operators.md

Interacting with machin learning models

Operators are one of the fundamental buidling blocks to interact with machine learning models on the Cortex network. A operator is a set of statements that performs a task or calculates a value. To use a function, you must define it somewhere in the scope from which you wish to call it.

CVM Operators

nn

OP	Name	Requirement	Default
Conv2D	channels
	kernel_size	2-D
	strides	2-D	(1, 1)
	padding	2-D	(0, 0)
	dilation	2-D	(1, 1)
	groups	1 or in_channels	1
	layout	NCHW	NCHW
	kernel_layout	OIHW	OIHW
	out_layout	undef or NCHW	undef
	out_dtype	-1 or 4(kInt32)	-1
	use_bias		TRUE
	input data	<= INT8
Dense	units	>=1
	input data	<= INT8
Upsampling	use_bias		TRUE
	scale	>0
	layout	NCHW	NCHW
	method	NEAREST_NEIGHBOR	NEAREST_NEIGHBOR
MaxPool2D	pool_size	2-D
	strides	2-D	(1, 1)
	padding	1-D or 2-D	(0, 0)
	layout	NCHW	NCHW
	ceil_mode	FALSE	FALSE

reduce

OP	Name	Requirement	Default
sum,max	axis	-ndim <= axis < ndim	()
	keepdims		FALSE
	exclude		FALSE
	dtype	kInt32	kInt32

transform

OP	Name	Requirement	Default
expand_dims	axis	-ndim-1 <= axis <= ndim
	num_newaxis	>=0	1
transpose	axes	NONE or =ndim	NONE
reshape	shape
squeeze	axis	-ndim <= axis < ndim	NONE
concatenate	axis	-ndim <= axis < ndim	1
take	axis	-ndim <= axis < ndim	NONE
strided_slice	begin
	end
	stride
repeat	repeats	>=1
	axis	-ndim <= axis < ndim	0
tile	reps		NONE
slice_like	axis	-src_ndim <= axis <= dest_ndim	NONE
cvm_lut
flatten

vision

OP	Name	Requirement	Default
get_valid_count	score_threshold		0
non_max_suppression	return_indices	FALSE	FALSE
	iou_threshold	multiply 100 by default	50
	force_suppress		FALSE
	top_k		-1
	id_index		0
	coord_start		2
	score_index		1
	max_output_size		-1
	invalid_to_bottom	TRUE	TRUE

broadcast

OP	Name	Requirement	Default
broadcast_add
broadcast_sub
broadcast_mul
broadcast_max

elemwise

OP	Name	Requirement	Default
abs
log2
elemwise_add
elemwise_sub
negative
clip
cvm_clip
cvm_right_shift
cvm_left_shift

CVM Operator Attributes

nn/convolution.cc

Name	Precision Check	Attribute Check
conv2d	✓	✓

nn/nn.cc

Name	Precision Check	Attribute Check
conv2d	✓	✓
dense	✓	✓
relu	✓	✓

nn/pooling.cc

Name	Precision Check	Attribute Check
max_pool2d	✓	✓

nn/upsampling.cc

Name	Precision Check	Attribute Check
upsampling	✓	✓

nn/nms.cc

Name	Precision Check	Attribute Check
non_max_suppression	✓	✓
get_valid_counts	✓	✓

tensor/broadcast.cc

Name	Precision Check	Attribute Check
broadcast_add	✓	✓
broadcast_sub	✓	✓
broadcast_mul	✓	✓
broadcast_max	✓	✓

tensor/reduce.cc

Name	Precision Check	Attribute Check
sum	✓	✓
max	✓	✓

tensor/elemwise.cc

Name	Precision Check	Attribute Check
abs	✓	✓
log2	✓	✓
elemwise_add	✓	✓
elemwise_sub	✓	✓
negative	✓	✓
clip	✓	✓
cvm_clip	✓	✓
cvm_right_shift	✓	✓
cvm_left_shift	✓	✓

tensor/transform.cc

Name	Precision Check	Attribute Check
repeat	✓	✓
tile	✓	✓
flatten	✓	✓
concatenate	✓	✓
expand_dims	✓	✓
reshape	✓	✓
squeeze	✓	✓
transpose	✓	✓
slice|strided_slice	✓	✓
take	✓	✓
cvm_lut	✓	✓
slice_like	✓	✓