torch¶

The torch package contains data structures for multi-dimensional tensors and defines mathematical operations over these tensors. Additionally, it provides many utilities for efficient serialization of Tensors and arbitrary types, and other useful utilities.

It has a CUDA counterpart, that enables you to run your tensor computations on an NVIDIA GPU with compute capability >= 3.0.

Tensors¶

`is_tensor`	Returns True if obj is a PyTorch tensor.
`is_storage`	Returns True if obj is a PyTorch storage object.
`is_complex`	Returns True if the data type of `input` is a complex data type i.e., one of `torch.complex64`, and `torch.complex128`.
`is_conj`	Returns True if the `input` is a conjugated tensor, i.e. its conjugate bit is set to True.
`is_floating_point`	Returns True if the data type of `input` is a floating point data type i.e., one of `torch.float64`, `torch.float32`, `torch.float16`, and `torch.bfloat16`.
`is_nonzero`	Returns True if the `input` is a single element tensor which is not equal to zero after type conversions.
`set_default_dtype`	Sets the default floating point dtype to `d`.
`get_default_dtype`	Get the current default floating point `torch.dtype`.
`set_default_device`	Sets the default `torch.Tensor` to be allocated on `device`.
`get_default_device`	Gets the default `torch.Tensor` to be allocated on `device`
`set_default_tensor_type`
`numel`	Returns the total number of elements in the `input` tensor.
`set_printoptions`	Set options for printing.
`set_flush_denormal`	Disables denormal floating numbers on CPU.

Creation Ops¶

Note

Random sampling creation ops are listed under Random sampling and include: torch.rand() torch.rand_like() torch.randn() torch.randn_like() torch.randint() torch.randint_like() torch.randperm() You may also use torch.empty() with the In-place random sampling methods to create torch.Tensor s with values sampled from a broader range of distributions.

`tensor`	Constructs a tensor with no autograd history (also known as a "leaf tensor", see Autograd mechanics) by copying `data`.
`sparse_coo_tensor`	Constructs a sparse tensor in COO(rdinate) format with specified values at the given `indices`.
`sparse_csr_tensor`	Constructs a sparse tensor in CSR (Compressed Sparse Row) with specified values at the given `crow_indices` and `col_indices`.
`sparse_csc_tensor`	Constructs a sparse tensor in CSC (Compressed Sparse Column) with specified values at the given `ccol_indices` and `row_indices`.
`sparse_bsr_tensor`	Constructs a sparse tensor in BSR (Block Compressed Sparse Row)) with specified 2-dimensional blocks at the given `crow_indices` and `col_indices`.
`sparse_bsc_tensor`	Constructs a sparse tensor in BSC (Block Compressed Sparse Column)) with specified 2-dimensional blocks at the given `ccol_indices` and `row_indices`.
`asarray`	Converts `obj` to a tensor.
`as_tensor`	Converts `data` into a tensor, sharing data and preserving autograd history if possible.
`as_strided`	Create a view of an existing torch.Tensor `input` with specified `size`, `stride` and `storage_offset`.
`from_file`	Creates a CPU tensor with a storage backed by a memory-mapped file.
`from_numpy`	Creates a `Tensor` from a `numpy.ndarray`.
`from_dlpack`	Converts a tensor from an external library into a `torch.Tensor`.
`frombuffer`	Creates a 1-dimensional `Tensor` from an object that implements the Python buffer protocol.
`zeros`	Returns a tensor filled with the scalar value 0, with the shape defined by the variable argument `size`.
`zeros_like`	Returns a tensor filled with the scalar value 0, with the same size as `input`.
`ones`	Returns a tensor filled with the scalar value 1, with the shape defined by the variable argument `size`.
`ones_like`	Returns a tensor filled with the scalar value 1, with the same size as `input`.
`arange`	Returns a 1-D tensor of size $\left\lceil \frac{\text{end} - \text{start}}{\text{step}} \right\rceil$ with values from the interval `[start, end)` taken with common difference `step` beginning from start.
`range`	Returns a 1-D tensor of size $\left\lfloor \frac{\text{end} - \text{start}}{\text{step}} \right\rfloor + 1$ with values from `start` to `end` with step `step`.
`linspace`	Creates a one-dimensional tensor of size `steps` whose values are evenly spaced from `start` to `end`, inclusive.
`logspace`	Creates a one-dimensional tensor of size `steps` whose values are evenly spaced from ${{\text{{base}}}}^{{\text{{start}}}}$ to ${{\text{{base}}}}^{{\text{{end}}}}$ , inclusive, on a logarithmic scale with base `base`.
`eye`	Returns a 2-D tensor with ones on the diagonal and zeros elsewhere.
`empty`	Returns a tensor filled with uninitialized data.
`empty_like`	Returns an uninitialized tensor with the same size as `input`.
`empty_strided`	Creates a tensor with the specified `size` and `stride` and filled with undefined data.
`full`	Creates a tensor of size `size` filled with `fill_value`.
`full_like`	Returns a tensor with the same size as `input` filled with `fill_value`.
`quantize_per_tensor`	Converts a float tensor to a quantized tensor with given scale and zero point.
`quantize_per_channel`	Converts a float tensor to a per-channel quantized tensor with given scales and zero points.
`dequantize`	Returns an fp32 Tensor by dequantizing a quantized Tensor
`complex`	Constructs a complex tensor with its real part equal to `real` and its imaginary part equal to `imag`.
`polar`	Constructs a complex tensor whose elements are Cartesian coordinates corresponding to the polar coordinates with absolute value `abs` and angle `angle`.
`heaviside`	Computes the Heaviside step function for each element in `input`.

Indexing, Slicing, Joining, Mutating Ops¶

`adjoint`	Returns a view of the tensor conjugated and with the last two dimensions transposed.
`argwhere`	Returns a tensor containing the indices of all non-zero elements of `input`.
`cat`	Concatenates the given sequence of `seq` tensors in the given dimension.
`concat`	Alias of `torch.cat()`.
`concatenate`	Alias of `torch.cat()`.
`conj`	Returns a view of `input` with a flipped conjugate bit.
`chunk`	Attempts to split a tensor into the specified number of chunks.
`dsplit`	Splits `input`, a tensor with three or more dimensions, into multiple tensors depthwise according to `indices_or_sections`.
`column_stack`	Creates a new tensor by horizontally stacking the tensors in `tensors`.
`dstack`	Stack tensors in sequence depthwise (along third axis).
`gather`	Gathers values along an axis specified by dim.
`hsplit`	Splits `input`, a tensor with one or more dimensions, into multiple tensors horizontally according to `indices_or_sections`.
`hstack`	Stack tensors in sequence horizontally (column wise).
`index_add`	See `index_add_()` for function description.
`index_copy`	See `index_add_()` for function description.
`index_reduce`	See `index_reduce_()` for function description.
`index_select`	Returns a new tensor which indexes the `input` tensor along dimension `dim` using the entries in `index` which is a LongTensor.
`masked_select`	Returns a new 1-D tensor which indexes the `input` tensor according to the boolean mask `mask` which is a BoolTensor.
`movedim`	Moves the dimension(s) of `input` at the position(s) in `source` to the position(s) in `destination`.
`moveaxis`	Alias for `torch.movedim()`.
`narrow`	Returns a new tensor that is a narrowed version of `input` tensor.
`narrow_copy`	Same as `Tensor.narrow()` except this returns a copy rather than shared storage.
`nonzero`
`permute`	Returns a view of the original tensor `input` with its dimensions permuted.
`reshape`	Returns a tensor with the same data and number of elements as `input`, but with the specified shape.
`row_stack`	Alias of `torch.vstack()`.
`select`	Slices the `input` tensor along the selected dimension at the given index.
`scatter`	Out-of-place version of `torch.Tensor.scatter_()`
`diagonal_scatter`	Embeds the values of the `src` tensor into `input` along the diagonal elements of `input`, with respect to `dim1` and `dim2`.
`select_scatter`	Embeds the values of the `src` tensor into `input` at the given index.
`slice_scatter`	Embeds the values of the `src` tensor into `input` at the given dimension.
`scatter_add`	Out-of-place version of `torch.Tensor.scatter_add_()`
`scatter_reduce`	Out-of-place version of `torch.Tensor.scatter_reduce_()`
`split`	Splits the tensor into chunks.
`squeeze`	Returns a tensor with all specified dimensions of `input` of size 1 removed.
`stack`	Concatenates a sequence of tensors along a new dimension.
`swapaxes`	Alias for `torch.transpose()`.
`swapdims`	Alias for `torch.transpose()`.
`t`	Expects `input` to be <= 2-D tensor and transposes dimensions 0 and 1.
`take`	Returns a new tensor with the elements of `input` at the given indices.
`take_along_dim`	Selects values from `input` at the 1-dimensional indices from `indices` along the given `dim`.
`tensor_split`	Splits a tensor into multiple sub-tensors, all of which are views of `input`, along dimension `dim` according to the indices or number of sections specified by `indices_or_sections`.
`tile`	Constructs a tensor by repeating the elements of `input`.
`transpose`	Returns a tensor that is a transposed version of `input`.
`unbind`	Removes a tensor dimension.
`unravel_index`	Converts a tensor of flat indices into a tuple of coordinate tensors that index into an arbitrary tensor of the specified shape.
`unsqueeze`	Returns a new tensor with a dimension of size one inserted at the specified position.
`vsplit`	Splits `input`, a tensor with two or more dimensions, into multiple tensors vertically according to `indices_or_sections`.
`vstack`	Stack tensors in sequence vertically (row wise).
`where`	Return a tensor of elements selected from either `input` or `other`, depending on `condition`.

Accelerators¶

Within the PyTorch repo, we define an “Accelerator” as a torch.device that is being used alongside a CPU to speed up computation. These device use an asynchronous execution scheme, using torch.Stream and torch.Event as their main way to perform synchronization. We also assume that only one such accelerator can be available at once on a given host. This allows us to use the current accelerator as the default device for relevant concepts such as pinned memory, Stream device_type, FSDP, etc.

As of today, accelerator devices are (in no particular order) “CUDA”, “MTIA”, “XPU”, and PrivateUse1 (many device not in the PyTorch repo itself).

Stream

An in-order queue of executing the respective tasks asynchronously in first in first out (FIFO) order.

Event

Query and record Stream status to identify or control dependencies across Stream and measure timing.

Generators¶

Generator

Creates and returns a generator object that manages the state of the algorithm which produces pseudo random numbers.

Random sampling¶

`seed`	Sets the seed for generating random numbers to a non-deterministic random number on all devices.
`manual_seed`	Sets the seed for generating random numbers on all devices.
`initial_seed`	Returns the initial seed for generating random numbers as a Python long.
`get_rng_state`	Returns the random number generator state as a torch.ByteTensor.
`set_rng_state`	Sets the random number generator state.

torch.default_generator Returns the default CPU torch.Generator¶

`bernoulli`	Draws binary random numbers (0 or 1) from a Bernoulli distribution.
`multinomial`	Returns a tensor where each row contains `num_samples` indices sampled from the multinomial (a stricter definition would be multivariate, refer to `torch.distributions.multinomial.Multinomial` for more details) probability distribution located in the corresponding row of tensor `input`.
`normal`	Returns a tensor of random numbers drawn from separate normal distributions whose mean and standard deviation are given.
`poisson`	Returns a tensor of the same size as `input` with each element sampled from a Poisson distribution with rate parameter given by the corresponding element in `input` i.e.,
`rand`	Returns a tensor filled with random numbers from a uniform distribution on the interval $[0, 1)$
`rand_like`	Returns a tensor with the same size as `input` that is filled with random numbers from a uniform distribution on the interval $[0, 1)$ .
`randint`	Returns a tensor filled with random integers generated uniformly between `low` (inclusive) and `high` (exclusive).
`randint_like`	Returns a tensor with the same shape as Tensor `input` filled with random integers generated uniformly between `low` (inclusive) and `high` (exclusive).
`randn`	Returns a tensor filled with random numbers from a normal distribution with mean 0 and variance 1 (also called the standard normal distribution).
`randn_like`	Returns a tensor with the same size as `input` that is filled with random numbers from a normal distribution with mean 0 and variance 1.
`randperm`	Returns a random permutation of integers from `0` to `n - 1`.

In-place random sampling¶

There are a few more in-place random sampling functions defined on Tensors as well. Click through to refer to their documentation:

torch.Tensor.bernoulli_() - in-place version of torch.bernoulli()
torch.Tensor.cauchy_() - numbers drawn from the Cauchy distribution
torch.Tensor.exponential_() - numbers drawn from the exponential distribution
torch.Tensor.geometric_() - elements drawn from the geometric distribution
torch.Tensor.log_normal_() - samples from the log-normal distribution
torch.Tensor.normal_() - in-place version of torch.normal()
torch.Tensor.random_() - numbers sampled from the discrete uniform distribution
torch.Tensor.uniform_() - numbers sampled from the continuous uniform distribution

Quasi-random sampling¶

quasirandom.SobolEngine

The torch.quasirandom.SobolEngine is an engine for generating (scrambled) Sobol sequences.

Serialization¶

save

Saves an object to a disk file.

load

Loads an object saved with torch.save() from a file.

Parallelism¶

`get_num_threads`	Returns the number of threads used for parallelizing CPU operations
`set_num_threads`	Sets the number of threads used for intraop parallelism on CPU.
`get_num_interop_threads`	Returns the number of threads used for inter-op parallelism on CPU (e.g.
`set_num_interop_threads`	Sets the number of threads used for interop parallelism (e.g.

Locally disabling gradient computation¶

The context managers torch.no_grad(), torch.enable_grad(), and torch.set_grad_enabled() are helpful for locally disabling and enabling gradient computation. See Locally disabling gradient computation for more details on their usage. These context managers are thread local, so they won’t work if you send work to another thread using the threading module, etc.

Examples:

>>> x = torch.zeros(1, requires_grad=True)
>>> with torch.no_grad():
...     y = x * 2
>>> y.requires_grad
False

>>> is_train = False
>>> with torch.set_grad_enabled(is_train):
...     y = x * 2
>>> y.requires_grad
False

>>> torch.set_grad_enabled(True)  # this can also be used as a function
>>> y = x * 2
>>> y.requires_grad
True

>>> torch.set_grad_enabled(False)
>>> y = x * 2
>>> y.requires_grad
False

`no_grad`	Context-manager that disables gradient calculation.
`enable_grad`	Context-manager that enables gradient calculation.
`autograd.grad_mode.set_grad_enabled`	Context-manager that sets gradient calculation on or off.
`is_grad_enabled`	Returns True if grad mode is currently enabled.
`autograd.grad_mode.inference_mode`	Context-manager that enables or disables inference mode.
`is_inference_mode_enabled`	Returns True if inference mode is currently enabled.

Math operations¶

Pointwise Ops¶

`abs`	Computes the absolute value of each element in `input`.
`absolute`	Alias for `torch.abs()`
`acos`	Computes the inverse cosine of each element in `input`.
`arccos`	Alias for `torch.acos()`.
`acosh`	Returns a new tensor with the inverse hyperbolic cosine of the elements of `input`.
`arccosh`	Alias for `torch.acosh()`.
`add`	Adds `other`, scaled by `alpha`, to `input`.
`addcdiv`	Performs the element-wise division of `tensor1` by `tensor2`, multiplies the result by the scalar `value` and adds it to `input`.
`addcmul`	Performs the element-wise multiplication of `tensor1` by `tensor2`, multiplies the result by the scalar `value` and adds it to `input`.
`angle`	Computes the element-wise angle (in radians) of the given `input` tensor.
`asin`	Returns a new tensor with the arcsine of the elements of `input`.
`arcsin`	Alias for `torch.asin()`.
`asinh`	Returns a new tensor with the inverse hyperbolic sine of the elements of `input`.
`arcsinh`	Alias for `torch.asinh()`.
`atan`	Returns a new tensor with the arctangent of the elements of `input`.
`arctan`	Alias for `torch.atan()`.
`atanh`	Returns a new tensor with the inverse hyperbolic tangent of the elements of `input`.
`arctanh`	Alias for `torch.atanh()`.
`atan2`	Element-wise arctangent of $\text{input}_{i} / \text{other}_{i}$ with consideration of the quadrant.
`arctan2`	Alias for `torch.atan2()`.
`bitwise_not`	Computes the bitwise NOT of the given input tensor.
`bitwise_and`	Computes the bitwise AND of `input` and `other`.
`bitwise_or`	Computes the bitwise OR of `input` and `other`.
`bitwise_xor`	Computes the bitwise XOR of `input` and `other`.
`bitwise_left_shift`	Computes the left arithmetic shift of `input` by `other` bits.
`bitwise_right_shift`	Computes the right arithmetic shift of `input` by `other` bits.
`ceil`	Returns a new tensor with the ceil of the elements of `input`, the smallest integer greater than or equal to each element.
`clamp`	Clamps all elements in `input` into the range [ `min`, `max` ].
`clip`	Alias for `torch.clamp()`.
`conj_physical`	Computes the element-wise conjugate of the given `input` tensor.
`copysign`	Create a new floating-point tensor with the magnitude of `input` and the sign of `other`, elementwise.
`cos`	Returns a new tensor with the cosine of the elements of `input`.
`cosh`	Returns a new tensor with the hyperbolic cosine of the elements of `input`.
`deg2rad`	Returns a new tensor with each of the elements of `input` converted from angles in degrees to radians.
`div`	Divides each element of the input `input` by the corresponding element of `other`.
`divide`	Alias for `torch.div()`.
`digamma`	Alias for `torch.special.digamma()`.
`erf`	Alias for `torch.special.erf()`.
`erfc`	Alias for `torch.special.erfc()`.
`erfinv`	Alias for `torch.special.erfinv()`.
`exp`	Returns a new tensor with the exponential of the elements of the input tensor `input`.
`exp2`	Alias for `torch.special.exp2()`.
`expm1`	Alias for `torch.special.expm1()`.
`fake_quantize_per_channel_affine`	Returns a new tensor with the data in `input` fake quantized per channel using `scale`, `zero_point`, `quant_min` and `quant_max`, across the channel specified by `axis`.
`fake_quantize_per_tensor_affine`	Returns a new tensor with the data in `input` fake quantized using `scale`, `zero_point`, `quant_min` and `quant_max`.
`fix`	Alias for `torch.trunc()`
`float_power`	Raises `input` to the power of `exponent`, elementwise, in double precision.
`floor`	Returns a new tensor with the floor of the elements of `input`, the largest integer less than or equal to each element.
`floor_divide`
`fmod`	Applies C++'s std::fmod entrywise.
`frac`	Computes the fractional portion of each element in `input`.
`frexp`	Decomposes `input` into mantissa and exponent tensors such that $\text{input} = \text{mantissa} \times 2^{\text{exponent}}$ .
`gradient`	Estimates the gradient of a function $g : \mathbb{R}^n \rightarrow \mathbb{R}$ in one or more dimensions using the second-order accurate central differences method and either first or second order estimates at the boundaries.
`imag`	Returns a new tensor containing imaginary values of the `self` tensor.
`ldexp`	Multiplies `input` by 2 ** `other`.
`lerp`	Does a linear interpolation of two tensors `start` (given by `input`) and `end` based on a scalar or tensor `weight` and returns the resulting `out` tensor.
`lgamma`	Computes the natural logarithm of the absolute value of the gamma function on `input`.
`log`	Returns a new tensor with the natural logarithm of the elements of `input`.
`log10`	Returns a new tensor with the logarithm to the base 10 of the elements of `input`.
`log1p`	Returns a new tensor with the natural logarithm of (1 + `input`).
`log2`	Returns a new tensor with the logarithm to the base 2 of the elements of `input`.
`logaddexp`	Logarithm of the sum of exponentiations of the inputs.
`logaddexp2`	Logarithm of the sum of exponentiations of the inputs in base-2.
`logical_and`	Computes the element-wise logical AND of the given input tensors.
`logical_not`	Computes the element-wise logical NOT of the given input tensor.
`logical_or`	Computes the element-wise logical OR of the given input tensors.
`logical_xor`	Computes the element-wise logical XOR of the given input tensors.
`logit`	Alias for `torch.special.logit()`.
`hypot`	Given the legs of a right triangle, return its hypotenuse.
`i0`	Alias for `torch.special.i0()`.
`igamma`	Alias for `torch.special.gammainc()`.
`igammac`	Alias for `torch.special.gammaincc()`.
`mul`	Multiplies `input` by `other`.
`multiply`	Alias for `torch.mul()`.
`mvlgamma`	Alias for `torch.special.multigammaln()`.
`nan_to_num`	Replaces `NaN`, positive infinity, and negative infinity values in `input` with the values specified by `nan`, `posinf`, and `neginf`, respectively.
`neg`	Returns a new tensor with the negative of the elements of `input`.
`negative`	Alias for `torch.neg()`
`nextafter`	Return the next floating-point value after `input` towards `other`, elementwise.
`polygamma`	Alias for `torch.special.polygamma()`.
`positive`	Returns `input`.
`pow`	Takes the power of each element in `input` with `exponent` and returns a tensor with the result.
`quantized_batch_norm`	Applies batch normalization on a 4D (NCHW) quantized tensor.
`quantized_max_pool1d`	Applies a 1D max pooling over an input quantized tensor composed of several input planes.
`quantized_max_pool2d`	Applies a 2D max pooling over an input quantized tensor composed of several input planes.
`rad2deg`	Returns a new tensor with each of the elements of `input` converted from angles in radians to degrees.
`real`	Returns a new tensor containing real values of the `self` tensor.
`reciprocal`	Returns a new tensor with the reciprocal of the elements of `input`
`remainder`	Computes Python's modulus operation entrywise.
`round`	Rounds elements of `input` to the nearest integer.
`rsqrt`	Returns a new tensor with the reciprocal of the square-root of each of the elements of `input`.
`sigmoid`	Alias for `torch.special.expit()`.
`sign`	Returns a new tensor with the signs of the elements of `input`.
`sgn`	This function is an extension of torch.sign() to complex tensors.
`signbit`	Tests if each element of `input` has its sign bit set or not.
`sin`	Returns a new tensor with the sine of the elements of `input`.
`sinc`	Alias for `torch.special.sinc()`.
`sinh`	Returns a new tensor with the hyperbolic sine of the elements of `input`.
`softmax`	Alias for `torch.nn.functional.softmax()`.
`sqrt`	Returns a new tensor with the square-root of the elements of `input`.
`square`	Returns a new tensor with the square of the elements of `input`.
`sub`	Subtracts `other`, scaled by `alpha`, from `input`.
`subtract`	Alias for `torch.sub()`.
`tan`	Returns a new tensor with the tangent of the elements of `input`.
`tanh`	Returns a new tensor with the hyperbolic tangent of the elements of `input`.
`true_divide`	Alias for `torch.div()` with `rounding_mode=None`.
`trunc`	Returns a new tensor with the truncated integer values of the elements of `input`.
`xlogy`	Alias for `torch.special.xlogy()`.

Reduction Ops¶

`argmax`	Returns the indices of the maximum value of all elements in the `input` tensor.
`argmin`	Returns the indices of the minimum value(s) of the flattened tensor or along a dimension
`amax`	Returns the maximum value of each slice of the `input` tensor in the given dimension(s) `dim`.
`amin`	Returns the minimum value of each slice of the `input` tensor in the given dimension(s) `dim`.
`aminmax`	Computes the minimum and maximum values of the `input` tensor.
`all`	Tests if all elements in `input` evaluate to True.
`any`	Tests if any element in `input` evaluates to True.
`max`	Returns the maximum value of all elements in the `input` tensor.
`min`	Returns the minimum value of all elements in the `input` tensor.
`dist`	Returns the p-norm of (`input` - `other`)
`logsumexp`	Returns the log of summed exponentials of each row of the `input` tensor in the given dimension `dim`.
`mean`	Returns the mean value of all elements in the `input` tensor.
`nanmean`	Computes the mean of all non-NaN elements along the specified dimensions.
`median`	Returns the median of the values in `input`.
`nanmedian`	Returns the median of the values in `input`, ignoring `NaN` values.
`mode`	Returns a namedtuple `(values, indices)` where `values` is the mode value of each row of the `input` tensor in the given dimension `dim`, i.e. a value which appears most often in that row, and `indices` is the index location of each mode value found.
`norm`	Returns the matrix norm or vector norm of a given tensor.
`nansum`	Returns the sum of all elements, treating Not a Numbers (NaNs) as zero.
`prod`	Returns the product of all elements in the `input` tensor.
`quantile`	Computes the q-th quantiles of each row of the `input` tensor along the dimension `dim`.
`nanquantile`	This is a variant of `torch.quantile()` that "ignores" `NaN` values, computing the quantiles `q` as if `NaN` values in `input` did not exist.
`std`	Calculates the standard deviation over the dimensions specified by `dim`.
`std_mean`	Calculates the standard deviation and mean over the dimensions specified by `dim`.
`sum`	Returns the sum of all elements in the `input` tensor.
`unique`	Returns the unique elements of the input tensor.
`unique_consecutive`	Eliminates all but the first element from every consecutive group of equivalent elements.
`var`	Calculates the variance over the dimensions specified by `dim`.
`var_mean`	Calculates the variance and mean over the dimensions specified by `dim`.
`count_nonzero`	Counts the number of non-zero values in the tensor `input` along the given `dim`.

Comparison Ops¶

`allclose`	This function checks if `input` and `other` satisfy the condition:
`argsort`	Returns the indices that sort a tensor along a given dimension in ascending order by value.
`eq`	Computes element-wise equality
`equal`	`True` if two tensors have the same size and elements, `False` otherwise.
`ge`	Computes $\text{input} \geq \text{other}$ element-wise.
`greater_equal`	Alias for `torch.ge()`.
`gt`	Computes $\text{input} > \text{other}$ element-wise.
`greater`	Alias for `torch.gt()`.
`isclose`	Returns a new tensor with boolean elements representing if each element of `input` is "close" to the corresponding element of `other`.
`isfinite`	Returns a new tensor with boolean elements representing if each element is finite or not.
`isin`	Tests if each element of `elements` is in `test_elements`.
`isinf`	Tests if each element of `input` is infinite (positive or negative infinity) or not.
`isposinf`	Tests if each element of `input` is positive infinity or not.
`isneginf`	Tests if each element of `input` is negative infinity or not.
`isnan`	Returns a new tensor with boolean elements representing if each element of `input` is NaN or not.
`isreal`	Returns a new tensor with boolean elements representing if each element of `input` is real-valued or not.
`kthvalue`	Returns a namedtuple `(values, indices)` where `values` is the `k` th smallest element of each row of the `input` tensor in the given dimension `dim`.
`le`	Computes $\text{input} \leq \text{other}$ element-wise.
`less_equal`	Alias for `torch.le()`.
`lt`	Computes $\text{input} < \text{other}$ element-wise.
`less`	Alias for `torch.lt()`.
`maximum`	Computes the element-wise maximum of `input` and `other`.
`minimum`	Computes the element-wise minimum of `input` and `other`.
`fmax`	Computes the element-wise maximum of `input` and `other`.
`fmin`	Computes the element-wise minimum of `input` and `other`.
`ne`	Computes $\text{input} \neq \text{other}$ element-wise.
`not_equal`	Alias for `torch.ne()`.
`sort`	Sorts the elements of the `input` tensor along a given dimension in ascending order by value.
`topk`	Returns the `k` largest elements of the given `input` tensor along a given dimension.
`msort`	Sorts the elements of the `input` tensor along its first dimension in ascending order by value.

Spectral Ops¶

`stft`	Short-time Fourier transform (STFT).
`istft`	Inverse short time Fourier Transform.
`bartlett_window`	Bartlett window function.
`blackman_window`	Blackman window function.
`hamming_window`	Hamming window function.
`hann_window`	Hann window function.
`kaiser_window`	Computes the Kaiser window with window length `window_length` and shape parameter `beta`.

Other Operations¶

atleast_1d

Returns a 1-dimensional view of each input tensor with zero dimensions.

atleast_2d

Returns a 2-dimensional view of each input tensor with zero dimensions.

atleast_3d

Returns a 3-dimensional view of each input tensor with zero dimensions.

bincount

Count the frequency of each value in an array of non-negative ints.

block_diag

Create a block diagonal matrix from provided tensors.

broadcast_tensors

Broadcasts the given tensors according to Broadcasting semantics.

broadcast_to

Broadcasts input to the shape shape.

broadcast_shapes

Similar to broadcast_tensors() but for shapes.

bucketize

Returns the indices of the buckets to which each value in the input belongs, where the boundaries of the buckets are set by boundaries.

cartesian_prod

Do cartesian product of the given sequence of tensors.

cdist

Computes batched the p-norm distance between each pair of the two collections of row vectors.

clone

Returns a copy of input.

combinations

Compute combinations of length $r$ of the given tensor.

corrcoef

Estimates the Pearson product-moment correlation coefficient matrix of the variables given by the input matrix, where rows are the variables and columns are the observations.

cov

Estimates the covariance matrix of the variables given by the input matrix, where rows are the variables and columns are the observations.

cross

Returns the cross product of vectors in dimension dim of input and other.

cummax

Returns a namedtuple (values, indices) where values is the cumulative maximum of elements of input in the dimension dim.

cummin

Returns a namedtuple (values, indices) where values is the cumulative minimum of elements of input in the dimension dim.

cumprod

Returns the cumulative product of elements of input in the dimension dim.

cumsum

Returns the cumulative sum of elements of input in the dimension dim.

diag

If input is a vector (1-D tensor), then returns a 2-D square tensor

diag_embed

Creates a tensor whose diagonals of certain 2D planes (specified by dim1 and dim2) are filled by input.

diagflat

If input is a vector (1-D tensor), then returns a 2-D square tensor

diagonal

Returns a partial view of input with the its diagonal elements with respect to dim1 and dim2 appended as a dimension at the end of the shape.

diff

Computes the n-th forward difference along the given dimension.

einsum

Sums the product of the elements of the input operands along dimensions specified using a notation based on the Einstein summation convention.

flatten

Flattens input by reshaping it into a one-dimensional tensor.

flip

Reverse the order of an n-D tensor along given axis in dims.

fliplr

Flip tensor in the left/right direction, returning a new tensor.

flipud

Flip tensor in the up/down direction, returning a new tensor.

kron

Computes the Kronecker product, denoted by $\otimes$ , of input and other.

rot90

Rotate an n-D tensor by 90 degrees in the plane specified by dims axis.

gcd

Computes the element-wise greatest common divisor (GCD) of input and other.

histc

Computes the histogram of a tensor.

histogram

Computes a histogram of the values in a tensor.

histogramdd

Computes a multi-dimensional histogram of the values in a tensor.

meshgrid

Creates grids of coordinates specified by the 1D inputs in attr:tensors.

lcm

Computes the element-wise least common multiple (LCM) of input and other.

logcumsumexp

Returns the logarithm of the cumulative summation of the exponentiation of elements of input in the dimension dim.

ravel

Return a contiguous flattened tensor.

renorm

Returns a tensor where each sub-tensor of input along dimension dim is normalized such that the p-norm of the sub-tensor is lower than the value maxnorm

repeat_interleave

Repeat elements of a tensor.

roll

Roll the tensor input along the given dimension(s).

searchsorted

Find the indices from the innermost dimension of sorted_sequence such that, if the corresponding values in values were inserted before the indices, when sorted, the order of the corresponding innermost dimension within sorted_sequence would be preserved.

tensordot

Returns a contraction of a and b over multiple dimensions.

trace

Returns the sum of the elements of the diagonal of the input 2-D matrix.

tril

Returns the lower triangular part of the matrix (2-D tensor) or batch of matrices input, the other elements of the result tensor out are set to 0.

tril_indices

Returns the indices of the lower triangular part of a row-by- col matrix in a 2-by-N Tensor, where the first row contains row coordinates of all indices and the second row contains column coordinates.

triu

Returns the upper triangular part of a matrix (2-D tensor) or batch of matrices input, the other elements of the result tensor out are set to 0.

triu_indices

Returns the indices of the upper triangular part of a row by col matrix in a 2-by-N Tensor, where the first row contains row coordinates of all indices and the second row contains column coordinates.

unflatten

Expands a dimension of the input tensor over multiple dimensions.

vander

Generates a Vandermonde matrix.

view_as_real

Returns a view of input as a real tensor.

view_as_complex

Returns a view of input as a complex tensor.

resolve_conj

Returns a new tensor with materialized conjugation if input's conjugate bit is set to True, else returns input.

resolve_neg

Returns a new tensor with materialized negation if input's negative bit is set to True, else returns input.

BLAS and LAPACK Operations¶

addbmm

Performs a batch matrix-matrix product of matrices stored in batch1 and batch2, with a reduced add step (all matrix multiplications get accumulated along the first dimension).

addmm

Performs a matrix multiplication of the matrices mat1 and mat2.

addmv

Performs a matrix-vector product of the matrix mat and the vector vec.

addr

Performs the outer-product of vectors vec1 and vec2 and adds it to the matrix input.

baddbmm

Performs a batch matrix-matrix product of matrices in batch1 and batch2.

bmm

Performs a batch matrix-matrix product of matrices stored in input and mat2.

chain_matmul

Returns the matrix product of the $N$ 2-D tensors.

cholesky

Computes the Cholesky decomposition of a symmetric positive-definite matrix $A$ or for batches of symmetric positive-definite matrices.

cholesky_inverse

Computes the inverse of a complex Hermitian or real symmetric positive-definite matrix given its Cholesky decomposition.

cholesky_solve

Computes the solution of a system of linear equations with complex Hermitian or real symmetric positive-definite lhs given its Cholesky decomposition.

dot

Computes the dot product of two 1D tensors.

geqrf

This is a low-level function for calling LAPACK's geqrf directly.

ger

Alias of torch.outer().

inner

Computes the dot product for 1D tensors.

inverse

Alias for torch.linalg.inv()

det

Alias for torch.linalg.det()

logdet

Calculates log determinant of a square matrix or batches of square matrices.

slogdet

Alias for torch.linalg.slogdet()

lu

Computes the LU factorization of a matrix or batches of matrices A.

lu_solve

Returns the LU solve of the linear system $Ax = b$ using the partially pivoted LU factorization of A from lu_factor().

lu_unpack

Unpacks the LU decomposition returned by lu_factor() into the P, L, U matrices.

matmul

Matrix product of two tensors.

matrix_power

Alias for torch.linalg.matrix_power()

matrix_exp

Alias for torch.linalg.matrix_exp().

mm

Performs a matrix multiplication of the matrices input and mat2.

mv

Performs a matrix-vector product of the matrix input and the vector vec.

orgqr

Alias for torch.linalg.householder_product().

ormqr

Computes the matrix-matrix multiplication of a product of Householder matrices with a general matrix.

outer

Outer product of input and vec2.

pinverse

Alias for torch.linalg.pinv()

qr

Computes the QR decomposition of a matrix or a batch of matrices input, and returns a namedtuple (Q, R) of tensors such that $\text{input} = Q R$ with $Q$ being an orthogonal matrix or batch of orthogonal matrices and $R$ being an upper triangular matrix or batch of upper triangular matrices.

svd

Computes the singular value decomposition of either a matrix or batch of matrices input.

svd_lowrank

Return the singular value decomposition (U, S, V) of a matrix, batches of matrices, or a sparse matrix $A$ such that $A \approx U \operatorname{diag}(S) V^{\text{H}}$ .

pca_lowrank

Performs linear Principal Component Analysis (PCA) on a low-rank matrix, batches of such matrices, or sparse matrix.

lobpcg

Find the k largest (or smallest) eigenvalues and the corresponding eigenvectors of a symmetric positive definite generalized eigenvalue problem using matrix-free LOBPCG methods.

trapz

Alias for torch.trapezoid().

trapezoid

Computes the trapezoidal rule along dim.

cumulative_trapezoid

Cumulatively computes the trapezoidal rule along dim.

triangular_solve

Solves a system of equations with a square upper or lower triangular invertible matrix $A$ and multiple right-hand sides $b$ .

vdot

Computes the dot product of two 1D vectors along a dimension.

Foreach Operations¶

Warning

This API is in beta and subject to future changes. Forward-mode AD is not supported.

`_foreach_abs`	Apply `torch.abs()` to each Tensor of the input list.
`_foreach_abs_`	Apply `torch.abs()` to each Tensor of the input list.
`_foreach_acos`	Apply `torch.acos()` to each Tensor of the input list.
`_foreach_acos_`	Apply `torch.acos()` to each Tensor of the input list.
`_foreach_asin`	Apply `torch.asin()` to each Tensor of the input list.
`_foreach_asin_`	Apply `torch.asin()` to each Tensor of the input list.
`_foreach_atan`	Apply `torch.atan()` to each Tensor of the input list.
`_foreach_atan_`	Apply `torch.atan()` to each Tensor of the input list.
`_foreach_ceil`	Apply `torch.ceil()` to each Tensor of the input list.
`_foreach_ceil_`	Apply `torch.ceil()` to each Tensor of the input list.
`_foreach_cos`	Apply `torch.cos()` to each Tensor of the input list.
`_foreach_cos_`	Apply `torch.cos()` to each Tensor of the input list.
`_foreach_cosh`	Apply `torch.cosh()` to each Tensor of the input list.
`_foreach_cosh_`	Apply `torch.cosh()` to each Tensor of the input list.
`_foreach_erf`	Apply `torch.erf()` to each Tensor of the input list.
`_foreach_erf_`	Apply `torch.erf()` to each Tensor of the input list.
`_foreach_erfc`	Apply `torch.erfc()` to each Tensor of the input list.
`_foreach_erfc_`	Apply `torch.erfc()` to each Tensor of the input list.
`_foreach_exp`	Apply `torch.exp()` to each Tensor of the input list.
`_foreach_exp_`	Apply `torch.exp()` to each Tensor of the input list.
`_foreach_expm1`	Apply `torch.expm1()` to each Tensor of the input list.
`_foreach_expm1_`	Apply `torch.expm1()` to each Tensor of the input list.
`_foreach_floor`	Apply `torch.floor()` to each Tensor of the input list.
`_foreach_floor_`	Apply `torch.floor()` to each Tensor of the input list.
`_foreach_log`	Apply `torch.log()` to each Tensor of the input list.
`_foreach_log_`	Apply `torch.log()` to each Tensor of the input list.
`_foreach_log10`	Apply `torch.log10()` to each Tensor of the input list.
`_foreach_log10_`	Apply `torch.log10()` to each Tensor of the input list.
`_foreach_log1p`	Apply `torch.log1p()` to each Tensor of the input list.
`_foreach_log1p_`	Apply `torch.log1p()` to each Tensor of the input list.
`_foreach_log2`	Apply `torch.log2()` to each Tensor of the input list.
`_foreach_log2_`	Apply `torch.log2()` to each Tensor of the input list.
`_foreach_neg`	Apply `torch.neg()` to each Tensor of the input list.
`_foreach_neg_`	Apply `torch.neg()` to each Tensor of the input list.
`_foreach_tan`	Apply `torch.tan()` to each Tensor of the input list.
`_foreach_tan_`	Apply `torch.tan()` to each Tensor of the input list.
`_foreach_sin`	Apply `torch.sin()` to each Tensor of the input list.
`_foreach_sin_`	Apply `torch.sin()` to each Tensor of the input list.
`_foreach_sinh`	Apply `torch.sinh()` to each Tensor of the input list.
`_foreach_sinh_`	Apply `torch.sinh()` to each Tensor of the input list.
`_foreach_round`	Apply `torch.round()` to each Tensor of the input list.
`_foreach_round_`	Apply `torch.round()` to each Tensor of the input list.
`_foreach_sqrt`	Apply `torch.sqrt()` to each Tensor of the input list.
`_foreach_sqrt_`	Apply `torch.sqrt()` to each Tensor of the input list.
`_foreach_lgamma`	Apply `torch.lgamma()` to each Tensor of the input list.
`_foreach_lgamma_`	Apply `torch.lgamma()` to each Tensor of the input list.
`_foreach_frac`	Apply `torch.frac()` to each Tensor of the input list.
`_foreach_frac_`	Apply `torch.frac()` to each Tensor of the input list.
`_foreach_reciprocal`	Apply `torch.reciprocal()` to each Tensor of the input list.
`_foreach_reciprocal_`	Apply `torch.reciprocal()` to each Tensor of the input list.
`_foreach_sigmoid`	Apply `torch.sigmoid()` to each Tensor of the input list.
`_foreach_sigmoid_`	Apply `torch.sigmoid()` to each Tensor of the input list.
`_foreach_trunc`	Apply `torch.trunc()` to each Tensor of the input list.
`_foreach_trunc_`	Apply `torch.trunc()` to each Tensor of the input list.
`_foreach_zero_`	Apply `torch.zero()` to each Tensor of the input list.

Utilities¶

`compiled_with_cxx11_abi`	Returns whether PyTorch was built with _GLIBCXX_USE_CXX11_ABI=1
`result_type`	Returns the `torch.dtype` that would result from performing an arithmetic operation on the provided input tensors.
`can_cast`	Determines if a type conversion is allowed under PyTorch casting rules described in the type promotion documentation.
`promote_types`	Returns the `torch.dtype` with the smallest size and scalar kind that is not smaller nor of lower kind than either type1 or type2.
`use_deterministic_algorithms`	Sets whether PyTorch operations must use "deterministic" algorithms.
`are_deterministic_algorithms_enabled`	Returns True if the global deterministic flag is turned on.
`is_deterministic_algorithms_warn_only_enabled`	Returns True if the global deterministic flag is set to warn only.
`set_deterministic_debug_mode`	Sets the debug mode for deterministic operations.
`get_deterministic_debug_mode`	Returns the current value of the debug mode for deterministic operations.
`set_float32_matmul_precision`	Sets the internal precision of float32 matrix multiplications.
`get_float32_matmul_precision`	Returns the current value of float32 matrix multiplication precision.
`set_warn_always`	When this flag is False (default) then some PyTorch warnings may only appear once per process.
`get_device_module`	Returns the module associated with a given device(e.g., torch.device('cuda'), "mtia:0", "xpu", ...).
`is_warn_always_enabled`	Returns True if the global warn_always flag is turned on.
`vmap`	vmap is the vectorizing map; `vmap(func)` returns a new function that maps `func` over some dimension of the inputs.
`_assert`	A wrapper around Python's assert which is symbolically traceable.

Symbolic Numbers¶

class torch.SymInt(node)[source]¶

Like an int (including magic methods), but redirects all operations on the wrapped node. This is used in particular to symbolically record operations in the symbolic shape workflow.

as_integer_ratio()[source]¶

Represent this int as an exact integer ratio

Return type: Tuple[SymInt, int]

class torch.SymFloat(node)[source]¶

Like an float (including magic methods), but redirects all operations on the wrapped node. This is used in particular to symbolically record operations in the symbolic shape workflow.

as_integer_ratio()[source]¶

Represent this float as an exact integer ratio

Return type: Tuple[int, int]

is_integer()[source]¶: Return True if the float is an integer.

class torch.SymBool(node)[source]¶

Like an bool (including magic methods), but redirects all operations on the wrapped node. This is used in particular to symbolically record operations in the symbolic shape workflow.

Unlike regular bools, regular boolean operators will force extra guards instead of symbolically evaluate. Use the bitwise operators instead to handle this.

`sym_float`	SymInt-aware utility for float casting.
`sym_int`	SymInt-aware utility for int casting.
`sym_max`	SymInt-aware utility for max which avoids branching on a < b.
`sym_min`	SymInt-aware utility for min().
`sym_not`	SymInt-aware utility for logical negation.
`sym_ite`

Export Path¶

Warning

This feature is a prototype and may have compatibility breaking changes in the future.

export generated/exportdb/index

Control Flow¶

Warning

This feature is a prototype and may have compatibility breaking changes in the future.

cond

Conditionally applies true_fn or false_fn.

Optimizations¶

compile

Optimizes given model/function using TorchDynamo and specified backend.

torch.compile documentation

Operator Tags¶

class torch.Tag¶

Members:

core

data_dependent_output

dynamic_output_shape

flexible_layout

generated

inplace_view

needs_fixed_stride_order

nondeterministic_bitwise

nondeterministic_seeded

pointwise

pt2_compliant_tag

view_copy

property name¶

torch¶

Tensors¶

Creation Ops¶

Indexing, Slicing, Joining, Mutating Ops¶

Accelerators¶

Generators¶

Random sampling¶

In-place random sampling¶

Quasi-random sampling¶

Serialization¶

Parallelism¶

Locally disabling gradient computation¶

Math operations¶

Pointwise Ops¶

Reduction Ops¶

Comparison Ops¶

Spectral Ops¶

Other Operations¶

BLAS and LAPACK Operations¶

Foreach Operations¶

Utilities¶

Symbolic Numbers¶

Export Path¶

Control Flow¶

Optimizations¶

Operator Tags¶

Docs

Tutorials

Resources