Functions — github.com/TuckerBMorgan/can-t

hub / github.com/TuckerBMorgan/can-t / functions

Functions825 in github.com/TuckerBMorgan/can-t

⨍ Functions 825◇ Types & classes 48

↓ 353 callersMethoditem

Returns the underlaying tensor as an Array

src/central/tensor.rs:307

↓ 237 callersMethodgrad

returns the underlaying grad of this tensor as an Array

src/central/tensor.rs:312

↓ 226 callersMethodset_requires_grad

Helper function to set the requires_grad bool, also sets the internal tensor as well

src/central/tensor.rs:347

↓ 184 callersMethodsum

(&self, mut axes: Vec<usize>, keep_dimensions: bool)

src/central/sum_op.rs:7

↓ 178 callersMethodbackward

sends this node backwards though the network, adding to the grad of every node that feeds into this one

src/central/tensor.rs:360

↓ 134 callersFunctionzero_all_grads

()

src/central/mod.rs:91

↓ 126 callersMethoddimensions

Returns the dimensions of the shape

src/central/shape.rs:67

↓ 66 callersFunctionget_equation

()

src/central/mod.rs:86

↓ 43 callersMethodmean

Computes the mean along specified axes # Arguments 'axes' - Vector of axis indices along which to compute the mean

src/central/mean_op.rs:8

↓ 40 callersMethodget_data_flat_buffer

Gets the underlaying backing storage of the tensor, util function for working with data of a tensor # Arguments 'tensor_id' - The tensor we are workin

src/central/equation.rs:349

↓ 35 callersMethodforward

(&mut self, inputs: Tensor)

src/nn/layer_norm.rs:37

↓ 35 callersMethodrelu

(&mut self)

src/central/relu_op.rs:4

↓ 33 callersMethodadd_tensor_grad

Adds new grad into the grad of tensor_id # Arugments 'tensor_id' - Id of for the loopup on the tensor 'grad' - the grad we are copying in

src/central/equation.rs:454

↓ 32 callersMethodreshape

Reshapes the underlaying tensor into a new shape it preforms a full copy under the hood

src/central/reshape.rs:6

↓ 30 callersMethodget_tensor_shape

Helper function to get the shape of a tensor, without getting the entire tensor # Arguments 'tensor_id' - the tensor you are getting the shape for

src/central/equation.rs:356

↓ 28 callersMethodtanh

Applies the hyperbolic tangent function element-wise to the tensor Returns a new tensor with the same shape where each element x is replaced with tanh

src/central/tanh_op.rs:8

↓ 27 callersMethodset_keep_alive

Helper function to set the keep_alive bool, also sets the internal tensor as well

src/central/tensor.rs:354

↓ 26 callersMethodtranspose

Swaps two provides axis for a tensor # Arguments 'first_index' : first index we are swapping 'second_index' : second index we are swapping

src/central/transpose_op.rs:53

↓ 24 callersFunctionapprox_eq

(a: &[f32], b: &[f32], eps: f32)

src/central/topk_op.rs:222

↓ 24 callersMethodforward

(&mut self, inputs: Tensor)

src/nn/linear.rs:61

↓ 24 callersMethodget_grad_flat_buffer

Gets the underlaying backing storage of the tensor, util function for working with grad of a tensor # Arguments 'tensor_id' - The tensor we are workin

src/central/equation.rs:342

↓ 23 callersMethodmasked_fill

(&self, mask: Tensor, value: f32)

src/central/masked_fill.rs:4

↓ 23 callersMethodpow

returns a tensors with each element raised to the provided power arguments 'power' - the power we are will raise each element to

src/central/pow_op.rs:7

↓ 22 callersMethodsoftmax

Computes the softmax function along specified axis # Arguments `axis` - Axis along which to compute the softmax

src/central/softmax_op.rs:57

↓ 20 callersMethodstd

Computes the standard deviation along specified axes # Arguments 'axes' - Vector of axis indices along which to compute the standard deviation

src/central/std_op.rs:8

↓ 18 callersMethodlog

Returns the natural log(ln, not log10) of each element it uses natural log because that is what pytorch does when you call .log on a tensor

src/central/log.rs:6

↓ 18 callersMethodselect

(&self, indices_id: TensorID)

src/central/select_op.rs:6

↓ 18 callersMethodtotal_size

Returns the total number of elements in this shape

src/central/shape.rs:56

↓ 17 callersMethodforward

(&mut self, inputs: Tensor)

src/nn/embedding.rs:24

↓ 17 callersMethodgather

(&self, dimension: usize, index: Tensor)

src/central/gather_op.rs:28

↓ 16 callersFunctioncompare_tensors

(a: Tensor, b: Tensor)

src/central/matmul_op.rs:150

↓ 16 callersMethodmax

(&self, dimension: usize, keep_dim: bool)

src/central/max_op.rs:8

↓ 15 callersMethodforward

(&mut self, inputs: Tensor)

src/nn/multihead_attention.rs:161

↓ 15 callersMethodforward

(&mut self, input: Tensor)

src/nn/tanh.rs:13

↓ 15 callersFunctionread

(id: TensorID)

src/central/topk_op.rs:384

↓ 14 callersMethodget_grad

Returns the underlaying grad of a tensor, as an array # Arugments 'id' - Id for lookup of the tensor

src/central/equation.rs:330

↓ 13 callersMethodbroadcast

Broadcasts two tensors togethers to form a new broadcast shape # Aruguments 'shape' - The shape we want to broadcast to

src/central/tensor.rs:319

↓ 13 callersMethodforward

(&mut self, inputs: Tensor)

src/nn/gpt_block.rs:96

↓ 13 callersMethodmatmul_shape

Returns a new shape that would be the result of two matrices of the provided shapes being matmul together # Arguments 'other' - the shape of ther righ

src/central/shape.rs:244

↓ 13 callersMethodnumber_of_dimension

Returns the number of dimenions of the shape

src/central/shape.rs:51

↓ 12 callersMethodtopk

( &self, k: usize, dimension: usize, _largest: bool, _sorted: bool,

src/central/topk_op.rs:87

↓ 11 callersMethodstack

(&self, others: Vec<Tensor>, dimension: usize)

src/central/stack_op.rs:6

↓ 10 callersMethodforward

( &self, query: Tensor, key: Tensor, value: Tensor, mask: Option<Tenso

src/nn/scaled_dot_project_attention.rs:19

↓ 10 callersMethodforward

(&mut self, input: Tensor)

src/nn/sequential.rs:16

↓ 10 callersMethodget_item

Returns the underlaying data of a tensor, as an array # Arugments 'id' - Id for lookup of the tensor

src/central/equation.rs:318

↓ 10 callersFunctiontensor_from

(data: Vec<f32>, dims: &[usize])

src/central/topk_op.rs:379

↓ 10 callersMethodunsqueeze

adds a dimension of 1 at the provided dimensions

src/central/shape.rs:443

↓ 10 callersFunctionupdate_parameters

(learning_rate: f32)

src/central/mod.rs:96

↓ 9 callersMethodas_ndarray_shape

Returns the cant lib shape type in the right format to passed to an ndarray ArrayD

src/central/shape.rs:424

↓ 9 callersMethodset_index

(&self, index: Indexable, new_data: f32)

src/central/index.rs:27

↓ 8 callersMethodcos

Applies the cosine function element-wise to the tensor Returns a new tensor with the same shape where each element x is replaced with cos(x)

src/central/cos_op.rs:8

↓ 8 callersMethoddiagonal

Interesting things I have noticed first and second dimensions seem to end up being sorted and have to be unique

src/central/diagonal_op.rs:6

↓ 8 callersFunctionpadding_dimenions_to_max

(in_dimension: Vec<usize>)

src/utils/mod.rs:42

↓ 8 callersMethodsigmoid

Applies the sigmoid function element-wise to the tensor Returns a new tensor with the same shape where each element x is replaced with sigmoid(x)

src/central/sigmoid_op.rs:10

↓ 8 callersMethodsin

Applies the sin function element-wise to the tensor Returns a new tensor with the same shape where each element x is replaced with sin(x)

src/central/sin_op.rs:8

↓ 7 callersMethodclamp

clamps each element of the tensor between min and max It does this into a new tensor, not in place

src/central/clamp_op.rs:6

↓ 7 callersMethodl1_loss

Computes the element-wise L1 loss (mean absolute error) between two tensors. All dimensions are reduced so the return value is a scalar loss.

src/central/l1_loss.rs:6

↓ 6 callersFunctionassert_softmax_properties

(tensor: &Tensor, _axis: usize)

src/central/softmax_op.rs:115

↓ 6 callersMethodcat

(&self, other: Tensor, dimension: usize)

src/central/cat_op.rs:6

↓ 5 callersMethodallocate_tensor

Allocates a tensor, and returns the ID of it, so it can be looked up later # Arguments 'shape' - the shape of the tensor. shape.total_size * 2 memory

src/central/equation.rs:81

↓ 5 callersMethodapply_rotary_embedding

(&self, input: Tensor, cos: Tensor, sin: Tensor)

src/nn/rotary_embedding.rs:114

↓ 5 callersMethoddetach

(&self)

src/central/tensor.rs:169

↓ 5 callersMethodgarbage_collect

(&mut self)

src/central/equation.rs:920

↓ 5 callersMethodget_parameters

(&self)

src/nn/gpt_block.rs:125

↓ 5 callersMethodmovedim

Moves the dimension at `original_location` to `new_destination`.

src/central/movedim_op.rs:19

↓ 5 callersMethodremove_index

Returns a new shape with the dimension at the provided index removed # Arguments 'index' - The 0base index to be removed from the shape

src/central/shape.rs:79

↓ 4 callersMethodadd_dimension_at_index

Returns a new shape with the provided dimension added at the provided index # Arguments 'dimension' - the dimension to be inserted 'index' - the 0base

src/central/shape.rs:161

↓ 4 callersMethodcompact_tensor_store

(&mut self)

src/central/equation.rs:886

↓ 4 callersMethodget_value

(&self, name: impl AsRef<str>)

src/utils/gguf.rs:19

↓ 4 callersFunctionlinear_index

(point: &[usize], dims: &[usize])

src/central/gather_op.rs:4

↓ 4 callersMethodmul_vector

Takes two tensors as flat buffers, multiplies them together at an elementwise level and returns the result # Arguments 'a' : The first tensor 'b' : th

src/central/equation.rs:309

↓ 4 callersMethodpermute

Reorders the tensor axes according to the provided permutation.

src/central/permute.rs:73

↓ 4 callersMethodpermute

Returns a new shape where the existing dimensions are rearranged according to the order set in permutation # Arguments 'premutation' - The order the

src/central/shape.rs:114

↓ 4 callersFunctionpermute_flat_data

(data: &[f32], dims: &[usize], permutation: &[usize])

src/central/permute.rs:20

↓ 4 callersMethodset_is_grequires_grad

Helper function for setting the internal tensor to know if it needs gradient or not # Arguments 'tensor_id' : which Tensor we are setting 'requires_gr

src/central/equation.rs:845

↓ 3 callersMethodbroadcast_shape

Returns the shape of that results from two shapes being brodcasted # Arguments 'other' - The shape we are testing against

src/central/shape.rs:382

↓ 3 callersFunctionbuild_batch_norm_dataset_from_subset

( words: &[String], stoi: &HashMap<char, usize>, )

src/lib.rs:120

↓ 3 callersMethodchunk

(&self, dimensions: usize, chunks: usize)

src/central/chunk_op.rs:7

↓ 3 callersFunctioncompare_tensors

(a: Tensor, b: Tensor)

src/central/pow_op.rs:49

↓ 3 callersMethodcompute_cos_sin

Placeholder for gathering per-position cosine/sine slices.

src/nn/rotary_embedding.rs:105

↓ 3 callersMethodcross_entropy_loss

Computes cross-entropy loss between logits and targets # Arguments 'targets' - One-hot encoded target tensor with same shape as logits

src/central/cross_entropy_op.rs:7

↓ 3 callersMethodexp

Each element is returned as e ^ x

src/central/tensor.rs:369

↓ 3 callersMethodforward

(&mut self, input: Tensor)

src/nn/gpt2.rs:133

↓ 3 callersMethodforward

(&mut self, inputs: Tensor)

src/nn/rmsnorm.rs:20

↓ 3 callersFunctionhandle_broadcasting

Handles broadcasting two operands to the right size, was doing this in a few places, put it into a util function Arugments lhs* - the left hand opeara

src/utils/mod.rs:17

↓ 3 callersFunctionreference_rmsnorm

(sample: &[f32], scale: &[f32])

src/nn/rmsnorm.rs:55

↓ 2 callersMethodallocate_zero_tensor

Allocates a tensor, and returns the ID of it, so it can be looked up later, is all zeroes # Arguments 'shape' - the shape of the tensor. shape.total_s

src/central/equation.rs:117

↓ 2 callersFunctionbuild_move_permutation

(origin: usize, destination: usize, rank: usize)

src/central/movedim_op.rs:4

↓ 2 callersMethodcan_broadcast

Checks if the two shapes are broadcastable # Arguments 'other' - The shape we are testing against

src/central/shape.rs:354

↓ 2 callersFunctioncausal_mask

(seq_len: usize)

src/nn/gpt_block.rs:5

↓ 2 callersFunctioncompare_tensors

(a: Tensor, b: Tensor)

src/central/reshape.rs:43

↓ 2 callersFunctioncompute_strides

(dims: &[usize])

src/central/permute.rs:6

↓ 2 callersFunctionfirst_image_from_array2

(arr: &ArrayD<f32>)

examples/autoencoder.rs:238

↓ 2 callersMethodgenerate_all_positions

(&self)

src/central/shape.rs:465

↓ 2 callersMethodget_parameters

(&self)

src/nn/linear.rs:69

↓ 2 callersMethodget_parameters

(&self)

src/nn/sequential.rs:24

↓ 2 callersFunctionimage_to_ascii

(image: &[f32])

examples/autoencoder.rs:220

↓ 2 callersFunctioninvert_permutation

(permutation: &[usize])

src/central/permute.rs:59

↓ 2 callersMethodlength

(&self)

src/utils/data_set.rs:69

↓ 2 callersMethodmatmul_vector

takes two matrices as flat vectors and preforms matmul of them 'a' - the flat array of the first matrix 'a_shape' - the shape of a 'b' - the flat arra

src/central/equation.rs:252

next →1–100 of 825, ranked by callers