ONNX operators

A node’s op_type names an operator of the operator set that its domain and the model’s opset_import select. The set itself is not in the file; these tables, from ONNX’s Operators.md and Operators-ml.md and the operator pages at onnx.ai, say what each name means. The “Since” column is the opset of the operator’s latest definition, as listed in November 2026. An operator named here but changed by a later opset keeps its name. A name not listed here belongs to a custom domain or to a function local to the model.

Attributes are listed by name; those names mean the same across operators and are explained in the attribute table at the end. The inputs, outputs and full type constraints of each operator are in the ONNX operator documents and are not repeated here. The explorer adds each operator’s line here to its note on op_type and on the attribute’s name, and to its table of operators used.

ai.onnx: the default set

Operator	Since	What it does	Attributes
Abs	13	Absolute value of each element.
Acos	22	Arccosine of each element.
Acosh	22	Inverse hyperbolic cosine of each element.
Add	14	Adds two tensors element by element, with broadcasting.
And	7	Logical AND, element by element, with broadcasting.
ArgMax	13	Indices of the largest elements along an axis.	axis, keepdims, select_last_index
ArgMin	13	Indices of the smallest elements along an axis.	axis, keepdims, select_last_index
Asin	22	Arcsine of each element.
Asinh	22	Inverse hyperbolic sine of each element.
Atan	22	Arctangent of each element.
Atanh	22	Inverse hyperbolic tangent of each element.
AveragePool	22	Average pooling over windows of the spatial dimensions.	auto_pad, ceil_mode, count_include_pad, dilations, kernel_shape, pads, strides
BatchNormalization	15	Batch normalisation, with stored or computed mean and variance.	epsilon, momentum, training_mode
BitCast	26	Reinterprets a tensor’s bits as another type of the same width.	to
BitShift	28	Shifts unsigned integers left or right.	direction
BitwiseAnd	18	Bitwise AND of integers, with broadcasting.
BitwiseNot	18	Bitwise NOT of integers.
BitwiseOr	18	Bitwise OR of integers, with broadcasting.
BitwiseXor	18	Bitwise XOR of integers, with broadcasting.
Cast	28	Converts a tensor to another data type.	to, saturate, round_mode
Ceil	13	Rounds each element up.
Col2Im	18	Rebuilds images from columns of blocks.	dilations, pads, strides
Compress	28	Selects elements where a boolean condition is true.	axis
Concat	13	Joins tensors along an existing axis.	axis
ConcatFromSequence	11	Joins a sequence of tensors, optionally along a new axis.	axis, new_axis
Constant	25	Produces a constant given in its attributes.	sparse_value, value, value_float, value_floats, value_int, value_ints, value_string, value_strings
ConstantOfShape	25	A tensor of a given shape, filled with one value.	value
Conv	22	Convolution over the spatial dimensions.	auto_pad, dilations, group, kernel_shape, pads, strides
ConvInteger	10	Convolution of integers with zero points.	auto_pad, dilations, group, kernel_shape, pads, strides
ConvTranspose	22	Transposed convolution.	auto_pad, dilations, group, kernel_shape, output_padding, output_shape, pads, strides
Cos	22	Cosine of each element.
Cosh	22	Hyperbolic cosine of each element.
CumProd	26	Running product along an axis.	exclusive, reverse
CumSum	14	Running sum along an axis.	exclusive, reverse
DFT	20	Discrete Fourier transform.	inverse, onesided
DeformConv	22	Deformable convolution.	dilations, group, kernel_shape, pads, strides
DequantizeLinear	28	Converts a quantised tensor back to full precision.	axis, block_size
Det	22	Determinant of square matrices.
Div	14	Divides two tensors element by element.
Dropout	22	Randomly zeroes elements when training.	seed
Einsum	28	Einstein summation.	equation
Equal	19	True where elements are equal.
Erf	13	Error function.
Exp	13	Exponential.
Expand	13	Broadcasts to a given shape.
EyeLike	22	Ones on a diagonal, shaped like the input.	dtype, k
Flatten	25	Flattens a tensor into a matrix.	axis
Floor	13	Rounds each element down.
GRU	22	Gated recurrent unit network.	activation_alpha, activation_beta, activations, clip, direction, hidden_size, layout, linear_before_reset
Gather	13	Takes slices along an axis.	axis
GatherElements	13	Takes elements at per-element indices.	axis
GatherND	13	Takes slices by multi-dimensional indices.	batch_dims
GlobalAveragePool	22	Averages over all spatial dimensions.
GlobalLpPool	22	Lp norm over all spatial dimensions.	p
GlobalMaxPool	22	Maximum over all spatial dimensions.
Greater	13	True where the first is larger.
GridSample	22	Samples a tensor at positions given by a grid.	align_corners, mode, padding_mode
Hardmax	13	One at the largest element of an axis.	axis
Identity	25	Passes its input through.
If	25	Runs one of two subgraphs.	else_branch, then_branch
ImageDecoder	20	Decodes encoded image bytes.	pixel_format
InstanceNormalization	22	Normalises each sample and channel.	epsilon
IsInf	20	True where an element is infinite.	detect_negative, detect_positive
IsNaN	20	True where an element is not a number.
LRN	13	Local response normalisation.	alpha, beta, bias, size
LSTM	22	Long short-term memory network.	activation_alpha, activation_beta, activations, clip, direction, hidden_size, input_forget, layout
Less	13	True where the first is smaller.
Log	13	Natural logarithm.
Loop	25	Runs a subgraph repeatedly.	body
LpNormalization	22	Normalises by the Lp norm.	axis, p
LpPool	22	Lp pooling over windows.	auto_pad, ceil_mode, dilations, kernel_shape, p, pads, strides
MatMul	13	Matrix multiplication, as NumPy’s matmul.
MatMulInteger	10	Matrix multiplication of integers.
Max	13	Element-wise maximum.
MaxPool	22	Maximum pooling over windows.	auto_pad, ceil_mode, dilations, kernel_shape, pads, storage_order, strides
MaxRoiPool	22	Maximum pooling over regions of interest.	pool_shape, roi_scale
MaxUnpool	22	Puts values back at the pooled positions.	kernel_shape, pads, strides
Mean	13	Element-wise mean.
MelWeightMatrix	17	Weights from a linear spectrum to mel bands.	output_datatype
Min	13	Element-wise minimum.
Mod	28	Remainder of division.	fmod
Mul	14	Multiplies two tensors element by element.
Multinomial	22	Samples class indices.	dtype, sample_size, seed
Neg	13	Negates each element.
NonMaxSuppression	11	Picks boxes by score, dropping overlaps.	center_point_box
NonZero	13	Indices of non-zero elements.
Not	1	Logical NOT.
OneHot	28	Turns indices into one-hot vectors.	axis
Optional	28	Makes an optional, holding the input or empty.	type
OptionalGetElement	28	Gets the value out of an optional.
OptionalHasElement	28	True if an optional holds a value.
Or	7	Logical OR.
Pad	25	Pads a tensor by given amounts.	mode
Pow	15	Raises to a power.
QLinearConv	10	Convolution of quantised tensors.	auto_pad, dilations, group, kernel_shape, pads, strides
QLinearMatMul	21	Matrix multiplication of quantised tensors.
QuantizeLinear	28	Converts to integers with a scale and zero point.	axis, block_size, output_dtype, precision
RNN	22	Plain recurrent network.	activation_alpha, activation_beta, activations, clip, direction, hidden_size, layout
RandomNormal	22	Normally distributed random values.	dtype, mean, scale, seed, shape
RandomNormalLike	22	Normal random values, shaped like the input.	dtype, mean, scale, seed
RandomUniform	22	Uniform random values.	dtype, high, low, seed, shape
RandomUniformLike	22	Uniform random values, shaped like the input.	dtype, high, low, seed
Reciprocal	13	One divided by each element.
ReduceMax	20	Largest along given axes.	keepdims, noop_with_empty_axes
ReduceMean	18	Mean along given axes.	keepdims, noop_with_empty_axes
ReduceMin	20	Smallest along given axes.	keepdims, noop_with_empty_axes
ReduceProd	18	Product along given axes.	keepdims, noop_with_empty_axes
ReduceSum	13	Sum along given axes.	keepdims, noop_with_empty_axes
RegexFullMatch	20	True where a whole string matches a pattern.	pattern
Reshape	25	Reshapes a tensor.	allowzero
Resize	19	Resizes a tensor by sizes or scales.	antialias, axes, coordinate_transformation_mode, cubic_coeff_a, exclude_outside, extrapolation_value, keep_aspect_ratio_policy, mode, nearest_mode
ReverseSequence	28	Reverses each sequence of a batch.	batch_axis, time_axis
RoiAlign	22	Crops regions of interest to one size.	coordinate_transformation_mode, mode, output_height, output_width, sampling_ratio
Round	22	Rounds to the nearest integer, halves to even.
STFT	17	Short-time Fourier transform.	onesided
Scan	25	Runs a subgraph over sequences, with state.	body, num_scan_inputs, scan_input_axes, scan_input_directions, scan_output_axes, scan_output_directions
Scatter	11	Deprecated: use ScatterElements.	axis
ScatterElements	18	Writes values at per-element indices.	axis, reduction
ScatterND	18	Writes values at multi-dimensional indices.	reduction
SequenceAt	11	Gets one tensor from a sequence.
SequenceConstruct	11	Makes a sequence from tensors.
SequenceEmpty	11	Makes an empty sequence.	dtype
SequenceErase	11	Removes one tensor from a sequence.
SequenceInsert	11	Inserts a tensor into a sequence.
SequenceLength	11	The number of tensors in a sequence.
Shape	25	The shape of a tensor, as a tensor.	end, start
Sigmoid	13	Sigmoid function.
Sign	13	The sign of each element.
Sin	22	Sine.
Sinh	22	Hyperbolic sine.
Size	25	The number of elements.
Slice	13	Takes a slice along given axes.
Split	18	Splits a tensor into pieces.	axis, num_outputs
SplitToSequence	24	Splits a tensor into a sequence.	axis, keepdims
Sqrt	13	Square root.
Squeeze	25	Removes dimensions of size 1.
StringConcat	20	Joins strings.
StringNormalizer	10	Strips stop words and changes case.	case_change_action, is_case_sensitive, locale, stopwords
StringSplit	20	Splits strings on a delimiter.	delimiter, maxsplit
Sub	14	Subtracts two tensors element by element.
Sum	13	Element-wise sum.
Tan	22	Tangent.
Tanh	13	Hyperbolic tangent.
TensorScatter	24	Writes updates into a fixed-size key/value cache.	axis, mode
TfIdfVectorizer	9	Counts n-grams, weighted by TF-IDF.	max_gram_length, max_skip_count, min_gram_length, mode, ngram_counts, ngram_indexes, pool_int64s, pool_strings, weights
Tile	13	Repeats a tensor along each dimension.
TopK	24	The k largest or smallest along an axis.	axis, largest, sorted
Transpose	25	Permutes dimensions.	perm
Trilu	14	The upper or lower triangle.	upper
Unique	28	The unique elements.	axis, sorted
Unsqueeze	25	Inserts dimensions of size 1.
Upsample	10	Deprecated: use Resize.	mode
Where	16	Picks elements by a condition.
Xor	7	Logical exclusive OR.

ai.onnx functions

These are defined by a body of other operators, so a runtime that lacks one can still run it by expanding it.

Operator	Since	What it does	Attributes
AffineGrid	20	A sampling grid from affine matrices.	align_corners
Attention	25	Scaled dot-product attention.	is_causal, kv_num_heads, left_window_size, q_num_heads, qk_matmul_output_mode, right_window_size, scale, softcap, softmax_precision
Bernoulli	22	Samples 0 or 1 by per-element probability.	dtype, seed
BlackmanWindow	17	A Blackman window.	output_datatype, periodic
CastLike	25	Casts to the type of a second tensor.	saturate
CausalConvWithState	27	A stateful, causal, depthwise 1-D convolution.	activation
Celu	28	Continuously differentiable exponential linear unit.	alpha
CenterCropPad	18	Crops or pads around the centre.	axes
Clip	13	Limits to a minimum and maximum.
DepthToSpace	28	Moves channel data into spatial blocks.	blocksize, mode
DynamicQuantizeLinear	11	Quantises, working out the scale.
Elu	22	Exponential linear unit.	alpha
Gelu	20	Gaussian error linear unit.	approximate
Gemm	29	General matrix multiplication: alpha × A × B + beta × C.	alpha, beta, transA, transB
GreaterOrEqual	16	True where the first is larger or equal.
GroupNormalization	21	Normalises over groups of channels.	epsilon, num_groups, stash_type
GroupedMatMul	29	Multiplies each token by its groups’ weight matrices.
HammingWindow	17	A Hamming window.	output_datatype, periodic
HannWindow	17	A Hann window.	output_datatype, periodic
HardSigmoid	22	Piecewise-linear sigmoid.	alpha, beta
HardSwish	22	Piecewise swish.
LayerNormalization	17	Normalises over the last dimensions.	axis, epsilon, stash_type
LeakyRelu	16	Rectified unit with a small slope below zero.	alpha
LessOrEqual	16	True where the first is smaller or equal.
LinearAttention	27	Linear attention recurrence for decoding and prefill.	chunk_size, kv_num_heads, q_num_heads, scale, update_rule
LogSoftmax	13	Logarithm of the softmax.	axis
MeanVarianceNormalization	13	Zero mean, unit variance over given axes.	axes
Mish	22	The mish activation function.
NegativeLogLikelihoodLoss	22	Negative log likelihood.	ignore_index, reduction
PRelu	16	Rectified unit with a learned slope.
RMSNormalization	23	Divides by the root mean square.	axis, epsilon, stash_type
Range	27	A sequence of numbers.
ReduceL1	18	L1 norm along axes.	keepdims, noop_with_empty_axes
ReduceL2	18	L2 norm along axes.	keepdims, noop_with_empty_axes
ReduceLogSum	28	Logarithm of the sum.	keepdims, noop_with_empty_axes
ReduceLogSumExp	28	Logarithm of the sum of exponentials.	keepdims, noop_with_empty_axes
ReduceSumSquare	18	Sum of squares.	keepdims, noop_with_empty_axes
Relu	14	Rectified linear unit.
RotaryEmbedding	23	Rotary positional embeddings.	interleaved, num_heads, rotary_embedding_dim
Selu	22	Scaled exponential linear unit.	alpha, gamma
SequenceMap	17	Runs a subgraph over sequences.	body
Shrink	9	Shrinks values towards zero.	bias, lambd
Softmax	13	Normalises to probabilities.	axis
SoftmaxCrossEntropyLoss	13	Softmax cross-entropy.	ignore_index, reduction
Softplus	22	The softplus function.
Softsign	22	The softsign function.
SpaceToDepth	28	Moves spatial blocks into channels.	blocksize
SwiGLU	28	Swish of a gate input, times a second input.	alpha
Swish	24	x × sigmoid(alpha × x).	alpha
ThresholdedRelu	22	Passes values above a threshold.	alpha

ai.onnx.preview and ai.onnx.preview.training

Domain	Operator	Since	What it does	Attributes
ai.onnx.preview	FlexAttention	1	Experimental attention whose scores and probabilities two subgraphs can modify.	prob_mod, scale, score_mod, softmax_precision
ai.onnx.preview.training	Adagrad	1	One Adagrad step.	decay_factor, epsilon, norm_coefficient
ai.onnx.preview.training	Adam	1	One Adam step.	alpha, beta, epsilon, norm_coefficient, norm_coefficient_post
ai.onnx.preview.training	Gradient	1	Gradients of outputs with respect to inputs.	xs, y, zs
ai.onnx.preview.training	Momentum	1	One momentum step.	alpha, beta, mode, norm_coefficient

ai.onnx.ml: classical machine learning

Operator	Since	What it does	Attributes
ArrayFeatureExtractor	1	Selects elements by indices from the last axes.
Binarizer	1	Maps values to 0 or 1 by a threshold.	threshold
CastMap	1	Turns a map into a tensor, by key order.	cast_to, map_form, max_map
CategoryMapper	1	Converts strings to integers and back.	cats_int64s, cats_strings, default_int64, default_string
DictVectorizer	1	Turns a dictionary into a 1-D array.	int64_vocabulary, string_vocabulary
FeatureVectorizer	1	Joins tensors into one output.	inputdimensions
Imputer	1	Replaces one value with another.	imputed_value_floats, imputed_value_int64s, replaced_value_float, replaced_value_int64
LabelEncoder	4	Maps each element through keys and values.	default_float, default_int64, default_string, default_tensor, keys_floats, keys_int64s, keys_strings, keys_tensor, values_floats, values_int64s, values_strings, values_tensor
LinearClassifier	1	A linear classifier.	classlabels_ints, classlabels_strings, coefficients, intercepts, multi_class, post_transform
LinearRegressor	1	A generalised linear regression.	coefficients, intercepts, post_transform, targets
Normalizer	1	Normalises by MAX, L1 or L2.	norm
OneHotEncoder	1	Turns a category into ones and zeros.	cats_int64s, cats_strings, zeros
SVMClassifier	1	A support vector machine classifier.	classlabels_ints, classlabels_strings, coefficients, kernel_params, kernel_type, post_transform, prob_a, prob_b, rho, support_vectors, vectors_per_class
SVMRegressor	1	Support vector regression and anomaly detection.	coefficients, kernel_params, kernel_type, n_supports, one_class, post_transform, rho, support_vectors
Scaler	1	Rescales data.	offset, scale
TreeEnsemble	5	Regression values from an ensemble of trees.	aggregate_function, leaf_targetids, leaf_weights, membership_values, n_targets, nodes_falseleafs, nodes_falsenodeids, nodes_featureids, nodes_hitrates, nodes_missing_value_tracks_true, nodes_modes, nodes_splits, nodes_trueleafs, nodes_truenodeids, post_transform, tree_roots
TreeEnsembleClassifier	5	Deprecated: the top class of an ensemble of trees. Use TreeEnsemble with ArgMax.	base_values, class_ids, class_nodeids, class_treeids, class_weights, classlabels_int64s, classlabels_strings, nodes_falsenodeids, nodes_featureids, nodes_hitrates, nodes_missing_value_tracks_true, nodes_modes, nodes_nodeids, nodes_treeids, nodes_truenodeids, nodes_values, post_transform, and the _as_tensor forms of the numeric lists
TreeEnsembleRegressor	5	Deprecated: use TreeEnsemble. Regression from an ensemble of trees.	aggregate_function, base_values, n_targets, nodes_falsenodeids, nodes_featureids, nodes_hitrates, nodes_missing_value_tracks_true, nodes_modes, nodes_nodeids, nodes_treeids, nodes_truenodeids, nodes_values, post_transform, target_ids, target_nodeids, target_treeids, target_weights, and the _as_tensor forms
ZipMap	1	Makes a map from input values and given keys.	classlabels_int64s, classlabels_strings

What the attributes mean

Each name above has one meaning across the operators that use it. Attribute names not listed below are specific to one operator and are described in its document.

Name	Meaning
auto_pad	How padding is worked out: NOTSET, SAME_UPPER, SAME_LOWER or VALID.
axis, axes	The axis or axes to act on; negative counts from the last.
batch_dims	How many leading dimensions are batch dimensions.
body	The subgraph to run.
ceil_mode	Whether to use ceiling rather than floor when computing the output shape.
coordinate_transformation_mode	How output coordinates map to input coordinates.
count_include_pad	Whether padded elements count in the average.
dilations	The spacing between kernel elements along each axis.
direction	RIGHT or LEFT for shifts; forward, reverse or bidirectional for recurrent cells.
dtype, output_datatype, output_dtype	The data type of the output, a TensorProto.DataType.
else_branch, then_branch	The subgraphs run when a condition is false or true.
epsilon	A small constant that keeps the arithmetic stable.
group	How many groups of channels the input is split into.
hidden_size	The number of units in the hidden layer.
keepdims	Whether reduced dimensions stay, with size 1.
kernel_shape	The size of the kernel along each axis.
mean, scale, seed	Parameters of a random distribution; for Scaler, the factor.
mode	Which variant of the operation to use.
noop_with_empty_axes	Whether empty axes mean doing nothing rather than all axes.
p	The order of the norm.
pads	Padding at the start and end of each axis.
perm	The new order of the dimensions.
post_transform	What to apply to the scores: NONE, SOFTMAX, LOGISTIC, SOFTMAX_ZERO or PROBIT.
reduction	How results combine: none, sum, mean, min, max or mul, as allowed.
round_mode, saturate	How out-of-range values are handled when converting.
shape	The shape of the output.
stash_type	The type used for the first stage of the calculation.
strides	The step of the window along each axis.
to	The target data type.
value, value_float, value_int, value_string and the plurals	The constant.

Versioning

A model selects operator definitions by its opset_import: for each domain, an operator means the definition in force at the version named there. An operator changed in opset 22 still means its old definition in a model that imports opset 13. The explorer does not know the historical definitions: it gives each operator’s latest description and its attribute names as listed above, which earlier opsets may differ from in detail (for example Slice’s bounds moved from attributes to inputs at opset 10, and axes attributes became inputs at opsets 13 and 18). For an operator’s exact inputs, outputs and types at a given opset, see ONNX’s Operators.md and Changelog.md.