ONNX

[szha]

We at ONNX are working on the roadmap and I'm proposing that we adopt the numpy array API definition and eventually the array API standard from the outcome of this work group. ONNX is a model exchange standard for the deep learning frameworks. The array API comparison exercise here is very helpful and it would be great if we could include the ONNX in the comparison here.

The definition of array API functions can be found here.

[kgryte]

@szha Thanks for the links! I'll look into the ONNX array API functions, and at a minimum add it to the "join" dataset, which compares across all libraries.

[szha]

@kgryte did you get a chance to look at ONNX? Let me know if you need anything else.

[rgommers]

I suspect part of the hold up here is that ONNX doesn't use Sphinx, so the tooling may not be as easily adjusted. Still would be good to include in the overview.

[szha]

	A	B	C	D	E	F	G	H	I
1	numpy	cupy	dask.array	jax	mxnet	pytorch	tensorflow	onnx runtime	ONNX Extra Operator
2	numpy.absolute	absolute		absolute	absolute	abs	math.abs	Abs	AveragePool
3	numpy.add	add		add	add	add	math.add	Add	BatchNormalization
4	numpy.all	all	all	alltrue			reduce_all		Cast
5	numpy.allclose	allclose	allclose	allclose		allclose			Compress
6	numpy.amax	max	max	max	amax		math.reduce_max		Constant
7	numpy.amin	min	min	min	amin		math.reduce_min		ConstantOfShape
8	numpy.angle	angle	angle	angle	angle	angle	math.angle		Conv
9	numpy.any	any	any	sometrue			math.reduce_any		ConvInteger
10	numpy.arange	arange	arange	arange	arange	range	range		ConvTranspose
11	numpy.arccos	arccos	arccos	arccos	arccos	acos	math.acos	Acos	DepthToSpace
12	numpy.arccosh	arccosh	arccosh	arccosh	arccosh		math.acosh	Acosh	DequantizeLinear
13	numpy.arcsin	arcsin	arcsin	arcsin	arcsin	asin	math.asin	Asin	Dropout
14	numpy.arcsinh	arcsinh	arcsinh	arcsinh	arcsinh		math.asinh	Asinh	Elu
15	numpy.arctan	arctan	arctan	arctan	arctan	atan	math.atan	Atan	Erf
16	numpy.arctan2	arctan2	arctan2	arctan2	arctan2	atan2	math.atan2		EyeLike
17	numpy.arctanh	arctanh	arctanh	arctanh	arctanh		math.atanh	Atanh	Flatten
18	numpy.argmax	argmax	argmax	argmax	argmax	argmax	math.argmax	ArgMax	GRU
19	numpy.argmin	argmin	argmin	argmin	argmin	argmin	math.argmin	ArgMin	Gather
20	numpy.argsort	argsort		argsort	argsort	argsort	argsort		GatherElements
21	numpy.around	around	round	around	around				GatherND
22	numpy.array	array	array	array	array	tensor	make_ndarray		Gemm
23	numpy.asarray	asarray	asarray	asarray	asarray		convert_to_tensor		GlobalAveragePool
24	numpy.atleast_1d	atleast_1d	atleast_1d	atleast_1d	atleast_1d				GlobalLpPool
25	numpy.atleast_2d	atleast_2d	atleast_2d	atleast_2d	atleast_2d				GlobalMaxPool
26	numpy.atleast_3d	atleast_3d	atleast_3d	atleast_3d	atleast_3d				HardSigmoid
27	numpy.bincount	bincount	bincount		bincount	bincount	math.bincount		Hardmax
28	numpy.bitwise_and	bitwise_and	bitwise_and	bitwise_and		bitwise_and	bitwise.bitwise_and
29	numpy.bitwise_or	bitwise_or	bitwise_or	bitwise_or		bitwise_or	bitwise.bitwise_or		If
30	numpy.bitwise_xor	bitwise_xor	bitwise_xor	bitwise_xor		bitwise_xor	bitwise.bitwise_xor		InstanceNormalization
31	numpy.broadcast_arrays	broadcast_arrays	broadcast_arrays	broadcast_arrays	broadcast_arrays	broadcast_tensors			LRN
32	numpy.broadcast_to	broadcast_to	broadcast_to	broadcast_to	broadcast_to		broadcast_to	Expand	LSTM
33	numpy.ceil	ceil	ceil	ceil	ceil	ceil	math.ceil	Ceil	LeakyRelu
34	numpy.clip	clip	clip	clip	clip		clip_by_value	Clip	LogSoftmax
35	numpy.concatenate	concatenate	concatenate	concatenate	concatenate	cat	concat	Concat	Loop
36	numpy.conj	conj	conj	conj	conj	conj	math.conj		LpNormalization
37	numpy.copysign	copysign	copysign	copysign	copysign				LpPool
38	numpy.corrcoef	corrcoef	corrcoef	corrcoef	corrcoef				MaxPool
39	numpy.cos	cos	cos	cos	cos	cos	math.cos	Cos	MaxRoiPool
40	numpy.cosh	cosh	cosh	cosh	cosh	cosh	math.cosh	Cosh	MaxUnpool
41	numpy.count_nonzero	count_nonzero	count_nonzero	count_nonzero	count_nonzero		math.count_nonzero		Multinomial
42	numpy.cov	cov	cov	cov	cov				NonMaxSuppression
43	numpy.cross	cross		cross	cross	cross	linalg.cross		NonZero
44	numpy.cumprod	cumprod	cumprod	cumproduct	cumprod	cumprod	math.cumprod		OneHot
45	numpy.cumsum	cumsum	cumsum	cumsum	cumsum	cumsum	math.cumsum	CumSum	PRelu
46	numpy.deg2rad	deg2rad	deg2rad	deg2rad	deg2rad				QLinearConv
47	numpy.degrees	degrees	degrees	degrees	degrees				QLinearMatMul
48	numpy.diag	diag	diag	diag	diag	diag	linalg.diag		QuantizeLinear
49	numpy.diagonal	diagonal	diagonal	diagonal	diagonal	diagonal	linalg.diag_part		RNN
50	numpy.divide	divide		divide	divide	div	math.divide	Div	RandomNormal
51	numpy.dot	dot	dot	dot	dot				RandomNormalLike
52	numpy.dstack	dstack	dstack	dstack	dstack			ConcatFromSequence	RandomUniform
53	numpy.einsum	einsum	einsum	einsum	einsum	einsum	einsum	Einsum	RandomUniformLike
54	numpy.empty	empty	empty	empty	empty	empty			Reciprocal
55	numpy.empty_like	empty_like	empty_like	empty_like	empty_like	empty_like			ReduceL1
56	numpy.equal	equal		equal		eq	math.equal	Equal	ReduceL2
57	numpy.exp	exp	exp	exp	exp	exp	math.exp	Exp	ReduceLogSum
58	numpy.expand_dims	expand_dims		expand_dims	expand_dims	unsqueeze	expand_dims	Unsqueeze	ReduceLogSumExp
59	numpy.expm1	expm1	expm1	expm1	expm1	expm1	math.expm1		ReduceMax
60	numpy.eye	eye	eye	eye	eye	eye	eye	Identity	ReduceMean
61	numpy.fft.fft	fft.fft	fft.fft	fft		fft	signal.fft		ReduceMin
62	numpy.fft.fft2	fft.fft2	fft.fft2	fft2			signal.fft2d		ReduceProd
63	numpy.fft.fftn	fft.fftn	fft.fftn	fftn			signal.fft3d		ReduceSum
64	numpy.fft.fftshift	fft.fftshift	fft.fftshift	fftshift			signal.fftshift	BitShift	ReduceSumSquare
65	numpy.fft.ifft	fft.ifft	fft.ifft	ifft		ifft	signal.ifft		Relu
66	numpy.fft.ifft2	fft.ifft2	fft.ifft2	ifft2			signal.ifft2d		Resize
67	numpy.fft.ifftn	fft.ifftn	fft.ifftn	ifftn			signal.ifft3d		ReverseSequence
68	numpy.fft.ifftshift	fft.ifftshift	fft.ifftshift	ifftshift			signal.ifftshift	BitShift	RoiAlign
69	numpy.fft.irfft	fft.irfft	fft.irfft	irfft		irfft	signal.irfft		Scan
70	numpy.fft.irfft2	fft.irfft2	fft.irfft2	irfft2			signal.irfft2d		Scatter (deprecated)
71	numpy.fft.irfftn	fft.irfftn	fft.irfftn	irfftn			signal.irfft3d		ScatterElements
72	numpy.fft.rfft	fft.rfft	fft.rfft	rfft		rfft	signal.rfft		ScatterND
73	numpy.fft.rfft2	fft.rfft2	fft.rfft2	rfft2			signal.rfft2d		Selu
74	numpy.fft.rfftn	fft.rfftn	fft.rfftn	rfftn			signal.rfft3d		SequenceAt
75	numpy.fix	fix	fix	fix	fix				SequenceConstruct
76	numpy.flip	flip	flip	flip	flip	flip	reverse		SequenceEmpty
77	numpy.fliplr	fliplr	fliplr	fliplr	fliplr				SequenceErase
78	numpy.flipud	flipud	flipud	flipud	flipud				SequenceInsert
79	numpy.floor	floor	floor	floor	floor	floor	math.floor	Floor	SequenceLength
80	numpy.floor_divide	floor_divide		floor_divide	floor_divide	floor_divide	math.floordiv		Shape
81	numpy.fmax	fmax	fmax	fmax	fmax				Shrink
82	numpy.fmin	fmin	fmin	fmin	fmin				Sigmoid
83	numpy.fmod	fmod	fmod	fmod	fmod	fmod			Size
84	numpy.frexp	frexp	frexp	frexp	frexp				Slice
85	numpy.full	full	full	full	full	full	fill		Softmax
86	numpy.full_like	full_like	full_like	full_like	full_like	full_like			Softplus
87	numpy.greater	greater		greater		gt	math.greater	Greater	Softsign
88	numpy.greater_equal	greater_equal		greater_equal		ge	math.greater_equal		SpaceToDepth
89	numpy.histogram	histogram	histogram		histogram	histc			SplitToSequence
90	numpy.hstack	hstack	hstack	hstack	hstack			ConcatFromSequence	StringNormalizer
91	numpy.hypot	hypot	hypot	hypot	hypot				TfIdfVectorizer
92	numpy.imag	imag	imag	imag	imag	imag	math.imag		ThresholdedRelu
93	numpy.inner	inner		inner	inner	dot			TopK
94	numpy.iscomplex	iscomplex	iscomplex	iscomplex		is_complex			Upsample (deprecated)
95	numpy.isfinite	isfinite	isfinite	isfinite		isfinite	math.is_finite		Xor
96	numpy.isinf	isinf	isinf	isinf		isinf		IsInf	Function
97	numpy.isnan	isnan	isnan	isnan		isnan	math.is_nan	IsNaN	Celu
98	numpy.isreal	isreal	isreal	isreal		is_floating_point			DynamicQuantizeLinear
99	numpy.ldexp	ldexp	ldexp	ldexp	ldexp				GreaterOrEqual
100	numpy.less	less		less		lt	math.less	Less	LessOrEqual
101	numpy.less_equal	less_equal		less_equal		le	math.less_equal		MeanVarianceNormalization
102	numpy.linalg.cholesky	linalg.cholesky	linalg.cholesky	cholesky	linalg.cholesky	cholesky	linalg.cholesky		NegativeLogLikelihoodLoss
103	numpy.linalg.det	linalg.det		det	linalg.det	det	linalg.det	Det	Range
104	numpy.linalg.eig			eig	linalg.eig	eig	linalg.eigh		SoftmaxCrossEntropyLoss
105	numpy.linalg.eigvalsh	linalg.eigvalsh		eigvalsh		symeig	linalg.eigvalsh
106	numpy.linalg.inv	linalg.inv	linalg.inv	inv	linalg.inv	inverse	linalg.inv
107	numpy.linalg.lstsq	linalg.lstsq	linalg.lstsq		linalg.lstsq	lstsq	linalg.lstsq
108	numpy.linalg.matrix_power	linalg.matrix_power		matrix_power	linalg.matrix_power	matrix_power
109	numpy.linalg.matrix_rank	linalg.matrix_rank		matrix_rank	linalg.matrix_rank	matrix_rank	rank
110	numpy.linalg.norm	linalg.norm	linalg.norm	norm	linalg.norm	norm	norm
111	numpy.linalg.pinv	linalg.pinv		pinv	linalg.pinv		linalg.pinv
112	numpy.linalg.qr	linalg.qr	linalg.qr	qr	linalg.qr	qr	linalg.qr
113	numpy.linalg.slogdet	linalg.slogdet		slogdet	linalg.slogdet	slogdet	linalg.slogdet
114	numpy.linalg.solve	linalg.solve	linalg.solve	solve	linalg.solve	solve	linalg.solve
115	numpy.linalg.svd	linalg.svd	linalg.svd	svd	linalg.svd	svd	linalg.svd
116	numpy.linspace	linspace	linspace	linspace	linspace	linspace	linspace
117	numpy.log	log	log	log	log	log	math.log	Log
118	numpy.log10	log10	log10	log10	log10	log10
119	numpy.log1p	log1p	log1p	log1p	log1p	log1p	math.log1p
120	numpy.log2	log2	log2	log2	log2	log2
121	numpy.logaddexp	logaddexp	logaddexp	logaddexp	logaddexp	logsumexp	math.reduce_logsumexp
122	numpy.logaddexp2	logaddexp2	logaddexp2	logaddexp2	logaddexp2
123	numpy.logical_and	logical_and	logical_and	logical_and		logical_and	math.logical_and	And
124	numpy.logical_not	logical_not	logical_not	logical_not		logical_not	math.logical_not	Not
125	numpy.logical_or	logical_or	logical_or	logical_or		logical_or	math.logical_or	Or
126	numpy.logical_xor	logical_xor	logical_xor	logical_xor		logical_xor	math.logical_xor
127	numpy.logspace	logspace		logspace	logspace	logspace
128	numpy.matmul	matmul	matmul	matmul	matmul	mm	linalg.matmul	MatMul, MatMulInteger
129	numpy.maximum	maximum	maximum	maximum	maximum	max	math.maximum	Max
130	numpy.mean	mean	mean	mean	mean	mean	math.reduce_mean	Mean
131	numpy.median		median	median	median	median
132	numpy.meshgrid	meshgrid	meshgrid	meshgrid	meshgrid	meshgrid	meshgrid
133	numpy.minimum	minimum	minimum	minimum	minimum	min	math.minimum	Min
134	numpy.mod	mod		mod	mod		math.floormod	Mod
135	numpy.moveaxis	moveaxis	moveaxis	moveaxis	moveaxis
136	numpy.multiply	multiply		multiply	multiply	mul	math.multiply	Mul
137	numpy.nan_to_num	nan_to_num	nan_to_num	nan_to_num	nan_to_num
138	numpy.nanmax	nanmax	nanmax	nanmax	nanmax
139	numpy.nanmin	nanmin	nanmin	nanmin	nanmin
140	numpy.nanprod	nanprod	nanprod	nanprod	nanprod
141	numpy.nansum	nansum	nansum	nansum	nansum
142	numpy.negative	negative		negative	negative	neg	math.negative	Neg
143	numpy.nextafter	nextafter	nextafter	nextafter	nextafter		math.nextafter
144	numpy.nonzero	nonzero	nonzero	nonzero	nonzero	nonzero
145	numpy.not_equal	not_equal		not_equal		ne	math.not_equal
146	numpy.ones	ones	ones	ones	ones	ones	ones
147	numpy.ones_like	ones_like	ones_like	ones_like	ones_like	ones_like	ones_like
148	numpy.outer	outer	outer	outer	outer	ger
149	numpy.pad	pad	pad	pad			pad	Pad
150	numpy.percentile	percentile	percentile	percentile	percentile
151	numpy.power	power		power	power	pow	math.pow	Pow
152	numpy.prod	prod	prod	product	prod	prod	math.reduce_prod
153	numpy.rad2deg	rad2deg	rad2deg	rad2deg	rad2deg
154	numpy.radians	radians	radians	radians	radians
155	numpy.ravel	ravel	ravel	ravel	ravel
156	numpy.real	real	real	real	real	real	math.real
157	numpy.reciprocal	reciprocal		reciprocal	reciprocal		math.reciprocal
158	numpy.remainder	remainder		remainder	remainder	remainder
159	numpy.repeat	repeat	repeat	repeat	repeat		repeat
160	numpy.reshape	reshape	reshape	reshape	reshape	reshape	reshape	Reshape
161	numpy.result_type	result_type	result_type	result_type		result_type
162	numpy.rint	rint	rint	rint	rint		math.rint
163	numpy.roll	roll	roll	roll	roll	roll	roll
164	numpy.rollaxis	rollaxis	rollaxis	rollaxis	rollaxis
165	numpy.rot90	rot90		rot90	rot90	rot90
166	numpy.round_	round_		round	round_	round	math.round	Round
167	numpy.sign	sign	sign	sign	sign	sign	math.sign	Sign
168	numpy.signbit	signbit	signbit	signbit	signbit
169	numpy.sin	sin	sin	sin	sin	sin	math.sin	Sin
170	numpy.sinh	sinh	sinh	sinh	sinh	sinh	math.sinh	Sinh
171	numpy.sort	sort		sort	sort	sort	sort
172	numpy.split	split		split	split	split	split	Split
173	numpy.sqrt	sqrt	sqrt	sqrt	sqrt	sqrt	math.sqrt	Sqrt
174	numpy.square	square	square	square	square	square	math.square
175	numpy.squeeze	squeeze	squeeze	squeeze	squeeze	squeeze	squeeze	Squeeze
176	numpy.stack	stack	stack	stack	stack	stack	stack	ConcatFromSequence
177	numpy.std	std	std	std	std	std	math.reduce_std
178	numpy.subtract	subtract		subtract	subtract		math.subtract	Sub
179	numpy.sum	sum	sum	sum	sum	sum	math.reduce_sum	Sum
180	numpy.take	take	take	take		take
181	numpy.tan	tan	tan	tan	tan	tan	math.tan	Tan
182	numpy.tanh	tanh	tanh	tanh	tanh	tanh	math.tanh	Tanh
183	numpy.tensordot	tensordot	tensordot	tensordot	tensordot	tensordot	tensordot
184	numpy.tile	tile	tile	tile	tile		tile	Tile
185	numpy.trace	trace	trace	trace	trace	trace	linalg.trace
186	numpy.transpose	transpose	transpose	transpose	transpose	transpose	transpose	Transpose
187	numpy.tril	tril	tril	tril	tril	tril
188	numpy.triu	triu	triu	triu	triu	triu
189	numpy.true_divide	true_divide		true_divide	true_divide	true_divide	math.truediv
190	numpy.trunc	trunc	trunc	trunc	trunc	trunc	truncatediv
191	numpy.unique	unique	unique		unique	unique	unique	Unique
192	numpy.var	var	var	var	var	var	math.reduce_variance
193	numpy.vdot	vdot	vdot	vdot	vdot
194	numpy.vstack	vstack	vstack	vstack	vstack			ConcatFromSequence
195	numpy.where	where	where	where	where	where	where	Where
196	numpy.zeros	zeros	zeros	zeros	zeros	zeros	zeros
197	numpy.zeros_like	zeros_like	zeros_like	zeros_like	zeros_like	zeros_like	zeros_like

[rgommers]

Thanks @szha! Question about scope: is ONNX in principle interested to add all the listed functions/operators, or are there sets of functionality that are out of scope?

Scrolling through the list one possible correction: ONNX's Identity probably maps to eye.

[szha]

@rgommers good catch. thanks. In terms of scope, the ONNX steering committee and operator SIG lead are generally in favor of supporting the operators in the array library standard first. We haven't committed to implementing all the above operators yet, and the community will need to review the draft standard first to decide.