs an elementwise division.

Parameters:

  • self (Tensor) – Tensor to be divided by other.

  • other (Tensor | Number) – The tensor to be divided by this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = 6.0
2b = tp.Tensor([2.0, 3.0])
3output = a / b
Local Variables
>>> b
tensor([2.0000, 3.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> output
tensor([3.0000, 2.0000], dtype=float32, loc=gpu:0, shape=(2,))
__sub__(other: Tensor | Number) → Tensor

Performs an elementwise subtraction.

Parameters:

  • self (Tensor) – Tensor to be subtracted by other.

  • other (Tensor | Number) – The tensor to subtract from this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([2, 3])
2b = tp.Tensor([1, 2])
3output = a - b
Local Variables
>>> a
tensor([2, 3], dtype=int32, loc=gpu:0, shape=(2,))

>>> b
tensor([1, 2], dtype=int32, loc=gpu:0, shape=(2,))

>>> output
tensor([1, 1], dtype=int32, loc=gpu:0, shape=(2,))
__truediv__(other: Tensor | Number) → Tensor

Performs an elementwise division.

Parameters:

  • self (Tensor) – Tensor to be divided by other.

  • other (Tensor | Number) – The tensor by which to divide this one. It should be broadcast-compatible.

Returns:

A new tensor with the broadcasted shape.

Return type:

Tensor

Example
1a = tp.Tensor([4.0, 6.0])
2b = tp.Tensor([2.0, 3.0])
3output = a / b
Local Variables
>>> a
tensor([4.0000, 6.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> b
tensor([2.0000, 3.0000], dtype=float32, loc=gpu:0, shape=(2,))

>>> output
tensor([2.0000, 2.0000], dtype=float32, loc=gpu:0, shape=(2,))
property shape: Sequence[int | DimensionSize]

Represents the shape of the tensor.

Returns:

A sequence containing the shape of this tensor.

Example
1input = tp.ones((8, 2))
2shape = input.shape
Local Variables
>>> input
tensor(
    [[1.0000, 1.0000],
     [1.0000, 1.0000],
     [1.0000, 1.0000],
     [1.0000, 1.0000],
     [1.0000, 1.0000],
     [1.0000, 1.0000],
     [1.0000, 1.0000],
     [1.0000, 1.0000]], 
    dtype=float32, loc=gpu:0, shape=(8, 2))

>>> shape
[8, 2]

See also: