id	sid	tid	token	lemma	pos
fcis-27197	1	1	frontiers	frontier	NOUN
fcis-27197	1	2	in	in	ADP
fcis-27197	1	3	computing	computing	NOUN
fcis-27197	1	4	and	and	CCONJ
fcis-27197	1	5	intelligent	intelligent	ADJ
fcis-27197	1	6	systems	system	NOUN
fcis-27197	1	7	issn	issn	VERB
fcis-27197	1	8	:	:	PUNCT
fcis-27197	1	9	2832	2832	NUM
fcis-27197	1	10	-	-	SYM
fcis-27197	1	11	6024	6024	NUM
fcis-27197	1	12	|	|	NOUN
fcis-27197	1	13	vol	vol	NOUN
fcis-27197	1	14	.	.	PROPN
fcis-27197	2	1	10	10	NUM
fcis-27197	2	2	,	,	PUNCT
fcis-27197	2	3	no	no	INTJ
fcis-27197	2	4	.	.	NOUN
fcis-27197	2	5	2	2	NUM
fcis-27197	2	6	,	,	PUNCT
fcis-27197	2	7	2024	2024	NUM
fcis-27197	2	8	39	39	NUM
fcis-27197	2	9	improving	improve	VERB
fcis-27197	2	10	the	the	DET
fcis-27197	2	11	performance	performance	NOUN
fcis-27197	2	12	and	and	CCONJ
fcis-27197	2	13	efficiency	efficiency	NOUN
fcis-27197	2	14	of	of	ADP
fcis-27197	2	15	convolutional	convolutional	ADJ
fcis-27197	2	16	neural	neural	ADJ
fcis-27197	2	17	networks	network	NOUN
fcis-27197	2	18	(	(	PUNCT
fcis-27197	2	19	cnns	cnns	PROPN
fcis-27197	2	20	)	)	PUNCT
fcis-27197	2	21	on	on	ADP
fcis-27197	2	22	image	image	NOUN
fcis-27197	2	23	classification	classification	NOUN
fcis-27197	2	24	tasks	task	NOUN
fcis-27197	2	25	via	via	ADP
fcis-27197	2	26	fixed‐point	fixed‐point	NOUN
fcis-27197	2	27	quantization	quantization	NOUN
fcis-27197	2	28	and	and	CCONJ
fcis-27197	2	29	structured	structure	VERB
fcis-27197	2	30	pruning	pruning	NOUN
fcis-27197	2	31	yingjie	yingjie	NOUN
fcis-27197	2	32	song	song	NOUN
fcis-27197	2	33	*	*	PUNCT
fcis-27197	2	34	lappeenranta	lappeenranta	PROPN
fcis-27197	2	35	university	university	NOUN
fcis-27197	2	36	of	of	ADP
fcis-27197	2	37	technology	technology	NOUN
fcis-27197	2	38	,	,	PUNCT
fcis-27197	2	39	lahti	lahti	NOUN
fcis-27197	2	40	,	,	PUNCT
fcis-27197	2	41	53850	53850	NUM
fcis-27197	2	42	,	,	PUNCT
fcis-27197	2	43	finland	finland	PROPN
fcis-27197	2	44	*	*	PUNCT
fcis-27197	2	45	corresponding	correspond	VERB
fcis-27197	2	46	author	author	NOUN
fcis-27197	2	47	email	email	NOUN
fcis-27197	2	48	:	:	PUNCT
fcis-27197	2	49	yingjie.song@student.lut.fi	yingjie.song@student.lut.fi	ADJ
fcis-27197	2	50	abstract	abstract	NOUN
fcis-27197	2	51	:	:	PUNCT
fcis-27197	2	52	this	this	DET
fcis-27197	2	53	study	study	NOUN
fcis-27197	2	54	explores	explore	VERB
fcis-27197	2	55	the	the	DET
fcis-27197	2	56	combined	combined	ADJ
fcis-27197	2	57	use	use	NOUN
fcis-27197	2	58	of	of	ADP
fcis-27197	2	59	fixed	fix	VERB
fcis-27197	2	60	-	-	PUNCT
fcis-27197	2	61	point	point	NOUN
fcis-27197	2	62	quantization	quantization	NOUN
fcis-27197	2	63	and	and	CCONJ
fcis-27197	2	64	structured	structure	VERB
fcis-27197	2	65	pruning	prune	VERB
fcis-27197	2	66	to	to	PART
fcis-27197	2	67	optimize	optimize	VERB
fcis-27197	2	68	the	the	DET
fcis-27197	2	69	performance	performance	NOUN
fcis-27197	2	70	and	and	CCONJ
fcis-27197	2	71	efficiency	efficiency	NOUN
fcis-27197	2	72	of	of	ADP
fcis-27197	2	73	convolutional	convolutional	ADJ
fcis-27197	2	74	neural	neural	ADJ
fcis-27197	2	75	networks	network	NOUN
fcis-27197	2	76	(	(	PUNCT
fcis-27197	2	77	cnns	cnns	PROPN
fcis-27197	2	78	)	)	PUNCT
fcis-27197	2	79	in	in	ADP
fcis-27197	2	80	image	image	NOUN
fcis-27197	2	81	classification	classification	NOUN
fcis-27197	2	82	tasks	task	NOUN
fcis-27197	2	83	.	.	PUNCT
fcis-27197	3	1	these	these	DET
fcis-27197	3	2	techniques	technique	NOUN
fcis-27197	3	3	can	can	AUX
fcis-27197	3	4	be	be	AUX
fcis-27197	3	5	used	use	VERB
fcis-27197	3	6	to	to	PART
fcis-27197	3	7	reduce	reduce	VERB
fcis-27197	3	8	model	model	NOUN
fcis-27197	3	9	size	size	NOUN
fcis-27197	3	10	and	and	CCONJ
fcis-27197	3	11	computational	computational	ADJ
fcis-27197	3	12	complexity	complexity	NOUN
fcis-27197	3	13	,	,	PUNCT
fcis-27197	3	14	making	make	VERB
fcis-27197	3	15	cnns	cnn	NOUN
fcis-27197	3	16	more	more	ADV
fcis-27197	3	17	suitable	suitable	ADJ
fcis-27197	3	18	for	for	ADP
fcis-27197	3	19	deployment	deployment	NOUN
fcis-27197	3	20	in	in	ADP
fcis-27197	3	21	resource	resource	NOUN
fcis-27197	3	22	-	-	PUNCT
fcis-27197	3	23	constrained	constrain	VERB
fcis-27197	3	24	environments	environment	NOUN
fcis-27197	3	25	such	such	ADJ
fcis-27197	3	26	as	as	ADP
fcis-27197	3	27	mobile	mobile	ADJ
fcis-27197	3	28	devices	device	NOUN
fcis-27197	3	29	and	and	CCONJ
fcis-27197	3	30	embedded	embed	VERB
fcis-27197	3	31	systems	system	NOUN
fcis-27197	3	32	.	.	PUNCT
fcis-27197	4	1	fixed	fix	VERB
fcis-27197	4	2	-	-	PUNCT
fcis-27197	4	3	point	point	NOUN
fcis-27197	4	4	quantization	quantization	NOUN
fcis-27197	4	5	methods	method	NOUN
fcis-27197	4	6	can	can	AUX
fcis-27197	4	7	reduce	reduce	VERB
fcis-27197	4	8	the	the	DET
fcis-27197	4	9	bit	bit	NOUN
fcis-27197	4	10	-	-	PUNCT
fcis-27197	4	11	width	width	NOUN
fcis-27197	4	12	of	of	ADP
fcis-27197	4	13	weights	weight	NOUN
fcis-27197	4	14	and	and	CCONJ
fcis-27197	4	15	activations	activation	NOUN
fcis-27197	4	16	,	,	PUNCT
fcis-27197	4	17	thereby	thereby	ADV
fcis-27197	4	18	reducing	reduce	VERB
fcis-27197	4	19	the	the	DET
fcis-27197	4	20	computational	computational	ADJ
fcis-27197	4	21	load	load	NOUN
fcis-27197	4	22	and	and	CCONJ
fcis-27197	4	23	memory	memory	NOUN
fcis-27197	4	24	footprint	footprint	NOUN
fcis-27197	4	25	.	.	PUNCT
fcis-27197	5	1	on	on	ADP
fcis-27197	5	2	the	the	DET
fcis-27197	5	3	other	other	ADJ
fcis-27197	5	4	hand	hand	NOUN
fcis-27197	5	5	,	,	PUNCT
fcis-27197	5	6	structured	structure	VERB
fcis-27197	5	7	pruning	pruning	NOUN
fcis-27197	5	8	systematically	systematically	ADV
fcis-27197	5	9	removes	remove	VERB
fcis-27197	5	10	unimportant	unimportant	ADJ
fcis-27197	5	11	convolutional	convolutional	ADJ
fcis-27197	5	12	filters	filter	NOUN
fcis-27197	5	13	or	or	CCONJ
fcis-27197	5	14	channels	channel	NOUN
fcis-27197	5	15	,	,	PUNCT
fcis-27197	5	16	which	which	PRON
fcis-27197	5	17	further	far	ADV
fcis-27197	5	18	reduces	reduce	VERB
fcis-27197	5	19	the	the	DET
fcis-27197	5	20	model	model	NOUN
fcis-27197	5	21	size	size	NOUN
fcis-27197	5	22	and	and	CCONJ
fcis-27197	5	23	increases	increase	VERB
fcis-27197	5	24	the	the	DET
fcis-27197	5	25	inference	inference	NOUN
fcis-27197	5	26	speed	speed	NOUN
fcis-27197	5	27	.	.	PUNCT
fcis-27197	6	1	an	an	DET
fcis-27197	6	2	experimental	experimental	ADJ
fcis-27197	6	3	evaluation	evaluation	NOUN
fcis-27197	6	4	was	be	AUX
fcis-27197	6	5	performed	perform	VERB
fcis-27197	6	6	on	on	ADP
fcis-27197	6	7	the	the	DET
fcis-27197	6	8	imagenet	imagenet	NOUN
fcis-27197	6	9	dataset	dataset	NOUN
fcis-27197	6	10	using	use	VERB
fcis-27197	6	11	the	the	DET
fcis-27197	6	12	resnet-50	resnet-50	ADJ
fcis-27197	6	13	architecture	architecture	NOUN
fcis-27197	6	14	.	.	PUNCT
fcis-27197	7	1	the	the	DET
fcis-27197	7	2	results	result	NOUN
fcis-27197	7	3	show	show	VERB
fcis-27197	7	4	that	that	SCONJ
fcis-27197	7	5	the	the	DET
fcis-27197	7	6	combined	combined	ADJ
fcis-27197	7	7	strategy	strategy	NOUN
fcis-27197	7	8	of	of	ADP
fcis-27197	7	9	quantization	quantization	NOUN
fcis-27197	7	10	and	and	CCONJ
fcis-27197	7	11	pruning	pruning	NOUN
fcis-27197	7	12	reduces	reduce	VERB
fcis-27197	7	13	the	the	DET
fcis-27197	7	14	model	model	NOUN
fcis-27197	7	15	size	size	NOUN
fcis-27197	7	16	by	by	ADP
fcis-27197	7	17	up	up	ADP
fcis-27197	7	18	to	to	PART
fcis-27197	7	19	75	75	NUM
fcis-27197	7	20	%	%	NOUN
fcis-27197	7	21	and	and	CCONJ
fcis-27197	7	22	increases	increase	VERB
fcis-27197	7	23	the	the	DET
fcis-27197	7	24	inference	inference	NOUN
fcis-27197	7	25	speed	speed	NOUN
fcis-27197	7	26	by	by	ADP
fcis-27197	7	27	50	50	NUM
fcis-27197	7	28	%	%	NOUN
fcis-27197	7	29	,	,	PUNCT
fcis-27197	7	30	while	while	SCONJ
fcis-27197	7	31	maintaining	maintain	VERB
fcis-27197	7	32	a	a	DET
fcis-27197	7	33	classification	classification	NOUN
fcis-27197	7	34	accuracy	accuracy	NOUN
fcis-27197	7	35	of	of	ADP
fcis-27197	7	36	74.5	74.5	NUM
fcis-27197	7	37	%	%	NOUN
fcis-27197	7	38	,	,	PUNCT
fcis-27197	7	39	compared	compare	VERB
fcis-27197	7	40	to	to	ADP
fcis-27197	7	41	76.4	76.4	NUM
fcis-27197	7	42	%	%	NOUN
fcis-27197	7	43	for	for	ADP
fcis-27197	7	44	the	the	DET
fcis-27197	7	45	baseline	baseline	PROPN
fcis-27197	7	46	model	model	NOUN
fcis-27197	7	47	.	.	PUNCT
fcis-27197	8	1	considering	consider	VERB
fcis-27197	8	2	the	the	DET
fcis-27197	8	3	significant	significant	ADJ
fcis-27197	8	4	increase	increase	NOUN
fcis-27197	8	5	in	in	ADP
fcis-27197	8	6	efficiency	efficiency	NOUN
fcis-27197	8	7	,	,	PUNCT
fcis-27197	8	8	a	a	DET
fcis-27197	8	9	slight	slight	ADJ
fcis-27197	8	10	decrease	decrease	NOUN
fcis-27197	8	11	in	in	ADP
fcis-27197	8	12	accuracy	accuracy	NOUN
fcis-27197	8	13	is	be	AUX
fcis-27197	8	14	acceptable	acceptable	ADJ
fcis-27197	8	15	.	.	PUNCT
fcis-27197	9	1	the	the	DET
fcis-27197	9	2	results	result	NOUN
fcis-27197	9	3	show	show	VERB
fcis-27197	9	4	that	that	SCONJ
fcis-27197	9	5	the	the	DET
fcis-27197	9	6	integrated	integrate	VERB
fcis-27197	9	7	approach	approach	NOUN
fcis-27197	9	8	effectively	effectively	ADV
fcis-27197	9	9	compresses	compress	VERB
fcis-27197	9	10	and	and	CCONJ
fcis-27197	9	11	accelerates	accelerate	VERB
fcis-27197	9	12	the	the	DET
fcis-27197	9	13	cnn	cnn	PROPN
fcis-27197	9	14	model	model	NOUN
fcis-27197	9	15	without	without	ADP
fcis-27197	9	16	a	a	DET
fcis-27197	9	17	significant	significant	ADJ
fcis-27197	9	18	drop	drop	NOUN
fcis-27197	9	19	in	in	ADP
fcis-27197	9	20	accuracy	accuracy	NOUN
fcis-27197	9	21	,	,	PUNCT
fcis-27197	9	22	making	make	VERB
fcis-27197	9	23	it	it	PRON
fcis-27197	9	24	ideal	ideal	ADJ
fcis-27197	9	25	for	for	ADP
fcis-27197	9	26	real	real	ADJ
fcis-27197	9	27	-	-	PUNCT
fcis-27197	9	28	time	time	NOUN
fcis-27197	9	29	applications	application	NOUN
fcis-27197	9	30	.	.	PUNCT
fcis-27197	10	1	keywords	keyword	NOUN
fcis-27197	10	2	:	:	PUNCT
fcis-27197	10	3	convolutional	convolutional	ADJ
fcis-27197	10	4	neural	neural	ADJ
fcis-27197	10	5	networks	network	NOUN
fcis-27197	10	6	;	;	PUNCT
fcis-27197	10	7	fixed	fix	VERB
fcis-27197	10	8	-	-	PUNCT
fcis-27197	10	9	point	point	NOUN
fcis-27197	10	10	quantization	quantization	NOUN
fcis-27197	10	11	;	;	PUNCT
fcis-27197	10	12	structured	structured	ADJ
fcis-27197	10	13	pruning	pruning	NOUN
fcis-27197	10	14	;	;	PUNCT
fcis-27197	10	15	model	model	NOUN
fcis-27197	10	16	compression	compression	NOUN
fcis-27197	10	17	;	;	PUNCT
fcis-27197	10	18	image	image	NOUN
fcis-27197	10	19	classification	classification	NOUN
fcis-27197	10	20	.	.	PUNCT
fcis-27197	11	1	1	1	X
fcis-27197	11	2	.	.	X
fcis-27197	11	3	introduction	introduction	NOUN
fcis-27197	11	4	convolutional	convolutional	ADJ
fcis-27197	11	5	neural	neural	ADJ
fcis-27197	11	6	networks	network	NOUN
fcis-27197	11	7	(	(	PUNCT
fcis-27197	11	8	cnns	cnns	PROPN
fcis-27197	11	9	)	)	PUNCT
fcis-27197	11	10	have	have	VERB
fcis-27197	11	11	a	a	DET
fcis-27197	11	12	wide	wide	ADJ
fcis-27197	11	13	range	range	NOUN
fcis-27197	11	14	of	of	ADP
fcis-27197	11	15	application	application	NOUN
fcis-27197	11	16	scenarios	scenario	NOUN
fcis-27197	11	17	in	in	ADP
fcis-27197	11	18	the	the	DET
fcis-27197	11	19	field	field	NOUN
fcis-27197	11	20	of	of	ADP
fcis-27197	11	21	computer	computer	NOUN
fcis-27197	11	22	vision	vision	NOUN
fcis-27197	11	23	.	.	PUNCT
fcis-27197	12	1	for	for	ADP
fcis-27197	12	2	example	example	NOUN
fcis-27197	12	3	,	,	PUNCT
fcis-27197	12	4	image	image	NOUN
fcis-27197	12	5	classification	classification	NOUN
fcis-27197	12	6	,	,	PUNCT
fcis-27197	12	7	target	target	NOUN
fcis-27197	12	8	detection	detection	NOUN
fcis-27197	12	9	,	,	PUNCT
fcis-27197	12	10	face	face	NOUN
fcis-27197	12	11	recognition	recognition	NOUN
fcis-27197	12	12	,	,	PUNCT
fcis-27197	12	13	image	image	NOUN
fcis-27197	12	14	generation	generation	NOUN
fcis-27197	12	15	,	,	PUNCT
fcis-27197	12	16	etc	etc	X
fcis-27197	12	17	.	.	X
fcis-27197	12	18	cnns	cnns	PROPN
fcis-27197	12	19	can	can	AUX
fcis-27197	12	20	extract	extract	VERB
fcis-27197	12	21	multilayered	multilayered	ADJ
fcis-27197	12	22	features	feature	NOUN
fcis-27197	12	23	of	of	ADP
fcis-27197	12	24	an	an	DET
fcis-27197	12	25	image	image	NOUN
fcis-27197	12	26	through	through	ADP
fcis-27197	12	27	layers	layer	NOUN
fcis-27197	12	28	of	of	ADP
fcis-27197	12	29	stacked	stack	VERB
fcis-27197	12	30	convolutional	convolutional	ADJ
fcis-27197	12	31	kernels	kernel	NOUN
fcis-27197	12	32	,	,	PUNCT
fcis-27197	12	33	thus	thus	ADV
fcis-27197	12	34	realizing	realize	VERB
fcis-27197	12	35	highly	highly	ADV
fcis-27197	12	36	complex	complex	ADJ
fcis-27197	12	37	vision	vision	NOUN
fcis-27197	12	38	tasks	task	NOUN
fcis-27197	12	39	.	.	PUNCT
fcis-27197	13	1	although	although	SCONJ
fcis-27197	13	2	cnns	cnn	NOUN
fcis-27197	13	3	are	be	AUX
fcis-27197	13	4	excellent	excellent	ADJ
fcis-27197	13	5	in	in	ADP
fcis-27197	13	6	performance	performance	NOUN
fcis-27197	13	7	,	,	PUNCT
fcis-27197	13	8	the	the	DET
fcis-27197	13	9	high	high	ADJ
fcis-27197	13	10	demand	demand	NOUN
fcis-27197	13	11	of	of	ADP
fcis-27197	13	12	computational	computational	ADJ
fcis-27197	13	13	and	and	CCONJ
fcis-27197	13	14	storage	storage	NOUN
fcis-27197	13	15	resources	resource	NOUN
fcis-27197	13	16	has	have	AUX
fcis-27197	13	17	become	become	VERB
fcis-27197	13	18	an	an	DET
fcis-27197	13	19	urgent	urgent	ADJ
fcis-27197	13	20	problem	problem	NOUN
fcis-27197	13	21	in	in	ADP
fcis-27197	13	22	applications	application	NOUN
fcis-27197	13	23	.	.	PUNCT
fcis-27197	14	1	facing	face	VERB
fcis-27197	14	2	this	this	DET
fcis-27197	14	3	challenge	challenge	NOUN
fcis-27197	14	4	,	,	PUNCT
fcis-27197	14	5	how	how	SCONJ
fcis-27197	14	6	to	to	PART
fcis-27197	14	7	compress	compress	VERB
fcis-27197	14	8	the	the	DET
fcis-27197	14	9	model	model	NOUN
fcis-27197	14	10	parameter	parameter	NOUN
fcis-27197	14	11	size	size	NOUN
fcis-27197	14	12	and	and	CCONJ
fcis-27197	14	13	reduce	reduce	VERB
fcis-27197	14	14	the	the	DET
fcis-27197	14	15	computational	computational	ADJ
fcis-27197	14	16	complexity	complexity	NOUN
fcis-27197	14	17	without	without	ADP
fcis-27197	14	18	significantly	significantly	ADV
fcis-27197	14	19	degrading	degrade	VERB
fcis-27197	14	20	the	the	DET
fcis-27197	14	21	model	model	NOUN
fcis-27197	14	22	performance	performance	NOUN
fcis-27197	14	23	has	have	AUX
fcis-27197	14	24	become	become	VERB
fcis-27197	14	25	an	an	DET
fcis-27197	14	26	important	important	ADJ
fcis-27197	14	27	research	research	NOUN
fcis-27197	14	28	direction	direction	NOUN
fcis-27197	14	29	.	.	PUNCT
fcis-27197	15	1	to	to	PART
fcis-27197	15	2	solve	solve	VERB
fcis-27197	15	3	this	this	DET
fcis-27197	15	4	problem	problem	NOUN
fcis-27197	15	5	,	,	PUNCT
fcis-27197	15	6	researchers	researcher	NOUN
fcis-27197	15	7	have	have	AUX
fcis-27197	15	8	proposed	propose	VERB
fcis-27197	15	9	a	a	DET
fcis-27197	15	10	variety	variety	NOUN
fcis-27197	15	11	of	of	ADP
fcis-27197	15	12	model	model	NOUN
fcis-27197	15	13	optimization	optimization	NOUN
fcis-27197	15	14	techniques	technique	NOUN
fcis-27197	15	15	,	,	PUNCT
fcis-27197	15	16	among	among	ADP
fcis-27197	15	17	which	which	DET
fcis-27197	15	18	quantization	quantization	NOUN
fcis-27197	15	19	and	and	CCONJ
fcis-27197	15	20	pruning	pruning	NOUN
fcis-27197	15	21	are	be	AUX
fcis-27197	15	22	the	the	DET
fcis-27197	15	23	two	two	NUM
fcis-27197	15	24	most	most	ADV
fcis-27197	15	25	used	use	VERB
fcis-27197	15	26	methods	method	NOUN
fcis-27197	15	27	.	.	PUNCT
fcis-27197	16	1	the	the	DET
fcis-27197	16	2	objective	objective	NOUN
fcis-27197	16	3	of	of	ADP
fcis-27197	16	4	this	this	DET
fcis-27197	16	5	study	study	NOUN
fcis-27197	16	6	is	be	AUX
fcis-27197	16	7	to	to	PART
fcis-27197	16	8	combine	combine	VERB
fcis-27197	16	9	fixed	fix	VERB
fcis-27197	16	10	-	-	PUNCT
fcis-27197	16	11	point	point	NOUN
fcis-27197	16	12	quantization	quantization	NOUN
fcis-27197	16	13	and	and	CCONJ
fcis-27197	16	14	structured	structure	VERB
fcis-27197	16	15	pruning	pruning	NOUN
fcis-27197	16	16	techniques	technique	NOUN
fcis-27197	16	17	to	to	PART
fcis-27197	16	18	design	design	VERB
fcis-27197	16	19	a	a	DET
fcis-27197	16	20	cnn	cnn	PROPN
fcis-27197	16	21	optimization	optimization	NOUN
fcis-27197	16	22	strategy	strategy	NOUN
fcis-27197	16	23	that	that	PRON
fcis-27197	16	24	can	can	AUX
fcis-27197	16	25	operate	operate	VERB
fcis-27197	16	26	efficiently	efficiently	ADV
fcis-27197	16	27	in	in	ADP
fcis-27197	16	28	resourceconstrained	resourceconstraine	VERB
fcis-27197	16	29	environments	environment	NOUN
fcis-27197	16	30	.	.	PUNCT
fcis-27197	17	1	2	2	X
fcis-27197	17	2	.	.	X
fcis-27197	17	3	quantitative	quantitative	ADJ
fcis-27197	17	4	techniques	technique	NOUN
fcis-27197	17	5	fixed	fix	VERB
fcis-27197	17	6	-	-	PUNCT
fcis-27197	17	7	point	point	NOUN
fcis-27197	17	8	quantization	quantization	NOUN
fcis-27197	17	9	converts	convert	VERB
fcis-27197	17	10	the	the	DET
fcis-27197	17	11	floating	float	VERB
fcis-27197	17	12	-	-	PUNCT
fcis-27197	17	13	point	point	NOUN
fcis-27197	17	14	weights	weight	NOUN
fcis-27197	17	15	and	and	CCONJ
fcis-27197	17	16	activation	activation	NOUN
fcis-27197	17	17	values	value	NOUN
fcis-27197	17	18	of	of	ADP
fcis-27197	17	19	a	a	DET
fcis-27197	17	20	model	model	NOUN
fcis-27197	17	21	from	from	ADP
fcis-27197	17	22	32	32	NUM
fcis-27197	17	23	-	-	PUNCT
fcis-27197	17	24	bit	bit	NOUN
fcis-27197	17	25	precision	precision	NOUN
fcis-27197	17	26	to	to	ADP
fcis-27197	17	27	8	8	NUM
fcis-27197	17	28	-	-	PUNCT
fcis-27197	17	29	bit	bit	NOUN
fcis-27197	17	30	or	or	CCONJ
fcis-27197	17	31	even	even	ADV
fcis-27197	17	32	fixed	fix	VERB
fcis-27197	17	33	-	-	PUNCT
fcis-27197	17	34	point	point	NOUN
fcis-27197	17	35	integral	integral	ADJ
fcis-27197	17	36	representations	representation	NOUN
fcis-27197	17	37	.	.	PUNCT
fcis-27197	18	1	this	this	DET
fcis-27197	18	2	quantization	quantization	NOUN
fcis-27197	18	3	technique	technique	NOUN
fcis-27197	18	4	can	can	AUX
fcis-27197	18	5	significantly	significantly	ADV
fcis-27197	18	6	reduce	reduce	VERB
fcis-27197	18	7	the	the	DET
fcis-27197	18	8	storage	storage	NOUN
fcis-27197	18	9	requirements	requirement	NOUN
fcis-27197	18	10	and	and	CCONJ
fcis-27197	18	11	computation	computation	NOUN
fcis-27197	18	12	of	of	ADP
fcis-27197	18	13	the	the	DET
fcis-27197	18	14	model	model	NOUN
fcis-27197	18	15	,	,	PUNCT
fcis-27197	18	16	thus	thus	ADV
fcis-27197	18	17	improving	improve	VERB
fcis-27197	18	18	inference	inference	NOUN
fcis-27197	18	19	speed	speed	NOUN
fcis-27197	18	20	[	[	X
fcis-27197	18	21	1	1	NUM
fcis-27197	18	22	]	]	PUNCT
fcis-27197	18	23	.	.	PUNCT
fcis-27197	19	1	this	this	DET
fcis-27197	19	2	technique	technique	NOUN
fcis-27197	19	3	is	be	AUX
fcis-27197	19	4	particularly	particularly	ADV
fcis-27197	19	5	suitable	suitable	ADJ
fcis-27197	19	6	for	for	ADP
fcis-27197	19	7	resource	resource	NOUN
fcis-27197	19	8	-	-	PUNCT
fcis-27197	19	9	constrained	constrain	VERB
fcis-27197	19	10	environments	environment	NOUN
fcis-27197	19	11	such	such	ADJ
fcis-27197	19	12	as	as	ADP
fcis-27197	19	13	mobile	mobile	ADJ
fcis-27197	19	14	devices	device	NOUN
fcis-27197	19	15	,	,	PUNCT
fcis-27197	19	16	and	and	CCONJ
fcis-27197	19	17	in	in	ADP
fcis-27197	19	18	network	network	NOUN
fcis-27197	19	19	architectures	architecture	NOUN
fcis-27197	19	20	such	such	ADJ
fcis-27197	19	21	as	as	ADP
fcis-27197	19	22	google	google	PROPN
fcis-27197	19	23	's	's	PART
fcis-27197	19	24	mobile	mobile	ADJ
fcis-27197	19	25	net	net	NOUN
fcis-27197	19	26	,	,	PUNCT
fcis-27197	19	27	especially	especially	ADV
fcis-27197	19	28	in	in	ADP
fcis-27197	19	29	resource	resource	NOUN
fcis-27197	19	30	-	-	PUNCT
fcis-27197	19	31	constrained	constrain	VERB
fcis-27197	19	32	environments	environment	NOUN
fcis-27197	19	33	such	such	ADJ
fcis-27197	19	34	as	as	ADP
fcis-27197	19	35	mobile	mobile	ADJ
fcis-27197	19	36	devices	device	NOUN
fcis-27197	19	37	and	and	CCONJ
fcis-27197	19	38	embedded	embed	VERB
fcis-27197	19	39	systems	system	NOUN
fcis-27197	19	40	,	,	PUNCT
fcis-27197	19	41	models	model	NOUN
fcis-27197	19	42	with	with	ADP
fcis-27197	19	43	fixed	fix	VERB
fcis-27197	19	44	-	-	PUNCT
fcis-27197	19	45	point	point	NOUN
fcis-27197	19	46	quantization	quantization	NOUN
fcis-27197	19	47	have	have	AUX
fcis-27197	19	48	shown	show	VERB
fcis-27197	19	49	significant	significant	ADJ
fcis-27197	19	50	improvement	improvement	NOUN
fcis-27197	19	51	in	in	ADP
fcis-27197	19	52	inference	inference	NOUN
fcis-27197	19	53	speed	speed	NOUN
fcis-27197	19	54	compared	compare	VERB
fcis-27197	19	55	to	to	ADP
fcis-27197	19	56	floating	float	VERB
fcis-27197	19	57	-	-	PUNCT
fcis-27197	19	58	point	point	NOUN
fcis-27197	19	59	models	model	NOUN
fcis-27197	19	60	[	[	X
fcis-27197	19	61	2	2	NUM
fcis-27197	19	62	]	]	PUNCT
fcis-27197	19	63	.	.	PUNCT
fcis-27197	20	1	3	3	X
fcis-27197	20	2	.	.	X
fcis-27197	20	3	pruning	prune	VERB
fcis-27197	20	4	techniques	technique	NOUN
fcis-27197	20	5	to	to	PART
fcis-27197	20	6	solve	solve	VERB
fcis-27197	20	7	the	the	DET
fcis-27197	20	8	sparsity	sparsity	NOUN
fcis-27197	20	9	problem	problem	NOUN
fcis-27197	20	10	of	of	ADP
fcis-27197	20	11	weighted	weighted	ADJ
fcis-27197	20	12	pruning	pruning	NOUN
fcis-27197	20	13	,	,	PUNCT
fcis-27197	20	14	structured	structure	VERB
fcis-27197	20	15	pruning	pruning	NOUN
fcis-27197	20	16	has	have	AUX
fcis-27197	20	17	emerged	emerge	VERB
fcis-27197	20	18	.	.	PUNCT
fcis-27197	21	1	structured	structured	ADJ
fcis-27197	21	2	pruning	pruning	NOUN
fcis-27197	21	3	compresses	compress	NOUN
fcis-27197	21	4	by	by	ADP
fcis-27197	21	5	removing	remove	VERB
fcis-27197	21	6	the	the	DET
fcis-27197	21	7	entire	entire	ADJ
fcis-27197	21	8	convolutional	convolutional	ADJ
fcis-27197	21	9	kernel	kernel	NOUN
fcis-27197	21	10	,	,	PUNCT
fcis-27197	21	11	channel	channel	NOUN
fcis-27197	21	12	,	,	PUNCT
fcis-27197	21	13	or	or	CCONJ
fcis-27197	21	14	even	even	ADV
fcis-27197	21	15	layer	layer	NOUN
fcis-27197	21	16	,	,	PUNCT
fcis-27197	21	17	which	which	PRON
fcis-27197	21	18	allows	allow	VERB
fcis-27197	21	19	the	the	DET
fcis-27197	21	20	pruned	pruned	ADJ
fcis-27197	21	21	model	model	NOUN
fcis-27197	21	22	to	to	PART
fcis-27197	21	23	still	still	ADV
fcis-27197	21	24	maintain	maintain	VERB
fcis-27197	21	25	structural	structural	ADJ
fcis-27197	21	26	integrity	integrity	NOUN
fcis-27197	21	27	and	and	CCONJ
fcis-27197	21	28	facilitates	facilitate	VERB
fcis-27197	21	29	efficient	efficient	ADJ
fcis-27197	21	30	acceleration	acceleration	NOUN
fcis-27197	21	31	on	on	ADP
fcis-27197	21	32	hardware	hardware	NOUN
fcis-27197	22	1	[	[	X
fcis-27197	22	2	3,4	3,4	NUM
fcis-27197	22	3	]	]	PUNCT
fcis-27197	22	4	.	.	PUNCT
fcis-27197	23	1	various	various	ADJ
fcis-27197	23	2	structured	structured	ADJ
fcis-27197	23	3	pruning	pruning	NOUN
fcis-27197	23	4	algorithms	algorithm	NOUN
fcis-27197	23	5	have	have	AUX
fcis-27197	23	6	been	be	AUX
fcis-27197	23	7	integrated	integrate	VERB
fcis-27197	23	8	into	into	ADP
fcis-27197	23	9	baidu	baidu	PROPN
fcis-27197	23	10	's	's	PART
fcis-27197	23	11	paddle	paddle	NOUN
fcis-27197	23	12	framework	framework	NOUN
fcis-27197	23	13	for	for	ADP
fcis-27197	23	14	optimizing	optimize	VERB
fcis-27197	23	15	the	the	DET
fcis-27197	23	16	deployment	deployment	NOUN
fcis-27197	23	17	of	of	ADP
fcis-27197	23	18	deep	deep	ADJ
fcis-27197	23	19	learning	learning	NOUN
fcis-27197	23	20	models	model	NOUN
fcis-27197	23	21	in	in	ADP
fcis-27197	23	22	production	production	NOUN
fcis-27197	23	23	environments	environment	NOUN
fcis-27197	23	24	[	[	X
fcis-27197	23	25	3	3	NUM
fcis-27197	23	26	]	]	PUNCT
fcis-27197	23	27	.	.	PUNCT
fcis-27197	24	1	the	the	DET
fcis-27197	24	2	advantage	advantage	NOUN
fcis-27197	24	3	of	of	ADP
fcis-27197	24	4	structured	structured	ADJ
fcis-27197	24	5	pruning	pruning	NOUN
fcis-27197	24	6	is	be	AUX
fcis-27197	24	7	that	that	SCONJ
fcis-27197	24	8	it	it	PRON
fcis-27197	24	9	not	not	PART
fcis-27197	24	10	only	only	ADV
fcis-27197	24	11	reduces	reduce	VERB
fcis-27197	24	12	the	the	DET
fcis-27197	24	13	number	number	NOUN
fcis-27197	24	14	of	of	ADP
fcis-27197	24	15	parameters	parameter	NOUN
fcis-27197	24	16	in	in	ADP
fcis-27197	24	17	the	the	DET
fcis-27197	24	18	model	model	NOUN
fcis-27197	24	19	,	,	PUNCT
fcis-27197	24	20	but	but	CCONJ
fcis-27197	24	21	also	also	ADV
fcis-27197	24	22	effectively	effectively	ADV
fcis-27197	24	23	improves	improve	VERB
fcis-27197	24	24	the	the	DET
fcis-27197	24	25	inference	inference	NOUN
fcis-27197	24	26	speed	speed	NOUN
fcis-27197	24	27	,	,	PUNCT
fcis-27197	24	28	especially	especially	ADV
fcis-27197	24	29	in	in	ADP
fcis-27197	24	30	embedded	embed	VERB
fcis-27197	24	31	or	or	CCONJ
fcis-27197	24	32	mobile	mobile	ADJ
fcis-27197	24	33	devices	device	NOUN
fcis-27197	24	34	.	.	PUNCT
fcis-27197	25	1	however	however	ADV
fcis-27197	25	2	,	,	PUNCT
fcis-27197	25	3	the	the	DET
fcis-27197	25	4	design	design	NOUN
fcis-27197	25	5	and	and	CCONJ
fcis-27197	25	6	implementation	implementation	NOUN
fcis-27197	25	7	of	of	ADP
fcis-27197	25	8	structured	structured	ADJ
fcis-27197	25	9	pruning	pruning	NOUN
fcis-27197	25	10	is	be	AUX
fcis-27197	25	11	more	more	ADV
fcis-27197	25	12	complex	complex	ADJ
fcis-27197	25	13	.	.	PUNCT
fcis-27197	26	1	pruning	prune	VERB
fcis-27197	26	2	needs	need	NOUN
fcis-27197	26	3	to	to	PART
fcis-27197	26	4	ensure	ensure	VERB
fcis-27197	26	5	that	that	SCONJ
fcis-27197	26	6	the	the	DET
fcis-27197	26	7	network	network	NOUN
fcis-27197	26	8	structure	structure	NOUN
fcis-27197	26	9	is	be	AUX
fcis-27197	26	10	not	not	PART
fcis-27197	26	11	damaged	damage	VERB
fcis-27197	26	12	,	,	PUNCT
fcis-27197	26	13	to	to	PART
fcis-27197	26	14	ensure	ensure	VERB
fcis-27197	26	15	that	that	SCONJ
fcis-27197	26	16	the	the	DET
fcis-27197	26	17	model	model	NOUN
fcis-27197	26	18	after	after	ADP
fcis-27197	26	19	pruning	prune	VERB
fcis-27197	26	20	still	still	ADV
fcis-27197	26	21	has	have	VERB
fcis-27197	26	22	good	good	ADJ
fcis-27197	26	23	training	training	NOUN
fcis-27197	26	24	stability	stability	NOUN
fcis-27197	26	25	and	and	CCONJ
fcis-27197	26	26	inference	inference	NOUN
fcis-27197	26	27	accuracy	accuracy	NOUN
fcis-27197	26	28	.	.	PUNCT
fcis-27197	27	1	4	4	X
fcis-27197	27	2	.	.	X
fcis-27197	27	3	design	design	NOUN
fcis-27197	27	4	and	and	CCONJ
fcis-27197	27	5	optimization	optimization	NOUN
fcis-27197	27	6	of	of	ADP
fcis-27197	27	7	quantitative	quantitative	ADJ
fcis-27197	27	8	methods	method	NOUN
fcis-27197	27	9	4.1	4.1	NUM
fcis-27197	27	10	.	.	PUNCT
fcis-27197	27	11	selection	selection	NOUN
fcis-27197	27	12	of	of	ADP
fcis-27197	27	13	quantization	quantization	NOUN
fcis-27197	27	14	bit	bit	NOUN
fcis-27197	27	15	width	width	ADJ
fcis-27197	27	16	8	8	NUM
fcis-27197	27	17	-	-	PUNCT
fcis-27197	27	18	bit	bit	NOUN
fcis-27197	27	19	is	be	AUX
fcis-27197	27	20	the	the	DET
fcis-27197	27	21	most	most	ADV
fcis-27197	27	22	common	common	ADJ
fcis-27197	27	23	choice	choice	NOUN
fcis-27197	27	24	of	of	ADP
fcis-27197	27	25	quantization	quantization	NOUN
fcis-27197	27	26	bit	bit	NOUN
fcis-27197	27	27	width	width	ADJ
fcis-27197	27	28	because	because	SCONJ
fcis-27197	27	29	it	it	PRON
fcis-27197	27	30	has	have	VERB
fcis-27197	27	31	high	high	ADJ
fcis-27197	27	32	computational	computational	ADJ
fcis-27197	27	33	efficiency	efficiency	NOUN
fcis-27197	27	34	in	in	ADP
fcis-27197	27	35	hardware	hardware	NOUN
fcis-27197	27	36	gas	gas	NOUN
fcis-27197	27	37	pedals	pedal	NOUN
fcis-27197	27	38	,	,	PUNCT
fcis-27197	27	39	while	while	SCONJ
fcis-27197	27	40	being	be	AUX
fcis-27197	27	41	able	able	ADJ
fcis-27197	27	42	to	to	PART
fcis-27197	27	43	guarantee	guarantee	VERB
fcis-27197	27	44	the	the	DET
fcis-27197	27	45	accuracy	accuracy	NOUN
fcis-27197	27	46	of	of	ADP
fcis-27197	27	47	most	most	ADJ
fcis-27197	27	48	models	model	NOUN
fcis-27197	27	49	[	[	X
fcis-27197	27	50	1	1	NUM
fcis-27197	27	51	]	]	PUNCT
fcis-27197	27	52	.	.	PUNCT
fcis-27197	28	1	however	however	ADV
fcis-27197	28	2	,	,	PUNCT
fcis-27197	28	3	in	in	ADP
fcis-27197	28	4	some	some	DET
fcis-27197	28	5	applications	application	NOUN
fcis-27197	28	6	that	that	PRON
fcis-27197	28	7	require	require	VERB
fcis-27197	28	8	less	less	ADV
fcis-27197	28	9	accuracy	accuracy	NOUN
fcis-27197	28	10	,	,	PUNCT
fcis-27197	28	11	lower	low	ADJ
fcis-27197	28	12	bit	bit	NOUN
fcis-27197	28	13	widths	width	NOUN
fcis-27197	28	14	can	can	AUX
fcis-27197	28	15	be	be	AUX
fcis-27197	28	16	chosen	choose	VERB
fcis-27197	28	17	to	to	PART
fcis-27197	28	18	further	far	ADV
fcis-27197	28	19	reduce	reduce	VERB
fcis-27197	28	20	the	the	DET
fcis-27197	28	21	40	40	NUM
fcis-27197	28	22	computational	computational	ADJ
fcis-27197	28	23	effort	effort	NOUN
fcis-27197	29	1	[	[	X
fcis-27197	29	2	4	4	NUM
fcis-27197	29	3	]	]	PUNCT
fcis-27197	29	4	.	.	PUNCT
fcis-27197	30	1	too	too	ADV
fcis-27197	30	2	low	low	ADJ
fcis-27197	30	3	a	a	DET
fcis-27197	30	4	bit	bit	NOUN
fcis-27197	30	5	width	width	NOUN
fcis-27197	30	6	can	can	AUX
fcis-27197	30	7	lead	lead	VERB
fcis-27197	30	8	to	to	ADP
fcis-27197	30	9	insufficient	insufficient	ADJ
fcis-27197	30	10	representation	representation	NOUN
fcis-27197	30	11	capability	capability	NOUN
fcis-27197	30	12	of	of	ADP
fcis-27197	30	13	the	the	DET
fcis-27197	30	14	model	model	NOUN
fcis-27197	30	15	,	,	PUNCT
fcis-27197	30	16	thus	thus	ADV
fcis-27197	30	17	affecting	affect	VERB
fcis-27197	30	18	the	the	DET
fcis-27197	30	19	classification	classification	NOUN
fcis-27197	30	20	accuracy	accuracy	NOUN
fcis-27197	30	21	.	.	PUNCT
fcis-27197	31	1	too	too	ADV
fcis-27197	31	2	high	high	ADJ
fcis-27197	31	3	a	a	DET
fcis-27197	31	4	bit	bit	NOUN
fcis-27197	31	5	width	width	NOUN
fcis-27197	31	6	can	can	AUX
fcis-27197	31	7	reduce	reduce	VERB
fcis-27197	31	8	the	the	DET
fcis-27197	31	9	accuracy	accuracy	NOUN
fcis-27197	31	10	loss	loss	NOUN
fcis-27197	31	11	,	,	PUNCT
fcis-27197	31	12	but	but	CCONJ
fcis-27197	31	13	it	it	PRON
fcis-27197	31	14	is	be	AUX
fcis-27197	31	15	difficult	difficult	ADJ
fcis-27197	31	16	to	to	PART
fcis-27197	31	17	fully	fully	ADV
fcis-27197	31	18	utilize	utilize	VERB
fcis-27197	31	19	the	the	DET
fcis-27197	31	20	advantages	advantage	NOUN
fcis-27197	31	21	of	of	ADP
fcis-27197	31	22	quantization	quantization	NOUN
fcis-27197	31	23	techniques	technique	NOUN
fcis-27197	31	24	.	.	PUNCT
fcis-27197	32	1	4.2	4.2	NUM
fcis-27197	32	2	.	.	PUNCT
fcis-27197	33	1	qat	qat	PROPN
fcis-27197	33	2	quantization	quantization	NOUN
fcis-27197	33	3	-	-	PUNCT
fcis-27197	33	4	aware	aware	ADJ
fcis-27197	33	5	training	training	NOUN
fcis-27197	33	6	(	(	PUNCT
fcis-27197	33	7	qat	qat	PROPN
fcis-27197	33	8	)	)	PUNCT
fcis-27197	33	9	is	be	AUX
fcis-27197	33	10	a	a	DET
fcis-27197	33	11	method	method	NOUN
fcis-27197	33	12	that	that	PRON
fcis-27197	33	13	simulates	simulate	VERB
fcis-27197	33	14	quantization	quantization	NOUN
fcis-27197	33	15	operations	operation	NOUN
fcis-27197	33	16	during	during	ADP
fcis-27197	33	17	training	training	NOUN
fcis-27197	33	18	.	.	PUNCT
fcis-27197	34	1	unlike	unlike	ADP
fcis-27197	34	2	direct	direct	ADJ
fcis-27197	34	3	quantization	quantization	NOUN
fcis-27197	34	4	of	of	ADP
fcis-27197	34	5	a	a	DET
fcis-27197	34	6	trained	train	VERB
fcis-27197	34	7	model	model	NOUN
fcis-27197	34	8	,	,	PUNCT
fcis-27197	34	9	qat	qat	PROPN
fcis-27197	34	10	introduces	introduce	VERB
fcis-27197	34	11	the	the	DET
fcis-27197	34	12	simulation	simulation	NOUN
fcis-27197	34	13	of	of	ADP
fcis-27197	34	14	quantization	quantization	NOUN
fcis-27197	34	15	errors	error	NOUN
fcis-27197	34	16	during	during	ADP
fcis-27197	34	17	the	the	DET
fcis-27197	34	18	training	training	NOUN
fcis-27197	34	19	phase	phase	NOUN
fcis-27197	34	20	,	,	PUNCT
fcis-27197	34	21	thus	thus	ADV
fcis-27197	34	22	allowing	allow	VERB
fcis-27197	34	23	the	the	DET
fcis-27197	34	24	model	model	NOUN
fcis-27197	34	25	to	to	PART
fcis-27197	34	26	adapt	adapt	VERB
fcis-27197	34	27	to	to	ADP
fcis-27197	34	28	changes	change	NOUN
fcis-27197	34	29	in	in	ADP
fcis-27197	34	30	accuracy	accuracy	NOUN
fcis-27197	34	31	after	after	ADP
fcis-27197	34	32	quantization	quantization	NOUN
fcis-27197	34	33	during	during	ADP
fcis-27197	34	34	training	training	NOUN
fcis-27197	34	35	[	[	X
fcis-27197	34	36	1,2	1,2	NUM
fcis-27197	34	37	]	]	PUNCT
fcis-27197	34	38	.	.	PUNCT
fcis-27197	35	1	we	we	PRON
fcis-27197	35	2	choose	choose	VERB
fcis-27197	35	3	a	a	DET
fcis-27197	35	4	layer	layer	NOUN
fcis-27197	35	5	-	-	PUNCT
fcis-27197	35	6	by	by	ADP
fcis-27197	35	7	-	-	PUNCT
fcis-27197	35	8	layer	layer	NOUN
fcis-27197	35	9	quantization	quantization	NOUN
fcis-27197	35	10	strategy	strategy	NOUN
fcis-27197	35	11	,	,	PUNCT
fcis-27197	35	12	where	where	SCONJ
fcis-27197	35	13	different	different	ADJ
fcis-27197	35	14	network	network	NOUN
fcis-27197	35	15	layers	layer	NOUN
fcis-27197	35	16	play	play	VERB
fcis-27197	35	17	different	different	ADJ
fcis-27197	35	18	roles	role	NOUN
fcis-27197	35	19	in	in	ADP
fcis-27197	35	20	the	the	DET
fcis-27197	35	21	model	model	NOUN
fcis-27197	35	22	,	,	PUNCT
fcis-27197	35	23	so	so	SCONJ
fcis-27197	35	24	that	that	SCONJ
fcis-27197	35	25	the	the	DET
fcis-27197	35	26	quantization	quantization	NOUN
fcis-27197	35	27	bit	bit	NOUN
fcis-27197	35	28	-	-	PUNCT
fcis-27197	35	29	width	width	NOUN
fcis-27197	35	30	can	can	AUX
fcis-27197	35	31	be	be	AUX
fcis-27197	35	32	chosen	choose	VERB
fcis-27197	35	33	to	to	PART
fcis-27197	35	34	be	be	AUX
fcis-27197	35	35	higher	high	ADJ
fcis-27197	35	36	at	at	ADP
fcis-27197	35	37	,	,	PUNCT
fcis-27197	35	38	for	for	ADP
fcis-27197	35	39	example	example	NOUN
fcis-27197	35	40	,	,	PUNCT
fcis-27197	35	41	8	8	NUM
fcis-27197	35	42	-	-	PUNCT
fcis-27197	35	43	bit	bit	NOUN
fcis-27197	35	44	for	for	ADP
fcis-27197	35	45	the	the	DET
fcis-27197	35	46	convolutional	convolutional	ADJ
fcis-27197	35	47	layer	layer	NOUN
fcis-27197	35	48	,	,	PUNCT
fcis-27197	35	49	and	and	CCONJ
fcis-27197	35	50	lower	low	ADJ
fcis-27197	35	51	for	for	ADP
fcis-27197	35	52	the	the	DET
fcis-27197	35	53	fully	fully	ADV
fcis-27197	35	54	connected	connect	VERB
fcis-27197	35	55	layer	layer	NOUN
fcis-27197	35	56	.	.	PUNCT
fcis-27197	36	1	this	this	PRON
fcis-27197	36	2	can	can	AUX
fcis-27197	36	3	further	far	ADV
fcis-27197	36	4	reduce	reduce	VERB
fcis-27197	36	5	the	the	DET
fcis-27197	36	6	computational	computational	ADJ
fcis-27197	36	7	effort	effort	NOUN
fcis-27197	36	8	of	of	ADP
fcis-27197	36	9	the	the	DET
fcis-27197	36	10	model	model	NOUN
fcis-27197	36	11	while	while	SCONJ
fcis-27197	36	12	minimizing	minimize	VERB
fcis-27197	36	13	the	the	DET
fcis-27197	36	14	loss	loss	NOUN
fcis-27197	36	15	of	of	ADP
fcis-27197	36	16	accuracy	accuracy	NOUN
fcis-27197	36	17	.	.	PUNCT
fcis-27197	37	1	5	5	X
fcis-27197	37	2	.	.	X
fcis-27197	37	3	optimization	optimization	NOUN
fcis-27197	37	4	strategies	strategy	NOUN
fcis-27197	37	5	for	for	ADP
fcis-27197	37	6	quantitative	quantitative	ADJ
fcis-27197	37	7	methods	method	NOUN
fcis-27197	37	8	we	we	PRON
fcis-27197	37	9	must	must	AUX
fcis-27197	37	10	design	design	VERB
fcis-27197	37	11	for	for	ADP
fcis-27197	37	12	inter	inter	ADJ
fcis-27197	37	13	-	-	ADJ
fcis-27197	37	14	layer	layer	ADJ
fcis-27197	37	15	quantization	quantization	NOUN
fcis-27197	37	16	differentiation	differentiation	NOUN
fcis-27197	37	17	.	.	PUNCT
fcis-27197	38	1	different	different	ADJ
fcis-27197	38	2	layers	layer	NOUN
fcis-27197	38	3	in	in	ADP
fcis-27197	38	4	a	a	DET
fcis-27197	38	5	cnn	cnn	PROPN
fcis-27197	38	6	have	have	VERB
fcis-27197	38	7	different	different	ADJ
fcis-27197	38	8	sensitivities	sensitivity	NOUN
fcis-27197	38	9	to	to	ADP
fcis-27197	38	10	quantization	quantization	NOUN
fcis-27197	38	11	.	.	PUNCT
fcis-27197	39	1	in	in	ADP
fcis-27197	39	2	general	general	ADJ
fcis-27197	39	3	,	,	PUNCT
fcis-27197	39	4	convolutional	convolutional	ADJ
fcis-27197	39	5	layers	layer	NOUN
fcis-27197	39	6	close	close	ADJ
fcis-27197	39	7	to	to	ADP
fcis-27197	39	8	the	the	DET
fcis-27197	39	9	input	input	NOUN
fcis-27197	39	10	layer	layer	NOUN
fcis-27197	39	11	are	be	AUX
fcis-27197	39	12	more	more	ADV
fcis-27197	39	13	sensitive	sensitive	ADJ
fcis-27197	39	14	to	to	ADP
fcis-27197	39	15	quantization	quantization	NOUN
fcis-27197	39	16	because	because	SCONJ
fcis-27197	39	17	these	these	DET
fcis-27197	39	18	layers	layer	NOUN
fcis-27197	39	19	extract	extract	VERB
fcis-27197	39	20	low	low	ADJ
fcis-27197	39	21	-	-	PUNCT
fcis-27197	39	22	level	level	NOUN
fcis-27197	39	23	image	image	NOUN
fcis-27197	39	24	features	feature	NOUN
fcis-27197	39	25	,	,	PUNCT
fcis-27197	39	26	while	while	SCONJ
fcis-27197	39	27	fully	fully	ADV
fcis-27197	39	28	connected	connected	ADJ
fcis-27197	39	29	layers	layer	NOUN
fcis-27197	39	30	close	close	ADV
fcis-27197	39	31	to	to	ADP
fcis-27197	39	32	the	the	DET
fcis-27197	39	33	output	output	NOUN
fcis-27197	39	34	layer	layer	NOUN
fcis-27197	39	35	are	be	AUX
fcis-27197	39	36	less	less	ADV
fcis-27197	39	37	sensitive	sensitive	ADJ
fcis-27197	39	38	to	to	ADP
fcis-27197	39	39	quantization	quantization	NOUN
fcis-27197	39	40	.	.	PUNCT
fcis-27197	40	1	a	a	DET
fcis-27197	40	2	balance	balance	NOUN
fcis-27197	40	3	between	between	ADP
fcis-27197	40	4	performance	performance	NOUN
fcis-27197	40	5	and	and	CCONJ
fcis-27197	40	6	accuracy	accuracy	NOUN
fcis-27197	40	7	can	can	AUX
fcis-27197	40	8	be	be	AUX
fcis-27197	40	9	achieved	achieve	VERB
fcis-27197	40	10	by	by	ADP
fcis-27197	40	11	using	use	VERB
fcis-27197	40	12	higher	high	ADJ
fcis-27197	40	13	quantization	quantization	NOUN
fcis-27197	40	14	accuracy	accuracy	NOUN
fcis-27197	40	15	in	in	ADP
fcis-27197	40	16	the	the	DET
fcis-27197	40	17	sensitive	sensitive	ADJ
fcis-27197	40	18	layers	layer	NOUN
fcis-27197	40	19	and	and	CCONJ
fcis-27197	40	20	lower	low	ADJ
fcis-27197	40	21	bit	bit	NOUN
fcis-27197	40	22	widths	width	NOUN
fcis-27197	40	23	in	in	ADP
fcis-27197	40	24	the	the	DET
fcis-27197	40	25	insensitive	insensitive	ADJ
fcis-27197	40	26	layers	layer	NOUN
fcis-27197	40	27	[	[	X
fcis-27197	40	28	1,2	1,2	NUM
fcis-27197	40	29	]	]	PUNCT
fcis-27197	40	30	.	.	PUNCT
fcis-27197	41	1	zero	zero	NUM
fcis-27197	41	2	offset	offset	VERB
fcis-27197	41	3	in	in	ADP
fcis-27197	41	4	the	the	DET
fcis-27197	41	5	quantization	quantization	NOUN
fcis-27197	41	6	process	process	NOUN
fcis-27197	41	7	is	be	AUX
fcis-27197	41	8	one	one	NUM
fcis-27197	41	9	of	of	ADP
fcis-27197	41	10	the	the	DET
fcis-27197	41	11	key	key	ADJ
fcis-27197	41	12	factors	factor	NOUN
fcis-27197	41	13	affecting	affect	VERB
fcis-27197	41	14	the	the	DET
fcis-27197	41	15	accuracy	accuracy	NOUN
fcis-27197	41	16	of	of	ADP
fcis-27197	41	17	the	the	DET
fcis-27197	41	18	model	model	NOUN
fcis-27197	41	19	.	.	PUNCT
fcis-27197	42	1	according	accord	VERB
fcis-27197	42	2	to	to	ADP
fcis-27197	42	3	the	the	DET
fcis-27197	42	4	distribution	distribution	NOUN
fcis-27197	42	5	characteristics	characteristic	NOUN
fcis-27197	42	6	of	of	ADP
fcis-27197	42	7	the	the	DET
fcis-27197	42	8	weights	weight	NOUN
fcis-27197	42	9	and	and	CCONJ
fcis-27197	42	10	activation	activation	NOUN
fcis-27197	42	11	values	value	NOUN
fcis-27197	42	12	of	of	ADP
fcis-27197	42	13	each	each	DET
fcis-27197	42	14	layer	layer	NOUN
fcis-27197	42	15	,	,	PUNCT
fcis-27197	42	16	the	the	DET
fcis-27197	42	17	position	position	NOUN
fcis-27197	42	18	of	of	ADP
fcis-27197	42	19	the	the	DET
fcis-27197	42	20	zero	zero	NUM
fcis-27197	42	21	point	point	NOUN
fcis-27197	42	22	is	be	AUX
fcis-27197	42	23	dynamically	dynamically	ADV
fcis-27197	42	24	adjusted	adjust	VERB
fcis-27197	42	25	so	so	SCONJ
fcis-27197	42	26	that	that	SCONJ
fcis-27197	42	27	the	the	DET
fcis-27197	42	28	quantized	quantize	VERB
fcis-27197	42	29	value	value	NOUN
fcis-27197	42	30	can	can	AUX
fcis-27197	42	31	better	well	ADV
fcis-27197	42	32	cover	cover	VERB
fcis-27197	42	33	the	the	DET
fcis-27197	42	34	actual	actual	ADJ
fcis-27197	42	35	data	datum	NOUN
fcis-27197	42	36	distribution	distribution	NOUN
fcis-27197	42	37	and	and	CCONJ
fcis-27197	42	38	reduce	reduce	VERB
fcis-27197	42	39	the	the	DET
fcis-27197	42	40	bias	bias	NOUN
fcis-27197	42	41	introduced	introduce	VERB
fcis-27197	42	42	by	by	ADP
fcis-27197	42	43	quantization	quantization	NOUN
fcis-27197	42	44	[	[	X
fcis-27197	42	45	1	1	NUM
fcis-27197	42	46	]	]	PUNCT
fcis-27197	42	47	.	.	PUNCT
fcis-27197	43	1	to	to	PART
fcis-27197	43	2	further	far	ADV
fcis-27197	43	3	improve	improve	VERB
fcis-27197	43	4	the	the	DET
fcis-27197	43	5	robustness	robustness	NOUN
fcis-27197	43	6	of	of	ADP
fcis-27197	43	7	the	the	DET
fcis-27197	43	8	quantization	quantization	NOUN
fcis-27197	43	9	model	model	NOUN
fcis-27197	43	10	,	,	PUNCT
fcis-27197	43	11	quantization	quantization	NOUN
fcis-27197	43	12	techniques	technique	NOUN
fcis-27197	43	13	can	can	AUX
fcis-27197	43	14	be	be	AUX
fcis-27197	43	15	used	use	VERB
fcis-27197	43	16	in	in	ADP
fcis-27197	43	17	conjunction	conjunction	NOUN
fcis-27197	43	18	with	with	ADP
fcis-27197	43	19	regularization	regularization	NOUN
fcis-27197	43	20	methods	method	NOUN
fcis-27197	43	21	.	.	PUNCT
fcis-27197	44	1	during	during	ADP
fcis-27197	44	2	the	the	DET
fcis-27197	44	3	training	training	NOUN
fcis-27197	44	4	process	process	NOUN
fcis-27197	44	5	,	,	PUNCT
fcis-27197	44	6	the	the	DET
fcis-27197	44	7	regularization	regularization	NOUN
fcis-27197	44	8	term	term	NOUN
fcis-27197	44	9	is	be	AUX
fcis-27197	44	10	used	use	VERB
fcis-27197	44	11	to	to	PART
fcis-27197	44	12	constrain	constrain	VERB
fcis-27197	44	13	the	the	DET
fcis-27197	44	14	weight	weight	NOUN
fcis-27197	44	15	distribution	distribution	NOUN
fcis-27197	44	16	after	after	ADP
fcis-27197	44	17	quantization	quantization	NOUN
fcis-27197	44	18	[	[	X
fcis-27197	44	19	1,2	1,2	NUM
fcis-27197	44	20	]	]	PUNCT
fcis-27197	44	21	,	,	PUNCT
fcis-27197	44	22	making	make	VERB
fcis-27197	44	23	the	the	DET
fcis-27197	44	24	model	model	NOUN
fcis-27197	44	25	more	more	ADV
fcis-27197	44	26	robust	robust	ADJ
fcis-27197	44	27	under	under	ADP
fcis-27197	44	28	fixed	fix	VERB
fcis-27197	44	29	-	-	PUNCT
fcis-27197	44	30	point	point	NOUN
fcis-27197	44	31	representation	representation	NOUN
fcis-27197	44	32	.	.	PUNCT
fcis-27197	45	1	this	this	DET
fcis-27197	45	2	method	method	NOUN
fcis-27197	45	3	can	can	AUX
fcis-27197	45	4	effectively	effectively	ADV
fcis-27197	45	5	reduce	reduce	VERB
fcis-27197	45	6	the	the	DET
fcis-27197	45	7	overfitting	overfitte	VERB
fcis-27197	45	8	phenomenon	phenomenon	NOUN
fcis-27197	45	9	and	and	CCONJ
fcis-27197	45	10	improve	improve	VERB
fcis-27197	45	11	the	the	DET
fcis-27197	45	12	generalization	generalization	NOUN
fcis-27197	45	13	ability	ability	NOUN
fcis-27197	45	14	of	of	ADP
fcis-27197	45	15	the	the	DET
fcis-27197	45	16	model	model	NOUN
fcis-27197	45	17	[	[	X
fcis-27197	45	18	4	4	NUM
fcis-27197	45	19	]	]	PUNCT
fcis-27197	45	20	.	.	PUNCT
fcis-27197	46	1	the	the	DET
fcis-27197	46	2	imagenet	imagenet	NOUN
fcis-27197	46	3	dataset	dataset	NOUN
fcis-27197	46	4	is	be	AUX
fcis-27197	46	5	selected	select	VERB
fcis-27197	46	6	for	for	ADP
fcis-27197	46	7	experiments	experiment	NOUN
fcis-27197	46	8	to	to	PART
fcis-27197	46	9	evaluate	evaluate	VERB
fcis-27197	46	10	the	the	DET
fcis-27197	46	11	effect	effect	NOUN
fcis-27197	46	12	of	of	ADP
fcis-27197	46	13	quantization	quantization	NOUN
fcis-27197	46	14	methods	method	NOUN
fcis-27197	46	15	on	on	ADP
fcis-27197	46	16	model	model	NOUN
fcis-27197	46	17	performance	performance	NOUN
fcis-27197	46	18	.	.	PUNCT
fcis-27197	47	1	the	the	DET
fcis-27197	47	2	cnn	cnn	PROPN
fcis-27197	47	3	architecture	architecture	NOUN
fcis-27197	47	4	is	be	AUX
fcis-27197	47	5	chosen	choose	VERB
fcis-27197	47	6	as	as	ADP
fcis-27197	47	7	the	the	DET
fcis-27197	47	8	experimental	experimental	ADJ
fcis-27197	47	9	object	object	NOUN
fcis-27197	47	10	to	to	PART
fcis-27197	47	11	analyze	analyze	VERB
fcis-27197	47	12	the	the	DET
fcis-27197	47	13	effect	effect	NOUN
fcis-27197	47	14	of	of	ADP
fcis-27197	47	15	quantization	quantization	NOUN
fcis-27197	47	16	methods	method	NOUN
fcis-27197	47	17	by	by	ADP
fcis-27197	47	18	comparing	compare	VERB
fcis-27197	47	19	the	the	DET
fcis-27197	47	20	model	model	NOUN
fcis-27197	47	21	performance	performance	NOUN
fcis-27197	47	22	before	before	ADP
fcis-27197	47	23	and	and	CCONJ
fcis-27197	47	24	after	after	ADP
fcis-27197	47	25	quantization	quantization	NOUN
fcis-27197	47	26	.	.	PUNCT
fcis-27197	48	1	the	the	DET
fcis-27197	48	2	metrics	metric	NOUN
fcis-27197	48	3	of	of	ADP
fcis-27197	48	4	model	model	NOUN
fcis-27197	48	5	accuracy	accuracy	NOUN
fcis-27197	48	6	,	,	PUNCT
fcis-27197	48	7	inference	inference	NOUN
fcis-27197	48	8	speed	speed	NOUN
fcis-27197	48	9	,	,	PUNCT
fcis-27197	48	10	model	model	NOUN
fcis-27197	48	11	size	size	NOUN
fcis-27197	48	12	,	,	PUNCT
fcis-27197	48	13	and	and	CCONJ
fcis-27197	48	14	memory	memory	NOUN
fcis-27197	48	15	consumption	consumption	NOUN
fcis-27197	48	16	after	after	ADP
fcis-27197	48	17	quantization	quantization	NOUN
fcis-27197	48	18	are	be	AUX
fcis-27197	48	19	evaluated	evaluate	VERB
fcis-27197	48	20	,	,	PUNCT
fcis-27197	48	21	focusing	focus	VERB
fcis-27197	48	22	on	on	ADP
fcis-27197	48	23	the	the	DET
fcis-27197	48	24	trade	trade	NOUN
fcis-27197	48	25	-	-	PUNCT
fcis-27197	48	26	off	off	NOUN
fcis-27197	48	27	between	between	ADP
fcis-27197	48	28	accuracy	accuracy	NOUN
fcis-27197	48	29	loss	loss	NOUN
fcis-27197	48	30	and	and	CCONJ
fcis-27197	48	31	computational	computational	ADJ
fcis-27197	48	32	complexity	complexity	NOUN
fcis-27197	48	33	.	.	PUNCT
fcis-27197	49	1	model	model	NOUN
fcis-27197	49	2	accuracy	accuracy	NOUN
fcis-27197	49	3	,	,	PUNCT
fcis-27197	49	4	the	the	DET
fcis-27197	49	5	resnet-50	resnet-50	PROPN
fcis-27197	49	6	model	model	NOUN
fcis-27197	49	7	has	have	VERB
fcis-27197	49	8	a	a	DET
fcis-27197	49	9	top-1	top-1	DET
fcis-27197	49	10	classification	classification	NOUN
fcis-27197	49	11	accuracy	accuracy	NOUN
fcis-27197	49	12	of	of	ADP
fcis-27197	49	13	76.4	76.4	NUM
fcis-27197	49	14	%	%	NOUN
fcis-27197	49	15	and	and	CCONJ
fcis-27197	49	16	a	a	DET
fcis-27197	49	17	top-5	top-5	NOUN
fcis-27197	49	18	accuracy	accuracy	NOUN
fcis-27197	49	19	of	of	ADP
fcis-27197	49	20	93.2	93.2	NUM
fcis-27197	49	21	%	%	NOUN
fcis-27197	49	22	on	on	ADP
fcis-27197	49	23	the	the	DET
fcis-27197	49	24	imagenet	imagenet	NOUN
fcis-27197	49	25	validation	validation	NOUN
fcis-27197	49	26	set	set	NOUN
fcis-27197	49	27	.	.	PUNCT
fcis-27197	50	1	this	this	PRON
fcis-27197	50	2	is	be	AUX
fcis-27197	50	3	the	the	DET
fcis-27197	50	4	best	good	ADJ
fcis-27197	50	5	performance	performance	NOUN
fcis-27197	50	6	metric	metric	ADJ
fcis-27197	50	7	for	for	ADP
fcis-27197	50	8	the	the	DET
fcis-27197	50	9	unquantized	unquantized	ADJ
fcis-27197	50	10	model	model	NOUN
fcis-27197	50	11	.	.	PUNCT
fcis-27197	51	1	inference	inference	NOUN
fcis-27197	51	2	speed	speed	NOUN
fcis-27197	51	3	,	,	PUNCT
fcis-27197	51	4	tested	test	VERB
fcis-27197	51	5	on	on	ADP
fcis-27197	51	6	a	a	DET
fcis-27197	51	7	gpu	gpu	NOUN
fcis-27197	51	8	(	(	PUNCT
fcis-27197	51	9	nvidia	nvidia	PROPN
fcis-27197	51	10	v100	v100	PROPN
fcis-27197	51	11	)	)	PUNCT
fcis-27197	51	12	,	,	PUNCT
fcis-27197	51	13	the	the	DET
fcis-27197	51	14	inference	inference	NOUN
fcis-27197	51	15	speed	speed	NOUN
fcis-27197	51	16	of	of	ADP
fcis-27197	51	17	the	the	DET
fcis-27197	51	18	unquantized	unquantized	ADJ
fcis-27197	51	19	model	model	NOUN
fcis-27197	51	20	is	be	AUX
fcis-27197	51	21	about	about	ADV
fcis-27197	51	22	90	90	NUM
fcis-27197	51	23	fps	fps	NOUN
fcis-27197	51	24	(	(	PUNCT
fcis-27197	51	25	frames	frame	NOUN
fcis-27197	51	26	per	per	ADP
fcis-27197	51	27	second	second	NOUN
fcis-27197	51	28	)	)	PUNCT
fcis-27197	51	29	,	,	PUNCT
fcis-27197	51	30	i.e.	i.e.	X
fcis-27197	51	31	,	,	PUNCT
fcis-27197	51	32	it	it	PRON
fcis-27197	51	33	can	can	AUX
fcis-27197	51	34	process	process	VERB
fcis-27197	51	35	90	90	NUM
fcis-27197	51	36	images	image	NOUN
fcis-27197	51	37	per	per	ADP
fcis-27197	51	38	second	second	ADJ
fcis-27197	51	39	.	.	PUNCT
fcis-27197	52	1	model	model	NOUN
fcis-27197	52	2	size	size	NOUN
fcis-27197	52	3	,	,	PUNCT
fcis-27197	52	4	the	the	DET
fcis-27197	52	5	storage	storage	NOUN
fcis-27197	52	6	requirement	requirement	NOUN
fcis-27197	52	7	of	of	ADP
fcis-27197	52	8	the	the	DET
fcis-27197	52	9	unquantized	unquantized	ADJ
fcis-27197	52	10	model	model	NOUN
fcis-27197	52	11	is	be	AUX
fcis-27197	52	12	98	98	NUM
fcis-27197	52	13	mb	mb	NOUN
fcis-27197	52	14	,	,	PUNCT
fcis-27197	52	15	which	which	PRON
fcis-27197	52	16	is	be	AUX
fcis-27197	52	17	mainly	mainly	ADV
fcis-27197	52	18	determined	determine	VERB
fcis-27197	52	19	by	by	ADP
fcis-27197	52	20	the	the	DET
fcis-27197	52	21	number	number	NOUN
fcis-27197	52	22	of	of	ADP
fcis-27197	52	23	parameters	parameter	NOUN
fcis-27197	52	24	(	(	PUNCT
fcis-27197	52	25	about	about	ADV
fcis-27197	52	26	25.5	25.5	NUM
fcis-27197	52	27	million	million	NUM
fcis-27197	52	28	parameters	parameter	NOUN
fcis-27197	52	29	)	)	PUNCT
fcis-27197	52	30	.	.	PUNCT
fcis-27197	53	1	memory	memory	NOUN
fcis-27197	53	2	consumption	consumption	NOUN
fcis-27197	53	3	,	,	PUNCT
fcis-27197	53	4	during	during	ADP
fcis-27197	53	5	the	the	DET
fcis-27197	53	6	test	test	NOUN
fcis-27197	53	7	,	,	PUNCT
fcis-27197	53	8	the	the	DET
fcis-27197	53	9	model	model	NOUN
fcis-27197	53	10	occupied	occupy	VERB
fcis-27197	53	11	about	about	ADV
fcis-27197	53	12	1.5	1.5	NUM
fcis-27197	53	13	gb	gb	NOUN
fcis-27197	53	14	of	of	ADP
fcis-27197	53	15	video	video	NOUN
fcis-27197	53	16	memory	memory	NOUN
fcis-27197	53	17	.	.	PUNCT
fcis-27197	54	1	after	after	ADP
fcis-27197	54	2	8	8	NUM
fcis-27197	54	3	-	-	PUNCT
fcis-27197	54	4	bit	bit	NOUN
fcis-27197	54	5	fixed	fix	VERB
fcis-27197	54	6	-	-	PUNCT
fcis-27197	54	7	point	point	NOUN
fcis-27197	54	8	quantization	quantization	NOUN
fcis-27197	54	9	,	,	PUNCT
fcis-27197	54	10	the	the	DET
fcis-27197	54	11	top-1	top-1	DET
fcis-27197	54	12	accuracy	accuracy	NOUN
fcis-27197	54	13	of	of	ADP
fcis-27197	54	14	the	the	DET
fcis-27197	54	15	model	model	NOUN
fcis-27197	54	16	decreases	decrease	VERB
fcis-27197	54	17	to	to	ADP
fcis-27197	54	18	74.8	74.8	NUM
fcis-27197	54	19	%	%	NOUN
fcis-27197	54	20	and	and	CCONJ
fcis-27197	54	21	the	the	DET
fcis-27197	54	22	top-5	top-5	VERB
fcis-27197	54	23	accuracy	accuracy	NOUN
fcis-27197	54	24	decreases	decrease	VERB
fcis-27197	54	25	to	to	ADP
fcis-27197	54	26	92.0	92.0	NUM
fcis-27197	54	27	%	%	NOUN
fcis-27197	54	28	.	.	PUNCT
fcis-27197	55	1	this	this	PRON
fcis-27197	55	2	indicates	indicate	VERB
fcis-27197	55	3	that	that	SCONJ
fcis-27197	55	4	quantization	quantization	NOUN
fcis-27197	55	5	brings	bring	VERB
fcis-27197	55	6	some	some	DET
fcis-27197	55	7	accuracy	accuracy	NOUN
fcis-27197	55	8	loss	loss	NOUN
fcis-27197	55	9	,	,	PUNCT
fcis-27197	55	10	which	which	PRON
fcis-27197	55	11	is	be	AUX
fcis-27197	55	12	mainly	mainly	ADV
fcis-27197	55	13	due	due	ADJ
fcis-27197	55	14	to	to	ADP
fcis-27197	55	15	the	the	DET
fcis-27197	55	16	decrease	decrease	NOUN
fcis-27197	55	17	in	in	ADP
fcis-27197	55	18	parameter	parameter	NOUN
fcis-27197	55	19	accuracy	accuracy	NOUN
fcis-27197	55	20	during	during	ADP
fcis-27197	55	21	the	the	DET
fcis-27197	55	22	quantization	quantization	NOUN
fcis-27197	55	23	process	process	NOUN
fcis-27197	55	24	,	,	PUNCT
fcis-27197	55	25	resulting	result	VERB
fcis-27197	55	26	in	in	ADP
fcis-27197	55	27	a	a	DET
fcis-27197	55	28	decrease	decrease	NOUN
fcis-27197	55	29	in	in	ADP
fcis-27197	55	30	the	the	DET
fcis-27197	55	31	model	model	NOUN
fcis-27197	55	32	's	's	PART
fcis-27197	55	33	ability	ability	NOUN
fcis-27197	55	34	to	to	PART
fcis-27197	55	35	capture	capture	VERB
fcis-27197	55	36	subtle	subtle	ADJ
fcis-27197	55	37	features	feature	NOUN
fcis-27197	55	38	.	.	PUNCT
fcis-27197	56	1	the	the	DET
fcis-27197	56	2	inference	inference	NOUN
fcis-27197	56	3	speed	speed	NOUN
fcis-27197	56	4	of	of	ADP
fcis-27197	56	5	the	the	DET
fcis-27197	56	6	quantized	quantize	VERB
fcis-27197	56	7	model	model	NOUN
fcis-27197	56	8	on	on	ADP
fcis-27197	56	9	the	the	DET
fcis-27197	56	10	same	same	ADJ
fcis-27197	56	11	gpu	gpu	NOUN
fcis-27197	56	12	increases	increase	NOUN
fcis-27197	56	13	to	to	ADP
fcis-27197	56	14	150	150	NUM
fcis-27197	56	15	fps	fps	PROPN
fcis-27197	56	16	,	,	PUNCT
fcis-27197	56	17	an	an	DET
fcis-27197	56	18	improvement	improvement	NOUN
fcis-27197	56	19	of	of	ADP
fcis-27197	56	20	about	about	ADV
fcis-27197	56	21	66	66	NUM
fcis-27197	56	22	%	%	NOUN
fcis-27197	56	23	.	.	PUNCT
fcis-27197	57	1	this	this	PRON
fcis-27197	57	2	is	be	AUX
fcis-27197	57	3	due	due	ADJ
fcis-27197	57	4	to	to	ADP
fcis-27197	57	5	the	the	DET
fcis-27197	57	6	higher	high	ADJ
fcis-27197	57	7	parallelism	parallelism	NOUN
fcis-27197	57	8	of	of	ADP
fcis-27197	57	9	8	8	NUM
fcis-27197	57	10	-	-	PUNCT
fcis-27197	57	11	bit	bit	NOUN
fcis-27197	57	12	operations	operation	NOUN
fcis-27197	57	13	on	on	ADP
fcis-27197	57	14	the	the	DET
fcis-27197	57	15	gpu	gpu	NOUN
fcis-27197	57	16	,	,	PUNCT
fcis-27197	57	17	as	as	ADV
fcis-27197	57	18	well	well	ADV
fcis-27197	57	19	as	as	ADP
fcis-27197	57	20	the	the	DET
fcis-27197	57	21	reduced	reduced	ADJ
fcis-27197	57	22	computation	computation	NOUN
fcis-27197	57	23	of	of	ADP
fcis-27197	57	24	each	each	DET
fcis-27197	57	25	weight	weight	NOUN
fcis-27197	57	26	and	and	CCONJ
fcis-27197	57	27	activation	activation	NOUN
fcis-27197	57	28	value	value	NOUN
fcis-27197	57	29	,	,	PUNCT
fcis-27197	57	30	which	which	PRON
fcis-27197	57	31	significantly	significantly	ADV
fcis-27197	57	32	speeds	speed	VERB
fcis-27197	57	33	up	up	ADP
fcis-27197	57	34	the	the	DET
fcis-27197	57	35	inference	inference	NOUN
fcis-27197	57	36	of	of	ADP
fcis-27197	57	37	the	the	DET
fcis-27197	57	38	model	model	NOUN
fcis-27197	57	39	.	.	PUNCT
fcis-27197	58	1	the	the	DET
fcis-27197	58	2	size	size	NOUN
fcis-27197	58	3	of	of	ADP
fcis-27197	58	4	the	the	DET
fcis-27197	58	5	8	8	NUM
fcis-27197	58	6	-	-	PUNCT
fcis-27197	58	7	bit	bit	NOUN
fcis-27197	58	8	quantized	quantize	VERB
fcis-27197	58	9	model	model	NOUN
fcis-27197	58	10	is	be	AUX
fcis-27197	58	11	about	about	ADV
fcis-27197	58	12	25	25	NUM
fcis-27197	58	13	mb	mb	NOUN
fcis-27197	58	14	,	,	PUNCT
fcis-27197	58	15	which	which	PRON
fcis-27197	58	16	is	be	AUX
fcis-27197	58	17	about	about	ADV
fcis-27197	58	18	a	a	PRON
fcis-27197	58	19	quarter	quarter	NOUN
fcis-27197	58	20	of	of	ADP
fcis-27197	58	21	the	the	DET
fcis-27197	58	22	original	original	ADJ
fcis-27197	58	23	model	model	NOUN
fcis-27197	58	24	.	.	PUNCT
fcis-27197	59	1	the	the	DET
fcis-27197	59	2	significant	significant	ADJ
fcis-27197	59	3	reduction	reduction	NOUN
fcis-27197	59	4	in	in	ADP
fcis-27197	59	5	the	the	DET
fcis-27197	59	6	size	size	NOUN
fcis-27197	59	7	of	of	ADP
fcis-27197	59	8	the	the	DET
fcis-27197	59	9	quantized	quantize	VERB
fcis-27197	59	10	model	model	NOUN
fcis-27197	59	11	means	mean	VERB
fcis-27197	59	12	that	that	SCONJ
fcis-27197	59	13	it	it	PRON
fcis-27197	59	14	is	be	AUX
fcis-27197	59	15	more	more	ADV
fcis-27197	59	16	efficiently	efficiently	ADV
fcis-27197	59	17	deployed	deploy	VERB
fcis-27197	59	18	in	in	ADP
fcis-27197	59	19	mobile	mobile	ADJ
fcis-27197	59	20	devices	device	NOUN
fcis-27197	59	21	and	and	CCONJ
fcis-27197	59	22	embedded	embed	VERB
fcis-27197	59	23	systems	system	NOUN
fcis-27197	59	24	,	,	PUNCT
fcis-27197	59	25	reducing	reduce	VERB
fcis-27197	59	26	storage	storage	NOUN
fcis-27197	59	27	requirements	requirement	NOUN
fcis-27197	59	28	.	.	PUNCT
fcis-27197	60	1	the	the	DET
fcis-27197	60	2	memory	memory	NOUN
fcis-27197	60	3	footprint	footprint	NOUN
fcis-27197	60	4	of	of	ADP
fcis-27197	60	5	the	the	DET
fcis-27197	60	6	quantized	quantize	VERB
fcis-27197	60	7	model	model	NOUN
fcis-27197	60	8	drops	drop	VERB
fcis-27197	60	9	to	to	ADP
fcis-27197	60	10	about	about	ADV
fcis-27197	60	11	600	600	NUM
fcis-27197	60	12	mb	mb	NOUN
fcis-27197	60	13	,	,	PUNCT
fcis-27197	60	14	a	a	DET
fcis-27197	60	15	reduction	reduction	NOUN
fcis-27197	60	16	of	of	ADP
fcis-27197	60	17	about	about	ADV
fcis-27197	60	18	60	60	NUM
fcis-27197	60	19	%	%	NOUN
fcis-27197	60	20	.	.	PUNCT
fcis-27197	61	1	this	this	PRON
fcis-27197	61	2	is	be	AUX
fcis-27197	61	3	because	because	SCONJ
fcis-27197	61	4	the	the	DET
fcis-27197	61	5	activation	activation	NOUN
fcis-27197	61	6	values	value	NOUN
fcis-27197	61	7	and	and	CCONJ
fcis-27197	61	8	intermediate	intermediate	ADJ
fcis-27197	61	9	feature	feature	NOUN
fcis-27197	61	10	maps	map	NOUN
fcis-27197	61	11	of	of	ADP
fcis-27197	61	12	each	each	DET
fcis-27197	61	13	layer	layer	NOUN
fcis-27197	61	14	are	be	AUX
fcis-27197	61	15	quantized	quantize	VERB
fcis-27197	61	16	to	to	ADP
fcis-27197	61	17	8	8	NUM
fcis-27197	61	18	-	-	PUNCT
fcis-27197	61	19	bits	bit	NOUN
fcis-27197	61	20	during	during	ADP
fcis-27197	61	21	the	the	DET
fcis-27197	61	22	forward	forward	ADJ
fcis-27197	61	23	propagation	propagation	NOUN
fcis-27197	61	24	process	process	NOUN
fcis-27197	61	25	,	,	PUNCT
fcis-27197	61	26	significantly	significantly	ADV
fcis-27197	61	27	reducing	reduce	VERB
fcis-27197	61	28	memory	memory	NOUN
fcis-27197	61	29	requirements	requirement	NOUN
fcis-27197	61	30	.	.	PUNCT
fcis-27197	62	1	the	the	DET
fcis-27197	62	2	8	8	NUM
fcis-27197	62	3	-	-	PUNCT
fcis-27197	62	4	bit	bit	NOUN
fcis-27197	62	5	quantization	quantization	NOUN
fcis-27197	62	6	brings	bring	VERB
fcis-27197	62	7	about	about	ADV
fcis-27197	62	8	1.6	1.6	NUM
fcis-27197	62	9	%	%	NOUN
fcis-27197	62	10	top	top	ADJ
fcis-27197	62	11	1	1	NUM
fcis-27197	62	12	accuracy	accuracy	NOUN
fcis-27197	62	13	degradation	degradation	NOUN
fcis-27197	62	14	,	,	PUNCT
fcis-27197	62	15	but	but	CCONJ
fcis-27197	62	16	this	this	DET
fcis-27197	62	17	accuracy	accuracy	NOUN
fcis-27197	62	18	loss	loss	NOUN
fcis-27197	62	19	is	be	AUX
fcis-27197	62	20	acceptable	acceptable	ADJ
fcis-27197	62	21	in	in	ADP
fcis-27197	62	22	real	real	ADJ
fcis-27197	62	23	application	application	NOUN
fcis-27197	62	24	scenarios	scenario	NOUN
fcis-27197	62	25	after	after	ADP
fcis-27197	62	26	model	model	NOUN
fcis-27197	62	27	compression	compression	NOUN
fcis-27197	62	28	.	.	PUNCT
fcis-27197	63	1	for	for	ADP
fcis-27197	63	2	certain	certain	ADJ
fcis-27197	63	3	application	application	NOUN
fcis-27197	63	4	scenarios	scenario	NOUN
fcis-27197	63	5	with	with	ADP
fcis-27197	63	6	low	low	ADJ
fcis-27197	63	7	accuracy	accuracy	NOUN
fcis-27197	63	8	requirements	requirement	NOUN
fcis-27197	63	9	,	,	PUNCT
fcis-27197	63	10	the	the	DET
fcis-27197	63	11	accuracy	accuracy	NOUN
fcis-27197	63	12	of	of	ADP
fcis-27197	63	13	the	the	DET
fcis-27197	63	14	quantized	quantize	VERB
fcis-27197	63	15	model	model	NOUN
fcis-27197	63	16	is	be	AUX
fcis-27197	63	17	sufficient	sufficient	ADJ
fcis-27197	63	18	to	to	PART
fcis-27197	63	19	meet	meet	VERB
fcis-27197	63	20	the	the	DET
fcis-27197	63	21	needs	need	NOUN
fcis-27197	63	22	.	.	PUNCT
fcis-27197	64	1	the	the	DET
fcis-27197	64	2	computational	computational	ADJ
fcis-27197	64	3	complexity	complexity	NOUN
fcis-27197	64	4	of	of	ADP
fcis-27197	64	5	the	the	DET
fcis-27197	64	6	quantized	quantize	VERB
fcis-27197	64	7	model	model	NOUN
fcis-27197	64	8	is	be	AUX
fcis-27197	64	9	significantly	significantly	ADV
fcis-27197	64	10	reduced	reduce	VERB
fcis-27197	64	11	,	,	PUNCT
fcis-27197	64	12	and	and	CCONJ
fcis-27197	64	13	the	the	DET
fcis-27197	64	14	inference	inference	NOUN
fcis-27197	64	15	speed	speed	NOUN
fcis-27197	64	16	is	be	AUX
fcis-27197	64	17	improved	improve	VERB
fcis-27197	64	18	by	by	ADP
fcis-27197	64	19	66	66	NUM
fcis-27197	64	20	%	%	NOUN
fcis-27197	64	21	.	.	PUNCT
fcis-27197	65	1	it	it	PRON
fcis-27197	65	2	is	be	AUX
fcis-27197	65	3	useful	useful	ADJ
fcis-27197	65	4	for	for	ADP
fcis-27197	65	5	real	real	ADJ
fcis-27197	65	6	-	-	PUNCT
fcis-27197	65	7	time	time	NOUN
fcis-27197	65	8	applications	application	NOUN
fcis-27197	65	9	in	in	ADP
fcis-27197	65	10	mobile	mobile	ADJ
fcis-27197	65	11	devices	device	NOUN
fcis-27197	65	12	,	,	PUNCT
fcis-27197	65	13	embedded	embed	VERB
fcis-27197	65	14	systems	system	NOUN
fcis-27197	65	15	,	,	PUNCT
fcis-27197	65	16	etc	etc	X
fcis-27197	65	17	.	.	X
fcis-27197	65	18	6	6	X
fcis-27197	65	19	.	.	PUNCT
fcis-27197	65	20	optimization	optimization	NOUN
fcis-27197	65	21	strategies	strategy	NOUN
fcis-27197	65	22	for	for	ADP
fcis-27197	65	23	pruning	prune	VERB
fcis-27197	65	24	methods	method	NOUN
fcis-27197	65	25	the	the	DET
fcis-27197	65	26	first	first	ADJ
fcis-27197	65	27	step	step	NOUN
fcis-27197	65	28	in	in	ADP
fcis-27197	65	29	choosing	choose	VERB
fcis-27197	65	30	a	a	DET
fcis-27197	65	31	pruning	pruning	NOUN
fcis-27197	65	32	method	method	NOUN
fcis-27197	65	33	is	be	AUX
fcis-27197	65	34	to	to	PART
fcis-27197	65	35	cut	cut	VERB
fcis-27197	65	36	out	out	ADP
fcis-27197	65	37	the	the	DET
fcis-27197	65	38	convolutional	convolutional	ADJ
fcis-27197	65	39	kernels	kernel	NOUN
fcis-27197	65	40	and	and	CCONJ
fcis-27197	65	41	selectively	selectively	ADV
fcis-27197	65	42	remove	remove	VERB
fcis-27197	65	43	unimportant	unimportant	ADJ
fcis-27197	65	44	convolutional	convolutional	ADJ
fcis-27197	65	45	kernels	kernel	NOUN
fcis-27197	65	46	,	,	PUNCT
fcis-27197	65	47	thus	thus	ADV
fcis-27197	65	48	reducing	reduce	VERB
fcis-27197	65	49	the	the	DET
fcis-27197	65	50	amount	amount	NOUN
fcis-27197	65	51	of	of	ADP
fcis-27197	65	52	computation	computation	NOUN
fcis-27197	65	53	and	and	CCONJ
fcis-27197	65	54	the	the	DET
fcis-27197	65	55	number	number	NOUN
fcis-27197	65	56	of	of	ADP
fcis-27197	65	57	parameters	parameter	NOUN
fcis-27197	65	58	.	.	PUNCT
fcis-27197	66	1	pruning	prune	VERB
fcis-27197	66	2	channels	channel	NOUN
fcis-27197	66	3	removes	remove	VERB
fcis-27197	66	4	the	the	DET
fcis-27197	66	5	entire	entire	ADJ
fcis-27197	66	6	channel	channel	NOUN
fcis-27197	66	7	and	and	CCONJ
fcis-27197	66	8	can	can	AUX
fcis-27197	66	9	significantly	significantly	ADV
fcis-27197	66	10	reduce	reduce	VERB
fcis-27197	66	11	the	the	DET
fcis-27197	66	12	amount	amount	NOUN
fcis-27197	66	13	of	of	ADP
fcis-27197	66	14	computation	computation	NOUN
fcis-27197	66	15	for	for	ADP
fcis-27197	66	16	the	the	DET
fcis-27197	66	17	network	network	NOUN
fcis-27197	66	18	,	,	PUNCT
fcis-27197	66	19	but	but	CCONJ
fcis-27197	66	20	make	make	VERB
fcis-27197	66	21	sure	sure	ADJ
fcis-27197	66	22	that	that	SCONJ
fcis-27197	66	23	the	the	DET
fcis-27197	66	24	network	network	NOUN
fcis-27197	66	25	is	be	AUX
fcis-27197	66	26	still	still	ADV
fcis-27197	66	27	able	able	ADJ
fcis-27197	66	28	to	to	PART
fcis-27197	66	29	effectively	effectively	ADV
fcis-27197	66	30	capture	capture	VERB
fcis-27197	66	31	the	the	DET
fcis-27197	66	32	features	feature	NOUN
fcis-27197	66	33	of	of	ADP
fcis-27197	66	34	the	the	DET
fcis-27197	66	35	input	input	NOUN
fcis-27197	66	36	image	image	NOUN
fcis-27197	66	37	after	after	ADP
fcis-27197	66	38	pruning	prune	VERB
fcis-27197	66	39	.	.	PUNCT
fcis-27197	67	1	clipping	clip	VERB
fcis-27197	67	2	layers	layer	NOUN
fcis-27197	67	3	can	can	AUX
fcis-27197	67	4	,	,	PUNCT
fcis-27197	67	5	in	in	ADP
fcis-27197	67	6	some	some	DET
fcis-27197	67	7	cases	case	NOUN
fcis-27197	67	8	,	,	PUNCT
fcis-27197	67	9	remove	remove	VERB
fcis-27197	67	10	entire	entire	ADJ
fcis-27197	67	11	convolutional	convolutional	ADJ
fcis-27197	67	12	layers	layer	NOUN
fcis-27197	67	13	,	,	PUNCT
fcis-27197	67	14	but	but	CCONJ
fcis-27197	67	15	this	this	PRON
fcis-27197	67	16	usually	usually	ADV
fcis-27197	67	17	significantly	significantly	ADV
fcis-27197	67	18	affects	affect	VERB
fcis-27197	67	19	the	the	DET
fcis-27197	67	20	feature	feature	NOUN
fcis-27197	67	21	extraction	extraction	NOUN
fcis-27197	67	22	capability	capability	NOUN
fcis-27197	67	23	of	of	ADP
fcis-27197	67	24	the	the	DET
fcis-27197	67	25	model	model	NOUN
fcis-27197	67	26	,	,	PUNCT
fcis-27197	67	27	so	so	SCONJ
fcis-27197	67	28	special	special	ADJ
fcis-27197	67	29	care	care	NOUN
fcis-27197	67	30	is	be	AUX
fcis-27197	67	31	needed	need	VERB
fcis-27197	67	32	.	.	PUNCT
fcis-27197	68	1	6.1	6.1	NUM
fcis-27197	68	2	.	.	PUNCT
fcis-27197	68	3	iterative	iterative	NOUN
fcis-27197	68	4	pruning	pruning	NOUN
fcis-27197	68	5	process	process	NOUN
fcis-27197	68	6	the	the	DET
fcis-27197	68	7	number	number	NOUN
fcis-27197	68	8	of	of	ADP
fcis-27197	68	9	parameters	parameter	NOUN
fcis-27197	68	10	and	and	CCONJ
fcis-27197	68	11	computation	computation	NOUN
fcis-27197	68	12	of	of	ADP
fcis-27197	68	13	the	the	DET
fcis-27197	68	14	model	model	NOUN
fcis-27197	68	15	will	will	AUX
fcis-27197	68	16	be	be	AUX
fcis-27197	68	17	significantly	significantly	ADV
fcis-27197	68	18	reduced	reduce	VERB
fcis-27197	68	19	after	after	ADP
fcis-27197	68	20	the	the	DET
fcis-27197	68	21	initial	initial	ADJ
fcis-27197	68	22	pruning	pruning	NOUN
fcis-27197	68	23	,	,	PUNCT
fcis-27197	68	24	but	but	CCONJ
fcis-27197	68	25	the	the	DET
fcis-27197	68	26	accuracy	accuracy	NOUN
fcis-27197	68	27	may	may	AUX
fcis-27197	68	28	be	be	AUX
fcis-27197	68	29	significantly	significantly	ADV
fcis-27197	68	30	reduced	reduce	VERB
fcis-27197	68	31	.	.	PUNCT
fcis-27197	69	1	retraining	retrain	VERB
fcis-27197	69	2	of	of	ADP
fcis-27197	69	3	the	the	DET
fcis-27197	69	4	pruned	prune	VERB
fcis-27197	69	5	model	model	NOUN
fcis-27197	69	6	is	be	AUX
fcis-27197	69	7	performed	perform	VERB
fcis-27197	69	8	to	to	PART
fcis-27197	69	9	retune	retune	VERB
fcis-27197	69	10	the	the	DET
fcis-27197	69	11	model	model	NOUN
fcis-27197	69	12	parameters	parameter	NOUN
fcis-27197	69	13	using	use	VERB
fcis-27197	69	14	the	the	DET
fcis-27197	69	15	training	training	NOUN
fcis-27197	69	16	data	datum	NOUN
fcis-27197	69	17	from	from	ADP
fcis-27197	69	18	the	the	DET
fcis-27197	69	19	unpruned	unpruned	ADJ
fcis-27197	69	20	time	time	NOUN
fcis-27197	69	21	to	to	PART
fcis-27197	69	22	restore	restore	VERB
fcis-27197	69	23	the	the	DET
fcis-27197	69	24	classification	classification	NOUN
fcis-27197	69	25	accuracy	accuracy	NOUN
fcis-27197	69	26	of	of	ADP
fcis-27197	69	27	the	the	DET
fcis-27197	69	28	model	model	NOUN
fcis-27197	69	29	[	[	X
fcis-27197	69	30	3	3	NUM
fcis-27197	69	31	]	]	PUNCT
fcis-27197	69	32	.	.	PUNCT
fcis-27197	70	1	a	a	DET
fcis-27197	70	2	lower	low	ADJ
fcis-27197	70	3	learning	learning	NOUN
fcis-27197	70	4	rate	rate	NOUN
fcis-27197	70	5	can	can	AUX
fcis-27197	70	6	be	be	AUX
fcis-27197	70	7	used	use	VERB
fcis-27197	70	8	for	for	ADP
fcis-27197	70	9	retraining	retrain	VERB
fcis-27197	70	10	to	to	PART
fcis-27197	70	11	avoid	avoid	VERB
fcis-27197	70	12	drastic	drastic	ADJ
fcis-27197	70	13	parameter	parameter	NOUN
fcis-27197	70	14	changes	change	NOUN
fcis-27197	70	15	in	in	ADP
fcis-27197	70	16	the	the	DET
fcis-27197	70	17	model	model	NOUN
fcis-27197	70	18	.	.	PUNCT
fcis-27197	71	1	the	the	DET
fcis-27197	71	2	pruning	prune	VERB
fcis-27197	71	3	and	and	CCONJ
fcis-27197	71	4	retraining	retrain	VERB
fcis-27197	71	5	process	process	NOUN
fcis-27197	71	6	described	describe	VERB
fcis-27197	71	7	above	above	ADV
fcis-27197	71	8	is	be	AUX
fcis-27197	71	9	repeated	repeat	VERB
fcis-27197	71	10	,	,	PUNCT
fcis-27197	71	11	with	with	ADP
fcis-27197	71	12	a	a	DET
fcis-27197	71	13	portion	portion	NOUN
fcis-27197	71	14	of	of	ADP
fcis-27197	71	15	unimportant	unimportant	ADJ
fcis-27197	71	16	convolutional	convolutional	ADJ
fcis-27197	71	17	kernels	kernel	NOUN
fcis-27197	71	18	or	or	CCONJ
fcis-27197	71	19	channels	channel	NOUN
fcis-27197	71	20	cut	cut	VERB
fcis-27197	71	21	out	out	ADP
fcis-27197	71	22	each	each	DET
fcis-27197	71	23	time	time	NOUN
fcis-27197	71	24	and	and	CCONJ
fcis-27197	71	25	retrained	retrain	VERB
fcis-27197	71	26	.	.	PUNCT
fcis-27197	72	1	through	through	ADP
fcis-27197	72	2	multiple	multiple	ADJ
fcis-27197	72	3	iterations	iteration	NOUN
fcis-27197	72	4	,	,	PUNCT
fcis-27197	72	5	the	the	DET
fcis-27197	72	6	computational	computational	ADJ
fcis-27197	72	7	and	and	CCONJ
fcis-27197	72	8	parametric	parametric	ADJ
fcis-27197	72	9	quantities	quantity	NOUN
fcis-27197	72	10	of	of	ADP
fcis-27197	72	11	the	the	DET
fcis-27197	72	12	model	model	NOUN
fcis-27197	72	13	are	be	AUX
fcis-27197	72	14	gradually	gradually	ADV
fcis-27197	72	15	reduced	reduce	VERB
fcis-27197	72	16	41	41	NUM
fcis-27197	72	17	while	while	SCONJ
fcis-27197	72	18	keeping	keep	VERB
fcis-27197	72	19	the	the	DET
fcis-27197	72	20	accuracy	accuracy	NOUN
fcis-27197	72	21	loss	loss	NOUN
fcis-27197	72	22	within	within	ADP
fcis-27197	72	23	acceptable	acceptable	ADJ
fcis-27197	72	24	limits	limit	NOUN
fcis-27197	72	25	[	[	X
fcis-27197	72	26	5	5	NUM
fcis-27197	72	27	]	]	PUNCT
fcis-27197	72	28	.	.	PUNCT
fcis-27197	73	1	6.2	6.2	NUM
fcis-27197	73	2	.	.	PUNCT
fcis-27197	74	1	hyperparameter	hyperparameter	NOUN
fcis-27197	74	2	selection	selection	NOUN
fcis-27197	74	3	during	during	ADP
fcis-27197	74	4	pruning	prune	VERB
fcis-27197	74	5	the	the	DET
fcis-27197	74	6	proportion	proportion	NOUN
fcis-27197	74	7	of	of	ADP
fcis-27197	74	8	convolutional	convolutional	ADJ
fcis-27197	74	9	kernels	kernel	NOUN
fcis-27197	74	10	or	or	CCONJ
fcis-27197	74	11	channels	channel	NOUN
fcis-27197	74	12	removed	remove	VERB
fcis-27197	74	13	at	at	ADP
fcis-27197	74	14	each	each	DET
fcis-27197	74	15	pruning	pruning	NOUN
fcis-27197	74	16	is	be	AUX
fcis-27197	74	17	an	an	DET
fcis-27197	74	18	important	important	ADJ
fcis-27197	74	19	hyperparameter	hyperparameter	NOUN
fcis-27197	74	20	in	in	ADP
fcis-27197	74	21	the	the	DET
fcis-27197	74	22	pruning	pruning	NOUN
fcis-27197	74	23	process	process	NOUN
fcis-27197	74	24	.	.	PUNCT
fcis-27197	75	1	it	it	PRON
fcis-27197	75	2	can	can	AUX
fcis-27197	75	3	usually	usually	ADV
fcis-27197	75	4	be	be	AUX
fcis-27197	75	5	started	start	VERB
fcis-27197	75	6	from	from	ADP
fcis-27197	75	7	10	10	NUM
fcis-27197	75	8	%	%	NOUN
fcis-27197	75	9	and	and	CCONJ
fcis-27197	75	10	gradually	gradually	ADV
fcis-27197	75	11	adjusted	adjust	VERB
fcis-27197	75	12	experimentally	experimentally	ADV
fcis-27197	75	13	to	to	PART
fcis-27197	75	14	find	find	VERB
fcis-27197	75	15	the	the	DET
fcis-27197	75	16	optimal	optimal	ADJ
fcis-27197	75	17	pruning	pruning	NOUN
fcis-27197	75	18	ratio	ratio	NOUN
fcis-27197	75	19	.	.	PUNCT
fcis-27197	76	1	the	the	DET
fcis-27197	76	2	number	number	NOUN
fcis-27197	76	3	of	of	ADP
fcis-27197	76	4	retraining	retrain	VERB
fcis-27197	76	5	sessions	session	NOUN
fcis-27197	76	6	after	after	ADP
fcis-27197	76	7	pruning	prune	VERB
fcis-27197	76	8	depends	depend	VERB
fcis-27197	76	9	on	on	ADP
fcis-27197	76	10	the	the	DET
fcis-27197	76	11	recovery	recovery	NOUN
fcis-27197	76	12	of	of	ADP
fcis-27197	76	13	model	model	NOUN
fcis-27197	76	14	accuracy	accuracy	NOUN
fcis-27197	76	15	.	.	PUNCT
fcis-27197	77	1	usually	usually	ADV
fcis-27197	77	2	,	,	PUNCT
fcis-27197	77	3	several	several	ADJ
fcis-27197	77	4	complete	complete	ADJ
fcis-27197	77	5	rounds	round	NOUN
fcis-27197	77	6	of	of	ADP
fcis-27197	77	7	retraining	retraining	NOUN
fcis-27197	77	8	are	be	AUX
fcis-27197	77	9	required	require	VERB
fcis-27197	77	10	after	after	ADP
fcis-27197	77	11	pruning	prune	VERB
fcis-27197	77	12	to	to	PART
fcis-27197	77	13	recover	recover	VERB
fcis-27197	77	14	the	the	DET
fcis-27197	77	15	model	model	NOUN
fcis-27197	77	16	accuracy	accuracy	NOUN
fcis-27197	77	17	[	[	X
fcis-27197	77	18	3,4	3,4	NUM
fcis-27197	77	19	]	]	PUNCT
fcis-27197	77	20	.	.	PUNCT
fcis-27197	78	1	6.3	6.3	NUM
fcis-27197	78	2	.	.	PUNCT
fcis-27197	78	3	experimental	experimental	ADJ
fcis-27197	78	4	design	design	NOUN
fcis-27197	78	5	and	and	CCONJ
fcis-27197	78	6	evaluation	evaluation	NOUN
fcis-27197	78	7	of	of	ADP
fcis-27197	78	8	effects	effect	NOUN
fcis-27197	78	9	the	the	DET
fcis-27197	78	10	imagenet	imagenet	NOUN
fcis-27197	78	11	dataset	dataset	NOUN
fcis-27197	78	12	is	be	AUX
fcis-27197	78	13	selected	select	VERB
fcis-27197	78	14	for	for	ADP
fcis-27197	78	15	the	the	DET
fcis-27197	78	16	experiment	experiment	NOUN
fcis-27197	78	17	and	and	CCONJ
fcis-27197	78	18	the	the	DET
fcis-27197	78	19	benchmark	benchmark	NOUN
fcis-27197	78	20	model	model	NOUN
fcis-27197	78	21	is	be	AUX
fcis-27197	78	22	resnet-50	resnet-50	PROPN
fcis-27197	78	23	.	.	PUNCT
fcis-27197	79	1	the	the	DET
fcis-27197	79	2	original	original	ADJ
fcis-27197	79	3	resnet-50	resnet-50	PROPN
fcis-27197	79	4	model	model	NOUN
fcis-27197	79	5	has	have	VERB
fcis-27197	79	6	a	a	DET
fcis-27197	79	7	top-1	top-1	NOUN
fcis-27197	79	8	accuracy	accuracy	NOUN
fcis-27197	79	9	of	of	ADP
fcis-27197	79	10	76.4	76.4	NUM
fcis-27197	79	11	%	%	NOUN
fcis-27197	79	12	and	and	CCONJ
fcis-27197	79	13	a	a	DET
fcis-27197	79	14	top-5	top-5	NOUN
fcis-27197	79	15	accuracy	accuracy	NOUN
fcis-27197	79	16	of	of	ADP
fcis-27197	79	17	93.2	93.2	NUM
fcis-27197	79	18	%	%	NOUN
fcis-27197	79	19	on	on	ADP
fcis-27197	79	20	the	the	DET
fcis-27197	79	21	imagenet	imagenet	NOUN
fcis-27197	79	22	validation	validation	NOUN
fcis-27197	79	23	set	set	NOUN
fcis-27197	79	24	.	.	PUNCT
fcis-27197	80	1	after	after	ADP
fcis-27197	80	2	several	several	ADJ
fcis-27197	80	3	rounds	round	NOUN
fcis-27197	80	4	of	of	ADP
fcis-27197	80	5	pruning	prune	VERB
fcis-27197	80	6	and	and	CCONJ
fcis-27197	80	7	retraining	retrain	VERB
fcis-27197	80	8	,	,	PUNCT
fcis-27197	80	9	the	the	DET
fcis-27197	80	10	model	model	NOUN
fcis-27197	80	11	with	with	ADP
fcis-27197	80	12	a	a	DET
fcis-27197	80	13	pruning	prune	VERB
fcis-27197	80	14	ratio	ratio	NOUN
fcis-27197	80	15	of	of	ADP
fcis-27197	80	16	50	50	NUM
fcis-27197	80	17	%	%	NOUN
fcis-27197	80	18	is	be	AUX
fcis-27197	80	19	able	able	ADJ
fcis-27197	80	20	to	to	PART
fcis-27197	80	21	control	control	VERB
fcis-27197	80	22	the	the	DET
fcis-27197	80	23	top-1	top-1	DET
fcis-27197	80	24	accuracy	accuracy	NOUN
fcis-27197	80	25	at	at	ADP
fcis-27197	80	26	about	about	ADV
fcis-27197	80	27	74.0	74.0	NUM
fcis-27197	80	28	%	%	NOUN
fcis-27197	80	29	and	and	CCONJ
fcis-27197	80	30	the	the	DET
fcis-27197	80	31	top-5	top-5	PUNCT
fcis-27197	80	32	accuracy	accuracy	NOUN
fcis-27197	80	33	at	at	ADP
fcis-27197	80	34	about	about	ADV
fcis-27197	80	35	91.5	91.5	NUM
fcis-27197	80	36	%	%	NOUN
fcis-27197	80	37	.	.	PUNCT
fcis-27197	81	1	the	the	DET
fcis-27197	81	2	size	size	NOUN
fcis-27197	81	3	of	of	ADP
fcis-27197	81	4	the	the	DET
fcis-27197	81	5	pruned	prune	VERB
fcis-27197	81	6	model	model	NOUN
fcis-27197	81	7	is	be	AUX
fcis-27197	81	8	reduced	reduce	VERB
fcis-27197	81	9	by	by	ADP
fcis-27197	81	10	about	about	ADV
fcis-27197	81	11	50	50	NUM
fcis-27197	81	12	%	%	NOUN
fcis-27197	81	13	,	,	PUNCT
fcis-27197	81	14	from	from	ADP
fcis-27197	81	15	98	98	NUM
fcis-27197	81	16	mb	mb	NOUN
fcis-27197	81	17	to	to	ADP
fcis-27197	81	18	about	about	ADP
fcis-27197	81	19	49	49	NUM
fcis-27197	81	20	mb	mb	NOUN
fcis-27197	81	21	,	,	PUNCT
fcis-27197	81	22	and	and	CCONJ
fcis-27197	81	23	the	the	DET
fcis-27197	81	24	inference	inference	NOUN
fcis-27197	81	25	speed	speed	NOUN
fcis-27197	81	26	of	of	ADP
fcis-27197	81	27	the	the	DET
fcis-27197	81	28	pruned	prune	VERB
fcis-27197	81	29	model	model	NOUN
fcis-27197	81	30	is	be	AUX
fcis-27197	81	31	improved	improve	VERB
fcis-27197	81	32	by	by	ADP
fcis-27197	81	33	about	about	ADV
fcis-27197	81	34	40	40	NUM
fcis-27197	81	35	%	%	NOUN
fcis-27197	81	36	on	on	ADP
fcis-27197	81	37	gpu	gpu	PROPN
fcis-27197	81	38	,	,	PUNCT
fcis-27197	81	39	indicating	indicate	VERB
fcis-27197	81	40	that	that	SCONJ
fcis-27197	81	41	structured	structured	ADJ
fcis-27197	81	42	pruning	pruning	NOUN
fcis-27197	81	43	has	have	VERB
fcis-27197	81	44	obvious	obvious	ADJ
fcis-27197	81	45	advantages	advantage	NOUN
fcis-27197	81	46	in	in	ADP
fcis-27197	81	47	improving	improve	VERB
fcis-27197	81	48	inference	inference	NOUN
fcis-27197	81	49	efficiency	efficiency	NOUN
fcis-27197	81	50	.	.	PUNCT
fcis-27197	82	1	7	7	X
fcis-27197	82	2	.	.	X
fcis-27197	82	3	combining	combine	VERB
fcis-27197	82	4	strategies	strategy	NOUN
fcis-27197	82	5	while	while	SCONJ
fcis-27197	82	6	quantization	quantization	NOUN
fcis-27197	82	7	or	or	CCONJ
fcis-27197	82	8	pruning	prune	VERB
fcis-27197	82	9	techniques	technique	NOUN
fcis-27197	82	10	alone	alone	ADV
fcis-27197	82	11	can	can	AUX
fcis-27197	82	12	be	be	AUX
fcis-27197	82	13	effective	effective	ADJ
fcis-27197	82	14	in	in	ADP
fcis-27197	82	15	compressing	compress	VERB
fcis-27197	82	16	models	model	NOUN
fcis-27197	82	17	,	,	PUNCT
fcis-27197	82	18	a	a	DET
fcis-27197	82	19	combination	combination	NOUN
fcis-27197	82	20	of	of	ADP
fcis-27197	82	21	the	the	DET
fcis-27197	82	22	two	two	NUM
fcis-27197	82	23	can	can	AUX
fcis-27197	82	24	further	far	ADV
fcis-27197	82	25	optimize	optimize	VERB
fcis-27197	82	26	model	model	NOUN
fcis-27197	82	27	performance	performance	NOUN
fcis-27197	82	28	.	.	PUNCT
fcis-27197	83	1	however	however	ADV
fcis-27197	83	2	,	,	PUNCT
fcis-27197	83	3	quantization	quantization	NOUN
fcis-27197	83	4	faces	face	VERB
fcis-27197	83	5	challenges	challenge	NOUN
fcis-27197	83	6	when	when	SCONJ
fcis-27197	83	7	combined	combine	VERB
fcis-27197	83	8	with	with	ADP
fcis-27197	83	9	pruning	prune	VERB
fcis-27197	83	10	.	.	PUNCT
fcis-27197	84	1	both	both	DET
fcis-27197	84	2	quantization	quantization	NOUN
fcis-27197	84	3	and	and	CCONJ
fcis-27197	84	4	pruning	prune	VERB
fcis-27197	84	5	introduce	introduce	NOUN
fcis-27197	84	6	a	a	DET
fcis-27197	84	7	certain	certain	ADJ
fcis-27197	84	8	loss	loss	NOUN
fcis-27197	84	9	of	of	ADP
fcis-27197	84	10	accuracy	accuracy	NOUN
fcis-27197	84	11	,	,	PUNCT
fcis-27197	84	12	and	and	CCONJ
fcis-27197	84	13	the	the	DET
fcis-27197	84	14	superposition	superposition	NOUN
fcis-27197	84	15	of	of	ADP
fcis-27197	84	16	the	the	DET
fcis-27197	84	17	two	two	NUM
fcis-27197	84	18	may	may	AUX
fcis-27197	84	19	lead	lead	VERB
fcis-27197	84	20	to	to	ADP
fcis-27197	84	21	an	an	DET
fcis-27197	84	22	excessive	excessive	ADJ
fcis-27197	84	23	decrease	decrease	NOUN
fcis-27197	84	24	in	in	ADP
fcis-27197	84	25	accuracy	accuracy	NOUN
fcis-27197	84	26	,	,	PUNCT
fcis-27197	84	27	making	make	VERB
fcis-27197	84	28	it	it	PRON
fcis-27197	84	29	difficult	difficult	ADJ
fcis-27197	84	30	to	to	PART
fcis-27197	84	31	meet	meet	VERB
fcis-27197	84	32	the	the	DET
fcis-27197	84	33	accuracy	accuracy	NOUN
fcis-27197	84	34	requirements	requirement	NOUN
fcis-27197	84	35	of	of	ADP
fcis-27197	84	36	practical	practical	ADJ
fcis-27197	84	37	applications	application	NOUN
fcis-27197	84	38	.	.	PUNCT
fcis-27197	85	1	choosing	choose	VERB
fcis-27197	85	2	the	the	DET
fcis-27197	85	3	appropriate	appropriate	ADJ
fcis-27197	85	4	combination	combination	NOUN
fcis-27197	85	5	order	order	NOUN
fcis-27197	85	6	is	be	AUX
fcis-27197	85	7	crucial	crucial	ADJ
fcis-27197	85	8	to	to	PART
fcis-27197	85	9	maximize	maximize	VERB
fcis-27197	85	10	the	the	DET
fcis-27197	85	11	model	model	NOUN
fcis-27197	85	12	performance	performance	NOUN
fcis-27197	85	13	[	[	X
fcis-27197	85	14	4	4	NUM
fcis-27197	85	15	]	]	PUNCT
fcis-27197	85	16	.	.	PUNCT
fcis-27197	86	1	the	the	DET
fcis-27197	86	2	structure	structure	NOUN
fcis-27197	86	3	of	of	ADP
fcis-27197	86	4	the	the	DET
fcis-27197	86	5	model	model	NOUN
fcis-27197	86	6	changes	change	NOUN
fcis-27197	86	7	after	after	ADP
fcis-27197	86	8	pruning	prune	VERB
fcis-27197	86	9	,	,	PUNCT
fcis-27197	86	10	while	while	SCONJ
fcis-27197	86	11	quantization	quantization	NOUN
fcis-27197	86	12	requires	require	VERB
fcis-27197	86	13	stable	stable	ADJ
fcis-27197	86	14	weight	weight	NOUN
fcis-27197	86	15	distribution	distribution	NOUN
fcis-27197	86	16	in	in	ADP
fcis-27197	86	17	each	each	DET
fcis-27197	86	18	layer	layer	NOUN
fcis-27197	86	19	,	,	PUNCT
fcis-27197	86	20	so	so	CCONJ
fcis-27197	86	21	the	the	DET
fcis-27197	86	22	training	training	NOUN
fcis-27197	86	23	stability	stability	NOUN
fcis-27197	86	24	of	of	ADP
fcis-27197	86	25	the	the	DET
fcis-27197	86	26	model	model	NOUN
fcis-27197	86	27	after	after	ADP
fcis-27197	86	28	pruning	prune	VERB
fcis-27197	86	29	becomes	become	VERB
fcis-27197	86	30	a	a	DET
fcis-27197	86	31	problem	problem	NOUN
fcis-27197	86	32	to	to	PART
fcis-27197	86	33	be	be	AUX
fcis-27197	86	34	solved	solve	VERB
fcis-27197	86	35	.	.	PUNCT
fcis-27197	87	1	7.1	7.1	NUM
fcis-27197	87	2	.	.	PUNCT
fcis-27197	87	3	design	design	NOUN
fcis-27197	87	4	and	and	CCONJ
fcis-27197	87	5	implementation	implementation	NOUN
fcis-27197	87	6	of	of	ADP
fcis-27197	87	7	combination	combination	NOUN
fcis-27197	87	8	strategies	strategy	NOUN
fcis-27197	87	9	the	the	DET
fcis-27197	87	10	first	first	ADJ
fcis-27197	87	11	strategy	strategy	NOUN
fcis-27197	87	12	is	be	AUX
fcis-27197	87	13	pruning	prune	VERB
fcis-27197	87	14	followed	follow	VERB
fcis-27197	87	15	by	by	ADP
fcis-27197	87	16	quantization	quantization	NOUN
fcis-27197	87	17	,	,	PUNCT
fcis-27197	87	18	where	where	SCONJ
fcis-27197	87	19	unimportant	unimportant	ADJ
fcis-27197	87	20	convolutional	convolutional	ADJ
fcis-27197	87	21	kernels	kernel	NOUN
fcis-27197	87	22	or	or	CCONJ
fcis-27197	87	23	channels	channel	NOUN
fcis-27197	87	24	are	be	AUX
fcis-27197	87	25	first	first	ADV
fcis-27197	87	26	removed	remove	VERB
fcis-27197	87	27	from	from	ADP
fcis-27197	87	28	the	the	DET
fcis-27197	87	29	model	model	NOUN
fcis-27197	87	30	based	base	VERB
fcis-27197	87	31	on	on	ADP
fcis-27197	87	32	weight	weight	NOUN
fcis-27197	87	33	importance	importance	NOUN
fcis-27197	87	34	or	or	CCONJ
fcis-27197	87	35	redundancy	redundancy	NOUN
fcis-27197	87	36	analysis	analysis	NOUN
fcis-27197	87	37	.	.	PUNCT
fcis-27197	88	1	this	this	DET
fcis-27197	88	2	stage	stage	NOUN
fcis-27197	88	3	can	can	AUX
fcis-27197	88	4	be	be	AUX
fcis-27197	88	5	done	do	VERB
fcis-27197	88	6	by	by	ADP
fcis-27197	88	7	using	use	VERB
fcis-27197	88	8	structured	structured	ADJ
fcis-27197	88	9	pruning	pruning	NOUN
fcis-27197	88	10	methods	method	NOUN
fcis-27197	88	11	to	to	PART
fcis-27197	88	12	remove	remove	VERB
fcis-27197	88	13	redundant	redundant	ADJ
fcis-27197	88	14	convolutional	convolutional	ADJ
fcis-27197	88	15	kernels	kernel	NOUN
fcis-27197	88	16	,	,	PUNCT
fcis-27197	88	17	channels	channel	NOUN
fcis-27197	88	18	,	,	PUNCT
fcis-27197	88	19	or	or	CCONJ
fcis-27197	88	20	entire	entire	ADJ
fcis-27197	88	21	convolutional	convolutional	ADJ
fcis-27197	88	22	layers	layer	NOUN
fcis-27197	88	23	.	.	PUNCT
fcis-27197	89	1	the	the	DET
fcis-27197	89	2	model	model	NOUN
fcis-27197	89	3	accuracy	accuracy	NOUN
fcis-27197	89	4	usually	usually	ADV
fcis-27197	89	5	decreases	decrease	VERB
fcis-27197	89	6	after	after	ADP
fcis-27197	89	7	pruning	prune	VERB
fcis-27197	89	8	,	,	PUNCT
fcis-27197	89	9	so	so	CCONJ
fcis-27197	89	10	the	the	DET
fcis-27197	89	11	pruned	prune	VERB
fcis-27197	89	12	model	model	NOUN
fcis-27197	89	13	needs	need	VERB
fcis-27197	89	14	to	to	PART
fcis-27197	89	15	be	be	AUX
fcis-27197	89	16	retrained	retrain	VERB
fcis-27197	89	17	to	to	PART
fcis-27197	89	18	restore	restore	VERB
fcis-27197	89	19	accuracy	accuracy	NOUN
fcis-27197	89	20	.	.	PUNCT
fcis-27197	90	1	at	at	ADP
fcis-27197	90	2	this	this	DET
fcis-27197	90	3	stage	stage	NOUN
fcis-27197	90	4	,	,	PUNCT
fcis-27197	90	5	the	the	DET
fcis-27197	90	6	structure	structure	NOUN
fcis-27197	90	7	of	of	ADP
fcis-27197	90	8	the	the	DET
fcis-27197	90	9	model	model	NOUN
fcis-27197	90	10	is	be	AUX
fcis-27197	90	11	fixed	fix	VERB
fcis-27197	90	12	,	,	PUNCT
fcis-27197	90	13	providing	provide	VERB
fcis-27197	90	14	a	a	DET
fcis-27197	90	15	basis	basis	NOUN
fcis-27197	90	16	for	for	ADP
fcis-27197	90	17	subsequent	subsequent	ADJ
fcis-27197	90	18	quantization	quantization	NOUN
fcis-27197	90	19	.	.	PUNCT
fcis-27197	91	1	quantization	quantization	NOUN
fcis-27197	91	2	aware	aware	ADJ
fcis-27197	91	3	training	training	NOUN
fcis-27197	91	4	of	of	ADP
fcis-27197	91	5	the	the	DET
fcis-27197	91	6	pruned	prune	VERB
fcis-27197	91	7	model	model	NOUN
fcis-27197	91	8	.	.	PUNCT
fcis-27197	92	1	since	since	SCONJ
fcis-27197	92	2	the	the	DET
fcis-27197	92	3	pruned	prune	VERB
fcis-27197	92	4	model	model	NOUN
fcis-27197	92	5	is	be	AUX
fcis-27197	92	6	already	already	ADV
fcis-27197	92	7	more	more	ADV
fcis-27197	92	8	streamlined	streamlined	ADJ
fcis-27197	92	9	,	,	PUNCT
fcis-27197	92	10	qat	qat	PROPN
fcis-27197	92	11	can	can	AUX
fcis-27197	92	12	further	far	ADV
fcis-27197	92	13	quantize	quantize	VERB
fcis-27197	92	14	the	the	DET
fcis-27197	92	15	weights	weight	NOUN
fcis-27197	92	16	and	and	CCONJ
fcis-27197	92	17	activation	activation	NOUN
fcis-27197	92	18	values	value	NOUN
fcis-27197	92	19	into	into	ADP
fcis-27197	92	20	low	low	ADJ
fcis-27197	92	21	precision	precision	NOUN
fcis-27197	92	22	fixed	fix	VERB
fcis-27197	92	23	-	-	PUNCT
fcis-27197	92	24	point	point	NOUN
fcis-27197	92	25	representations	representation	NOUN
fcis-27197	92	26	,	,	PUNCT
fcis-27197	92	27	thus	thus	ADV
fcis-27197	92	28	significantly	significantly	ADV
fcis-27197	92	29	reducing	reduce	VERB
fcis-27197	92	30	the	the	DET
fcis-27197	92	31	amount	amount	NOUN
fcis-27197	92	32	of	of	ADP
fcis-27197	92	33	computation	computation	NOUN
fcis-27197	92	34	and	and	CCONJ
fcis-27197	92	35	storage	storage	NOUN
fcis-27197	92	36	requirements	requirement	NOUN
fcis-27197	92	37	[	[	X
fcis-27197	92	38	1	1	NUM
fcis-27197	92	39	-	-	SYM
fcis-27197	92	40	4	4	NUM
fcis-27197	92	41	]	]	PUNCT
fcis-27197	92	42	.	.	PUNCT
fcis-27197	93	1	then	then	ADV
fcis-27197	93	2	there	there	PRON
fcis-27197	93	3	is	be	VERB
fcis-27197	93	4	a	a	DET
fcis-27197	93	5	post	post	ADJ
fcis-27197	93	6	-	-	ADJ
fcis-27197	93	7	quantization	quantization	NOUN
fcis-27197	93	8	pruning	pruning	NOUN
fcis-27197	93	9	strategy	strategy	NOUN
fcis-27197	93	10	,	,	PUNCT
fcis-27197	93	11	which	which	PRON
fcis-27197	93	12	first	first	ADV
fcis-27197	93	13	performs	perform	VERB
fcis-27197	93	14	preliminary	preliminary	ADJ
fcis-27197	93	15	quantization	quantization	NOUN
fcis-27197	93	16	,	,	PUNCT
fcis-27197	93	17	quantization	quantization	NOUN
fcis-27197	93	18	-	-	PUNCT
fcis-27197	93	19	aware	aware	ADJ
fcis-27197	93	20	training	training	NOUN
fcis-27197	93	21	on	on	ADP
fcis-27197	93	22	the	the	DET
fcis-27197	93	23	original	original	ADJ
fcis-27197	93	24	model	model	NOUN
fcis-27197	93	25	,	,	PUNCT
fcis-27197	93	26	and	and	CCONJ
fcis-27197	93	27	converts	convert	VERB
fcis-27197	93	28	the	the	DET
fcis-27197	93	29	weights	weight	NOUN
fcis-27197	93	30	and	and	CCONJ
fcis-27197	93	31	activation	activation	NOUN
fcis-27197	93	32	values	value	NOUN
fcis-27197	93	33	of	of	ADP
fcis-27197	93	34	the	the	DET
fcis-27197	93	35	model	model	NOUN
fcis-27197	93	36	into	into	ADP
fcis-27197	93	37	fixed	fix	VERB
fcis-27197	93	38	-	-	PUNCT
fcis-27197	93	39	point	point	NOUN
fcis-27197	93	40	representations	representation	NOUN
fcis-27197	93	41	.	.	PUNCT
fcis-27197	94	1	this	this	DET
fcis-27197	94	2	stage	stage	NOUN
fcis-27197	94	3	can	can	AUX
fcis-27197	94	4	reduce	reduce	VERB
fcis-27197	94	5	the	the	DET
fcis-27197	94	6	computational	computational	ADJ
fcis-27197	94	7	complexity	complexity	NOUN
fcis-27197	94	8	of	of	ADP
fcis-27197	94	9	the	the	DET
fcis-27197	94	10	model	model	NOUN
fcis-27197	94	11	and	and	CCONJ
fcis-27197	94	12	provide	provide	VERB
fcis-27197	94	13	an	an	DET
fcis-27197	94	14	optimization	optimization	NOUN
fcis-27197	94	15	basis	basis	NOUN
fcis-27197	94	16	for	for	ADP
fcis-27197	94	17	subsequent	subsequent	ADJ
fcis-27197	94	18	pruning	pruning	NOUN
fcis-27197	94	19	operations	operation	NOUN
fcis-27197	95	1	[	[	X
fcis-27197	95	2	1,2	1,2	NUM
fcis-27197	95	3	]	]	PUNCT
fcis-27197	95	4	.	.	PUNCT
fcis-27197	96	1	based	base	VERB
fcis-27197	96	2	on	on	ADP
fcis-27197	96	3	the	the	DET
fcis-27197	96	4	quantized	quantize	VERB
fcis-27197	96	5	model	model	NOUN
fcis-27197	96	6	,	,	PUNCT
fcis-27197	96	7	unimportant	unimportant	ADJ
fcis-27197	96	8	convolutional	convolutional	ADJ
fcis-27197	96	9	kernels	kernel	NOUN
fcis-27197	96	10	or	or	CCONJ
fcis-27197	96	11	channels	channel	NOUN
fcis-27197	96	12	are	be	AUX
fcis-27197	96	13	selectively	selectively	ADV
fcis-27197	96	14	removed	remove	VERB
fcis-27197	96	15	based	base	VERB
fcis-27197	96	16	on	on	ADP
fcis-27197	96	17	the	the	DET
fcis-27197	96	18	quantized	quantize	VERB
fcis-27197	96	19	weight	weight	NOUN
fcis-27197	96	20	distribution	distribution	NOUN
fcis-27197	96	21	and	and	CCONJ
fcis-27197	96	22	activation	activation	NOUN
fcis-27197	96	23	value	value	NOUN
fcis-27197	96	24	range	range	NOUN
fcis-27197	97	1	[	[	X
fcis-27197	97	2	1	1	NUM
fcis-27197	97	3	]	]	PUNCT
fcis-27197	97	4	.	.	PUNCT
fcis-27197	98	1	unlike	unlike	ADP
fcis-27197	98	2	traditional	traditional	ADJ
fcis-27197	98	3	pruning	pruning	NOUN
fcis-27197	98	4	,	,	PUNCT
fcis-27197	98	5	the	the	DET
fcis-27197	98	6	weights	weight	NOUN
fcis-27197	98	7	of	of	ADP
fcis-27197	98	8	the	the	DET
fcis-27197	98	9	model	model	NOUN
fcis-27197	98	10	have	have	AUX
fcis-27197	98	11	been	be	AUX
fcis-27197	98	12	quantized	quantize	VERB
fcis-27197	98	13	at	at	ADP
fcis-27197	98	14	this	this	DET
fcis-27197	98	15	point	point	NOUN
fcis-27197	98	16	,	,	PUNCT
fcis-27197	98	17	and	and	CCONJ
fcis-27197	98	18	the	the	DET
fcis-27197	98	19	accuracy	accuracy	NOUN
fcis-27197	98	20	of	of	ADP
fcis-27197	98	21	the	the	DET
fcis-27197	98	22	quantized	quantize	VERB
fcis-27197	98	23	weights	weight	NOUN
fcis-27197	98	24	needs	need	VERB
fcis-27197	98	25	to	to	PART
fcis-27197	98	26	be	be	AUX
fcis-27197	98	27	considered	consider	VERB
fcis-27197	98	28	when	when	SCONJ
fcis-27197	98	29	pruning	prune	VERB
fcis-27197	98	30	.	.	PUNCT
fcis-27197	99	1	the	the	DET
fcis-27197	99	2	pruned	prune	VERB
fcis-27197	99	3	model	model	NOUN
fcis-27197	99	4	needs	need	VERB
fcis-27197	99	5	to	to	PART
fcis-27197	99	6	be	be	AUX
fcis-27197	99	7	retrained	retrain	VERB
fcis-27197	99	8	to	to	PART
fcis-27197	99	9	recover	recover	VERB
fcis-27197	99	10	accuracy	accuracy	NOUN
fcis-27197	99	11	.	.	PUNCT
fcis-27197	100	1	at	at	ADP
fcis-27197	100	2	this	this	DET
fcis-27197	100	3	point	point	NOUN
fcis-27197	100	4	,	,	PUNCT
fcis-27197	100	5	the	the	DET
fcis-27197	100	6	retrained	retrain	VERB
fcis-27197	100	7	model	model	NOUN
fcis-27197	100	8	is	be	AUX
fcis-27197	100	9	already	already	ADV
fcis-27197	100	10	in	in	ADP
fcis-27197	100	11	its	its	PRON
fcis-27197	100	12	quantized	quantize	VERB
fcis-27197	100	13	form	form	NOUN
fcis-27197	100	14	,	,	PUNCT
fcis-27197	100	15	and	and	CCONJ
fcis-27197	100	16	the	the	DET
fcis-27197	100	17	quantization	quantization	NOUN
fcis-27197	100	18	operation	operation	NOUN
fcis-27197	100	19	simulated	simulate	VERB
fcis-27197	100	20	during	during	ADP
fcis-27197	100	21	training	training	NOUN
fcis-27197	100	22	makes	make	VERB
fcis-27197	100	23	the	the	DET
fcis-27197	100	24	model	model	NOUN
fcis-27197	100	25	more	more	ADV
fcis-27197	100	26	robust	robust	ADJ
fcis-27197	100	27	and	and	CCONJ
fcis-27197	100	28	better	well	ADV
fcis-27197	100	29	able	able	ADJ
fcis-27197	100	30	to	to	PART
fcis-27197	100	31	adapt	adapt	VERB
fcis-27197	100	32	to	to	ADP
fcis-27197	100	33	the	the	DET
fcis-27197	100	34	loss	loss	NOUN
fcis-27197	100	35	of	of	ADP
fcis-27197	100	36	accuracy	accuracy	NOUN
fcis-27197	100	37	after	after	ADP
fcis-27197	100	38	quantization	quantization	NOUN
fcis-27197	100	39	[	[	X
fcis-27197	100	40	1,3	1,3	NUM
fcis-27197	100	41	-	-	SYM
fcis-27197	100	42	4	4	NUM
fcis-27197	100	43	]	]	PUNCT
fcis-27197	100	44	.	.	PUNCT
fcis-27197	101	1	7.2	7.2	NUM
fcis-27197	101	2	.	.	PUNCT
fcis-27197	102	1	experimentation	experimentation	NOUN
fcis-27197	102	2	and	and	CCONJ
fcis-27197	102	3	effectiveness	effectiveness	NOUN
fcis-27197	102	4	evaluation	evaluation	NOUN
fcis-27197	102	5	of	of	ADP
fcis-27197	102	6	combination	combination	NOUN
fcis-27197	102	7	strategies	strategy	NOUN
fcis-27197	102	8	experiments	experiment	NOUN
fcis-27197	102	9	are	be	AUX
fcis-27197	102	10	conducted	conduct	VERB
fcis-27197	102	11	using	use	VERB
fcis-27197	102	12	the	the	DET
fcis-27197	102	13	imagenet	imagenet	NOUN
fcis-27197	102	14	dataset	dataset	VERB
fcis-27197	102	15	to	to	PART
fcis-27197	102	16	evaluate	evaluate	VERB
fcis-27197	102	17	the	the	DET
fcis-27197	102	18	impact	impact	NOUN
fcis-27197	102	19	of	of	ADP
fcis-27197	102	20	the	the	DET
fcis-27197	102	21	two	two	NUM
fcis-27197	102	22	combined	combine	VERB
fcis-27197	102	23	strategies	strategy	NOUN
fcis-27197	102	24	on	on	ADP
fcis-27197	102	25	the	the	DET
fcis-27197	102	26	model	model	NOUN
fcis-27197	102	27	performance	performance	NOUN
fcis-27197	102	28	.	.	PUNCT
fcis-27197	103	1	resnet-50	resnet-50	PROPN
fcis-27197	103	2	is	be	AUX
fcis-27197	103	3	chosen	choose	VERB
fcis-27197	103	4	as	as	ADP
fcis-27197	103	5	the	the	DET
fcis-27197	103	6	benchmark	benchmark	NOUN
fcis-27197	103	7	model	model	NOUN
fcis-27197	103	8	,	,	PUNCT
fcis-27197	103	9	and	and	CCONJ
fcis-27197	103	10	the	the	DET
fcis-27197	103	11	original	original	ADJ
fcis-27197	103	12	unpruned	unpruned	NOUN
fcis-27197	103	13	and	and	CCONJ
fcis-27197	103	14	unquantized	unquantized	ADJ
fcis-27197	103	15	model	model	NOUN
fcis-27197	103	16	is	be	AUX
fcis-27197	103	17	first	first	ADV
fcis-27197	103	18	trained	train	VERB
fcis-27197	103	19	and	and	CCONJ
fcis-27197	103	20	its	its	PRON
fcis-27197	103	21	top	top	ADJ
fcis-27197	103	22	1	1	NUM
fcis-27197	103	23	and	and	CCONJ
fcis-27197	103	24	top	top	ADJ
fcis-27197	103	25	5	5	NUM
fcis-27197	103	26	accuracies	accuracy	NOUN
fcis-27197	103	27	are	be	AUX
fcis-27197	103	28	recorded	record	VERB
fcis-27197	103	29	.	.	PUNCT
fcis-27197	104	1	this	this	PRON
fcis-27197	104	2	includes	include	VERB
fcis-27197	104	3	accuracy	accuracy	NOUN
fcis-27197	104	4	loss	loss	NOUN
fcis-27197	104	5	(	(	PUNCT
fcis-27197	104	6	top	top	NOUN
fcis-27197	104	7	1	1	NUM
fcis-27197	104	8	and	and	CCONJ
fcis-27197	104	9	top	top	ADJ
fcis-27197	104	10	5	5	NUM
fcis-27197	104	11	accuracy	accuracy	NOUN
fcis-27197	104	12	)	)	PUNCT
fcis-27197	104	13	,	,	PUNCT
fcis-27197	104	14	model	model	NOUN
fcis-27197	104	15	size	size	NOUN
fcis-27197	104	16	,	,	PUNCT
fcis-27197	104	17	inference	inference	NOUN
fcis-27197	104	18	speed	speed	NOUN
fcis-27197	104	19	,	,	PUNCT
fcis-27197	104	20	and	and	CCONJ
fcis-27197	104	21	memory	memory	NOUN
fcis-27197	104	22	usage	usage	NOUN
fcis-27197	104	23	.	.	PUNCT
fcis-27197	105	1	7.2.1	7.2.1	X
fcis-27197	105	2	.	.	PUNCT
fcis-27197	106	1	effects	effect	NOUN
fcis-27197	106	2	of	of	ADP
fcis-27197	106	3	quantification	quantification	NOUN
fcis-27197	106	4	after	after	ADP
fcis-27197	106	5	pruning	prune	VERB
fcis-27197	106	6	model	model	NOUN
fcis-27197	106	7	size	size	NOUN
fcis-27197	106	8	,	,	PUNCT
fcis-27197	106	9	the	the	DET
fcis-27197	106	10	size	size	NOUN
fcis-27197	106	11	of	of	ADP
fcis-27197	106	12	the	the	DET
fcis-27197	106	13	model	model	NOUN
fcis-27197	106	14	that	that	PRON
fcis-27197	106	15	is	be	AUX
fcis-27197	106	16	pruned	prune	VERB
fcis-27197	106	17	and	and	CCONJ
fcis-27197	106	18	then	then	ADV
fcis-27197	106	19	quantized	quantize	VERB
fcis-27197	106	20	is	be	AUX
fcis-27197	106	21	reduced	reduce	VERB
fcis-27197	106	22	by	by	ADP
fcis-27197	106	23	about	about	ADV
fcis-27197	106	24	75	75	NUM
fcis-27197	106	25	%	%	NOUN
fcis-27197	106	26	,	,	PUNCT
fcis-27197	106	27	from	from	ADP
fcis-27197	106	28	98	98	NUM
fcis-27197	106	29	mb	mb	NOUN
fcis-27197	106	30	in	in	ADP
fcis-27197	106	31	the	the	DET
fcis-27197	106	32	original	original	ADJ
fcis-27197	106	33	model	model	NOUN
fcis-27197	106	34	to	to	ADP
fcis-27197	106	35	about	about	ADP
fcis-27197	106	36	25mb.inference	25mb.inference	NUM
fcis-27197	106	37	speed	speed	NOUN
fcis-27197	106	38	,	,	PUNCT
fcis-27197	106	39	the	the	DET
fcis-27197	106	40	inference	inference	NOUN
fcis-27197	106	41	speed	speed	NOUN
fcis-27197	106	42	of	of	ADP
fcis-27197	106	43	the	the	DET
fcis-27197	106	44	model	model	NOUN
fcis-27197	106	45	that	that	PRON
fcis-27197	106	46	is	be	AUX
fcis-27197	106	47	pruned	prune	VERB
fcis-27197	106	48	and	and	CCONJ
fcis-27197	106	49	then	then	ADV
fcis-27197	106	50	quantized	quantize	VERB
fcis-27197	106	51	is	be	AUX
fcis-27197	106	52	increased	increase	VERB
fcis-27197	106	53	by	by	ADP
fcis-27197	106	54	about	about	ADV
fcis-27197	106	55	60	60	NUM
fcis-27197	106	56	%	%	NOUN
fcis-27197	106	57	,	,	PUNCT
fcis-27197	106	58	which	which	PRON
fcis-27197	106	59	can	can	AUX
fcis-27197	106	60	significantly	significantly	ADV
fcis-27197	106	61	improve	improve	VERB
fcis-27197	106	62	the	the	DET
fcis-27197	106	63	inference	inference	NOUN
fcis-27197	106	64	efficiency	efficiency	NOUN
fcis-27197	106	65	.	.	PUNCT
fcis-27197	107	1	accuracy	accuracy	NOUN
fcis-27197	107	2	,	,	PUNCT
fcis-27197	107	3	in	in	ADP
fcis-27197	107	4	terms	term	NOUN
fcis-27197	107	5	of	of	ADP
fcis-27197	107	6	top	top	ADJ
fcis-27197	107	7	1	1	NUM
fcis-27197	107	8	accuracy	accuracy	NOUN
fcis-27197	107	9	,	,	PUNCT
fcis-27197	107	10	the	the	DET
fcis-27197	107	11	model	model	NOUN
fcis-27197	107	12	accuracy	accuracy	NOUN
fcis-27197	107	13	decreases	decrease	VERB
fcis-27197	107	14	from	from	ADP
fcis-27197	107	15	the	the	DET
fcis-27197	107	16	original	original	ADJ
fcis-27197	107	17	76.4	76.4	NUM
fcis-27197	107	18	%	%	NOUN
fcis-27197	107	19	to	to	ADP
fcis-27197	107	20	74.5	74.5	NUM
fcis-27197	107	21	%	%	NOUN
fcis-27197	107	22	,	,	PUNCT
fcis-27197	107	23	and	and	CCONJ
fcis-27197	107	24	top-5	top-5	VERB
fcis-27197	107	25	accuracy	accuracy	NOUN
fcis-27197	107	26	decreases	decrease	VERB
fcis-27197	107	27	from	from	ADP
fcis-27197	107	28	93.2	93.2	NUM
fcis-27197	107	29	%	%	NOUN
fcis-27197	107	30	to	to	ADP
fcis-27197	107	31	91.0	91.0	NUM
fcis-27197	107	32	%	%	NOUN
fcis-27197	107	33	.	.	PUNCT
fcis-27197	108	1	the	the	DET
fcis-27197	108	2	accuracy	accuracy	NOUN
fcis-27197	108	3	loss	loss	NOUN
fcis-27197	108	4	is	be	AUX
fcis-27197	108	5	relatively	relatively	ADV
fcis-27197	108	6	controllable	controllable	ADJ
fcis-27197	108	7	.	.	PUNCT
fcis-27197	109	1	7.2.2	7.2.2	X
fcis-27197	109	2	.	.	PUNCT
fcis-27197	110	1	quantifying	quantify	VERB
fcis-27197	110	2	the	the	DET
fcis-27197	110	3	effect	effect	NOUN
fcis-27197	110	4	of	of	ADP
fcis-27197	110	5	post	post	NOUN
fcis-27197	110	6	pruning	prune	VERB
fcis-27197	110	7	model	model	NOUN
fcis-27197	110	8	size	size	NOUN
fcis-27197	110	9	,	,	PUNCT
fcis-27197	110	10	quantization	quantization	NOUN
fcis-27197	110	11	followed	follow	VERB
fcis-27197	110	12	by	by	ADP
fcis-27197	110	13	pruning	pruning	NOUN
fcis-27197	110	14	reduces	reduce	VERB
fcis-27197	110	15	the	the	DET
fcis-27197	110	16	model	model	NOUN
fcis-27197	110	17	size	size	NOUN
fcis-27197	110	18	by	by	ADP
fcis-27197	110	19	about	about	ADV
fcis-27197	110	20	70	70	NUM
fcis-27197	110	21	%	%	NOUN
fcis-27197	110	22	,	,	PUNCT
fcis-27197	110	23	from	from	ADP
fcis-27197	110	24	98	98	NUM
fcis-27197	110	25	mb	mb	NOUN
fcis-27197	110	26	in	in	ADP
fcis-27197	110	27	the	the	DET
fcis-27197	110	28	original	original	ADJ
fcis-27197	110	29	model	model	NOUN
fcis-27197	110	30	to	to	ADP
fcis-27197	110	31	about	about	ADP
fcis-27197	110	32	30	30	NUM
fcis-27197	110	33	mb	mb	NOUN
fcis-27197	110	34	.	.	PUNCT
fcis-27197	110	35	inference	inference	NOUN
fcis-27197	110	36	speed	speed	NOUN
fcis-27197	110	37	,	,	PUNCT
fcis-27197	110	38	quantization	quantization	NOUN
fcis-27197	110	39	followed	follow	VERB
fcis-27197	110	40	by	by	ADP
fcis-27197	110	41	pruning	prune	VERB
fcis-27197	110	42	improves	improve	VERB
fcis-27197	110	43	the	the	DET
fcis-27197	110	44	inference	inference	NOUN
fcis-27197	110	45	speed	speed	NOUN
fcis-27197	110	46	of	of	ADP
fcis-27197	110	47	the	the	DET
fcis-27197	110	48	model	model	NOUN
fcis-27197	110	49	by	by	ADP
fcis-27197	110	50	about	about	ADV
fcis-27197	110	51	50	50	NUM
fcis-27197	110	52	%	%	NOUN
fcis-27197	110	53	,	,	PUNCT
fcis-27197	110	54	which	which	PRON
fcis-27197	110	55	is	be	AUX
fcis-27197	110	56	slightly	slightly	ADV
fcis-27197	110	57	less	less	ADV
fcis-27197	110	58	effective	effective	ADJ
fcis-27197	110	59	than	than	ADP
fcis-27197	110	60	the	the	DET
fcis-27197	110	61	quantization	quantization	NOUN
fcis-27197	110	62	strategy	strategy	NOUN
fcis-27197	110	63	after	after	ADP
fcis-27197	110	64	pruning	prune	VERB
fcis-27197	110	65	.	.	PUNCT
fcis-27197	111	1	accuracy	accuracy	NOUN
fcis-27197	111	2	,	,	PUNCT
fcis-27197	111	3	top	top	ADJ
fcis-27197	111	4	1	1	NUM
fcis-27197	111	5	accuracy	accuracy	NOUN
fcis-27197	111	6	decreased	decrease	VERB
fcis-27197	111	7	from	from	ADP
fcis-27197	111	8	76.4	76.4	NUM
fcis-27197	111	9	%	%	NOUN
fcis-27197	111	10	to	to	ADP
fcis-27197	111	11	73.0	73.0	NUM
fcis-27197	111	12	%	%	NOUN
fcis-27197	111	13	and	and	CCONJ
fcis-27197	111	14	top-5	top-5	VERB
fcis-27197	111	15	accuracy	accuracy	NOUN
fcis-27197	111	16	decreased	decrease	VERB
fcis-27197	111	17	from	from	ADP
fcis-27197	111	18	93.2	93.2	NUM
fcis-27197	111	19	%	%	NOUN
fcis-27197	111	20	to	to	PART
fcis-27197	111	21	90.0	90.0	NUM
fcis-27197	111	22	%	%	NOUN
fcis-27197	111	23	.	.	PUNCT
fcis-27197	112	1	the	the	DET
fcis-27197	112	2	loss	loss	NOUN
fcis-27197	112	3	of	of	ADP
fcis-27197	112	4	accuracy	accuracy	NOUN
fcis-27197	112	5	is	be	AUX
fcis-27197	112	6	large	large	ADJ
fcis-27197	112	7	,	,	PUNCT
fcis-27197	112	8	and	and	CCONJ
fcis-27197	112	9	the	the	DET
fcis-27197	112	10	retraining	retrain	VERB
fcis-27197	112	11	process	process	NOUN
fcis-27197	112	12	is	be	AUX
fcis-27197	112	13	difficult	difficult	ADJ
fcis-27197	112	14	to	to	PART
fcis-27197	112	15	fully	fully	ADV
fcis-27197	112	16	recover	recover	VERB
fcis-27197	112	17	.	.	PUNCT
fcis-27197	113	1	7.2.3	7.2.3	NUM
fcis-27197	113	2	.	.	PUNCT
fcis-27197	113	3	optimization	optimization	NOUN
fcis-27197	113	4	strategies	strategy	NOUN
fcis-27197	113	5	and	and	CCONJ
fcis-27197	113	6	future	future	ADJ
fcis-27197	113	7	research	research	NOUN
fcis-27197	113	8	directions	direction	NOUN
fcis-27197	113	9	a	a	DET
fcis-27197	113	10	joint	joint	ADJ
fcis-27197	113	11	optimization	optimization	NOUN
fcis-27197	113	12	strategy	strategy	NOUN
fcis-27197	113	13	is	be	AUX
fcis-27197	113	14	used	use	VERB
fcis-27197	113	15	to	to	PART
fcis-27197	113	16	consider	consider	VERB
fcis-27197	113	17	quantization	quantization	NOUN
fcis-27197	113	18	and	and	CCONJ
fcis-27197	113	19	pruning	prune	VERB
fcis-27197	113	20	simultaneously	simultaneously	ADV
fcis-27197	113	21	,	,	PUNCT
fcis-27197	113	22	design	design	VERB
fcis-27197	113	23	a	a	DET
fcis-27197	113	24	joint	joint	ADJ
fcis-27197	113	25	loss	loss	NOUN
fcis-27197	113	26	function	function	NOUN
fcis-27197	113	27	,	,	PUNCT
fcis-27197	113	28	and	and	CCONJ
fcis-27197	113	29	dynamically	dynamically	ADV
fcis-27197	113	30	weigh	weigh	VERB
fcis-27197	113	31	the	the	DET
fcis-27197	113	32	effects	effect	NOUN
fcis-27197	113	33	of	of	ADP
fcis-27197	113	34	both	both	PRON
fcis-27197	113	35	during	during	ADP
fcis-27197	113	36	the	the	DET
fcis-27197	113	37	training	training	NOUN
fcis-27197	113	38	process	process	NOUN
fcis-27197	113	39	to	to	PART
fcis-27197	113	40	maximize	maximize	VERB
fcis-27197	113	41	the	the	DET
fcis-27197	113	42	model	model	NOUN
fcis-27197	113	43	compression	compression	NOUN
fcis-27197	113	44	rate	rate	NOUN
fcis-27197	113	45	while	while	SCONJ
fcis-27197	113	46	minimizing	minimize	VERB
fcis-27197	113	47	the	the	DET
fcis-27197	113	48	accuracy	accuracy	NOUN
fcis-27197	113	49	loss	loss	NOUN
fcis-27197	113	50	.	.	PUNCT
fcis-27197	114	1	explore	explore	VERB
fcis-27197	114	2	the	the	DET
fcis-27197	114	3	joint	joint	ADJ
fcis-27197	114	4	optimization	optimization	NOUN
fcis-27197	114	5	strategy	strategy	NOUN
fcis-27197	114	6	of	of	ADP
fcis-27197	114	7	quantization	quantization	NOUN
fcis-27197	114	8	and	and	CCONJ
fcis-27197	114	9	pruning	prune	VERB
fcis-27197	114	10	in	in	ADP
fcis-27197	114	11	multi	multi	ADJ
fcis-27197	114	12	-	-	NOUN
fcis-27197	114	13	task	task	ADJ
fcis-27197	114	14	learning	learning	NOUN
fcis-27197	114	15	for	for	ADP
fcis-27197	114	16	efficient	efficient	ADJ
fcis-27197	114	17	compression	compression	NOUN
fcis-27197	114	18	and	and	CCONJ
fcis-27197	114	19	acceleration	acceleration	NOUN
fcis-27197	114	20	of	of	ADP
fcis-27197	114	21	multi	multi	ADJ
fcis-27197	114	22	-	-	ADJ
fcis-27197	114	23	task	task	NOUN
fcis-27197	114	24	models	model	NOUN
fcis-27197	114	25	.	.	PUNCT
fcis-27197	115	1	design	design	VERB
fcis-27197	115	2	adaptive	adaptive	ADJ
fcis-27197	115	3	quantization	quantization	NOUN
fcis-27197	115	4	and	and	CCONJ
fcis-27197	115	5	pruning	prune	VERB
fcis-27197	115	6	strategies	strategy	NOUN
fcis-27197	115	7	based	base	VERB
fcis-27197	115	8	on	on	ADP
fcis-27197	115	9	the	the	DET
fcis-27197	115	10	characteristics	characteristic	NOUN
fcis-27197	115	11	of	of	ADP
fcis-27197	115	12	different	different	ADJ
fcis-27197	115	13	hardware	hardware	NOUN
fcis-27197	115	14	platforms	platform	NOUN
fcis-27197	115	15	.	.	PUNCT
fcis-27197	116	1	8	8	X
fcis-27197	116	2	.	.	X
fcis-27197	116	3	conclusion	conclusion	NOUN
fcis-27197	116	4	in	in	ADP
fcis-27197	116	5	this	this	DET
fcis-27197	116	6	study	study	NOUN
fcis-27197	116	7	,	,	PUNCT
fcis-27197	116	8	i	i	PRON
fcis-27197	116	9	explore	explore	VERB
fcis-27197	116	10	the	the	DET
fcis-27197	116	11	combined	combine	VERB
fcis-27197	116	12	use	use	NOUN
fcis-27197	116	13	of	of	ADP
fcis-27197	116	14	fixed	fix	VERB
fcis-27197	116	15	-	-	PUNCT
fcis-27197	116	16	point	point	NOUN
fcis-27197	116	17	quantisation	quantisation	NOUN
fcis-27197	116	18	and	and	CCONJ
fcis-27197	116	19	structured	structure	VERB
fcis-27197	116	20	pruning	pruning	NOUN
fcis-27197	116	21	techniques	technique	NOUN
fcis-27197	116	22	to	to	PART
fcis-27197	116	23	optimize	optimize	VERB
fcis-27197	116	24	the	the	DET
fcis-27197	116	25	performance	performance	NOUN
fcis-27197	116	26	and	and	CCONJ
fcis-27197	116	27	efficiency	efficiency	NOUN
fcis-27197	116	28	of	of	ADP
fcis-27197	116	29	convolutional	convolutional	ADJ
fcis-27197	116	30	neural	neural	ADJ
fcis-27197	116	31	networks	network	NOUN
fcis-27197	116	32	(	(	PUNCT
fcis-27197	116	33	cnns	cnns	PROPN
fcis-27197	116	34	)	)	PUNCT
fcis-27197	116	35	in	in	ADP
fcis-27197	116	36	image	image	NOUN
fcis-27197	116	37	classification	classification	NOUN
fcis-27197	116	38	tasks	task	NOUN
fcis-27197	116	39	.	.	PUNCT
fcis-27197	117	1	by	by	ADP
fcis-27197	117	2	reducing	reduce	VERB
fcis-27197	117	3	model	model	NOUN
fcis-27197	117	4	size	size	NOUN
fcis-27197	117	5	and	and	CCONJ
fcis-27197	117	6	computational	computational	ADJ
fcis-27197	117	7	complexity	complexity	NOUN
fcis-27197	117	8	,	,	PUNCT
fcis-27197	117	9	these	these	PRON
fcis-27197	117	10	approaches	approach	VERB
fcis-27197	117	11	42	42	NUM
fcis-27197	117	12	make	make	VERB
fcis-27197	117	13	cnns	cnn	NOUN
fcis-27197	117	14	more	more	ADV
fcis-27197	117	15	suitable	suitable	ADJ
fcis-27197	117	16	for	for	ADP
fcis-27197	117	17	deployment	deployment	NOUN
fcis-27197	117	18	in	in	ADP
fcis-27197	117	19	resourceconstrained	resourceconstraine	VERB
fcis-27197	117	20	environments	environment	NOUN
fcis-27197	117	21	such	such	ADJ
fcis-27197	117	22	as	as	ADP
fcis-27197	117	23	mobile	mobile	ADJ
fcis-27197	117	24	devices	device	NOUN
fcis-27197	117	25	and	and	CCONJ
fcis-27197	117	26	embedded	embed	VERB
fcis-27197	117	27	systems	system	NOUN
fcis-27197	117	28	.	.	PUNCT
fcis-27197	118	1	experimental	experimental	ADJ
fcis-27197	118	2	results	result	NOUN
fcis-27197	118	3	using	use	VERB
fcis-27197	118	4	resnet-50	resnet-50	PROPN
fcis-27197	118	5	on	on	ADP
fcis-27197	118	6	the	the	DET
fcis-27197	118	7	imagenet	imagenet	NOUN
fcis-27197	118	8	dataset	dataset	NOUN
fcis-27197	118	9	show	show	NOUN
fcis-27197	118	10	that	that	SCONJ
fcis-27197	118	11	the	the	DET
fcis-27197	118	12	combination	combination	NOUN
fcis-27197	118	13	of	of	ADP
fcis-27197	118	14	quantisation	quantisation	NOUN
fcis-27197	118	15	and	and	CCONJ
fcis-27197	118	16	pruning	pruning	NOUN
fcis-27197	118	17	can	can	AUX
fcis-27197	118	18	reduce	reduce	VERB
fcis-27197	118	19	model	model	NOUN
fcis-27197	118	20	size	size	NOUN
fcis-27197	118	21	by	by	ADP
fcis-27197	118	22	up	up	ADP
fcis-27197	118	23	to	to	PART
fcis-27197	118	24	75	75	NUM
fcis-27197	118	25	%	%	NOUN
fcis-27197	118	26	and	and	CCONJ
fcis-27197	118	27	increase	increase	VERB
fcis-27197	118	28	inference	inference	NOUN
fcis-27197	118	29	speed	speed	NOUN
fcis-27197	118	30	by	by	ADP
fcis-27197	118	31	50	50	NUM
fcis-27197	118	32	%	%	NOUN
fcis-27197	118	33	,	,	PUNCT
fcis-27197	118	34	while	while	SCONJ
fcis-27197	118	35	maintaining	maintain	VERB
fcis-27197	118	36	an	an	DET
fcis-27197	118	37	acceptable	acceptable	ADJ
fcis-27197	118	38	classification	classification	NOUN
fcis-27197	118	39	accuracy	accuracy	NOUN
fcis-27197	118	40	of	of	ADP
fcis-27197	118	41	74.5	74.5	NUM
fcis-27197	118	42	%	%	NOUN
fcis-27197	118	43	,	,	PUNCT
fcis-27197	118	44	compared	compare	VERB
fcis-27197	118	45	to	to	ADP
fcis-27197	118	46	76.4	76.4	NUM
fcis-27197	118	47	%	%	NOUN
fcis-27197	118	48	for	for	ADP
fcis-27197	118	49	the	the	DET
fcis-27197	118	50	baseline	baseline	PROPN
fcis-27197	118	51	model	model	NOUN
fcis-27197	118	52	.	.	PUNCT
fcis-27197	119	1	while	while	SCONJ
fcis-27197	119	2	the	the	DET
fcis-27197	119	3	accuracy	accuracy	NOUN
fcis-27197	119	4	drops	drop	VERB
fcis-27197	119	5	slightly	slightly	ADV
fcis-27197	119	6	,	,	PUNCT
fcis-27197	119	7	the	the	DET
fcis-27197	119	8	increased	increase	VERB
fcis-27197	119	9	efficiency	efficiency	NOUN
fcis-27197	119	10	makes	make	VERB
fcis-27197	119	11	this	this	DET
fcis-27197	119	12	approach	approach	NOUN
fcis-27197	119	13	ideal	ideal	NOUN
fcis-27197	119	14	for	for	ADP
fcis-27197	119	15	real	real	ADJ
fcis-27197	119	16	-	-	PUNCT
fcis-27197	119	17	time	time	NOUN
fcis-27197	119	18	applications	application	NOUN
fcis-27197	119	19	.	.	PUNCT
fcis-27197	120	1	future	future	ADJ
fcis-27197	120	2	work	work	NOUN
fcis-27197	120	3	will	will	AUX
fcis-27197	120	4	further	far	ADV
fcis-27197	120	5	focus	focus	VERB
fcis-27197	120	6	on	on	ADP
fcis-27197	120	7	improving	improve	VERB
fcis-27197	120	8	the	the	DET
fcis-27197	120	9	combined	combined	ADJ
fcis-27197	120	10	quantisation	quantisation	NOUN
fcis-27197	120	11	and	and	CCONJ
fcis-27197	120	12	pruning	pruning	NOUN
fcis-27197	120	13	strategies	strategy	NOUN
fcis-27197	120	14	,	,	PUNCT
fcis-27197	120	15	as	as	ADV
fcis-27197	120	16	well	well	ADV
fcis-27197	120	17	as	as	ADP
fcis-27197	120	18	exploring	explore	VERB
fcis-27197	120	19	their	their	PRON
fcis-27197	120	20	applicability	applicability	NOUN
fcis-27197	120	21	to	to	ADP
fcis-27197	120	22	other	other	ADJ
fcis-27197	120	23	cnn	cnn	PROPN
fcis-27197	120	24	architectures	architecture	NOUN
fcis-27197	120	25	and	and	CCONJ
fcis-27197	120	26	tasks	task	NOUN
fcis-27197	120	27	.	.	PUNCT
fcis-27197	121	1	references	reference	NOUN
fcis-27197	121	2	[	[	X
fcis-27197	121	3	1	1	NUM
fcis-27197	121	4	]	]	PUNCT
fcis-27197	121	5	azarpeyvand	azarpeyvand	PROPN
fcis-27197	121	6	,	,	PUNCT
fcis-27197	121	7	a.	a.	PROPN
fcis-27197	121	8	,	,	PUNCT
fcis-27197	121	9	rokh	rokh	PROPN
fcis-27197	121	10	,	,	PUNCT
fcis-27197	121	11	b.	b.	PROPN
fcis-27197	121	12	,	,	PUNCT
fcis-27197	121	13	&	&	CCONJ
fcis-27197	121	14	naderi	naderi	PROPN
fcis-27197	121	15	,	,	PUNCT
fcis-27197	121	16	s.	s.	PROPN
fcis-27197	121	17	(	(	PUNCT
fcis-27197	121	18	2023	2023	NUM
fcis-27197	121	19	)	)	PUNCT
fcis-27197	121	20	.	.	PUNCT
fcis-27197	122	1	a	a	DET
fcis-27197	122	2	comprehensive	comprehensive	ADJ
fcis-27197	122	3	survey	survey	NOUN
fcis-27197	122	4	on	on	ADP
fcis-27197	122	5	model	model	NOUN
fcis-27197	122	6	quantization	quantization	NOUN
fcis-27197	122	7	for	for	ADP
fcis-27197	122	8	deep	deep	ADJ
fcis-27197	122	9	neural	neural	ADJ
fcis-27197	122	10	networks	network	NOUN
fcis-27197	122	11	in	in	ADP
fcis-27197	122	12	image	image	NOUN
fcis-27197	122	13	classification	classification	NOUN
fcis-27197	122	14	.	.	PUNCT
fcis-27197	123	1	acm	acm	NOUN
fcis-27197	123	2	transactions	transaction	NOUN
fcis-27197	123	3	on	on	ADP
fcis-27197	123	4	intelligent	intelligent	ADJ
fcis-27197	123	5	systems	system	NOUN
fcis-27197	123	6	and	and	CCONJ
fcis-27197	123	7	technology	technology	NOUN
fcis-27197	123	8	,	,	PUNCT
fcis-27197	123	9	14(6	14(6	NOUN
fcis-27197	123	10	)	)	PUNCT
fcis-27197	123	11	,	,	PUNCT
fcis-27197	123	12	1	1	NUM
fcis-27197	123	13	-	-	SYM
fcis-27197	123	14	50	50	NUM
fcis-27197	123	15	.	.	PUNCT
fcis-27197	124	1	[	[	X
fcis-27197	124	2	2	2	NUM
fcis-27197	124	3	]	]	X
fcis-27197	124	4	cheng	cheng	PROPN
fcis-27197	124	5	,	,	PUNCT
fcis-27197	124	6	y.	y.	PROPN
fcis-27197	124	7	,	,	PUNCT
fcis-27197	124	8	wang	wang	PROPN
fcis-27197	124	9	,	,	PUNCT
fcis-27197	124	10	x.	x.	PROPN
fcis-27197	124	11	,	,	PUNCT
fcis-27197	124	12	xie	xie	PROPN
fcis-27197	124	13	,	,	PUNCT
fcis-27197	124	14	x.	x.	PROPN
fcis-27197	124	15	,	,	PUNCT
fcis-27197	124	16	li	li	PROPN
fcis-27197	124	17	,	,	PUNCT
fcis-27197	124	18	w.	w.	PROPN
fcis-27197	124	19	,	,	PUNCT
fcis-27197	124	20	&	&	CCONJ
fcis-27197	124	21	peng	peng	PROPN
fcis-27197	124	22	,	,	PUNCT
fcis-27197	124	23	s.	s.	PROPN
fcis-27197	124	24	(	(	PUNCT
fcis-27197	124	25	2022	2022	NUM
fcis-27197	124	26	)	)	PUNCT
fcis-27197	124	27	.	.	PUNCT
fcis-27197	125	1	channel	channel	NOUN
fcis-27197	125	2	pruning	pruning	NOUN
fcis-27197	125	3	guided	guide	VERB
fcis-27197	125	4	by	by	ADP
fcis-27197	125	5	global	global	ADJ
fcis-27197	125	6	channel	channel	PROPN
fcis-27197	125	7	relation	relation	PROPN
fcis-27197	125	8	.	.	PUNCT
fcis-27197	126	1	applied	apply	VERB
fcis-27197	126	2	intelligence	intelligence	NOUN
fcis-27197	126	3	,	,	PUNCT
fcis-27197	126	4	42(5	42(5	NUM
fcis-27197	126	5	)	)	PUNCT
fcis-27197	126	6	,	,	PUNCT
fcis-27197	126	7	1234	1234	NUM
fcis-27197	126	8	-	-	SYM
fcis-27197	126	9	1250	1250	NUM
fcis-27197	126	10	.	.	PUNCT
fcis-27197	127	1	[	[	X
fcis-27197	127	2	3	3	NUM
fcis-27197	127	3	]	]	X
fcis-27197	127	4	dai	dai	PROPN
fcis-27197	127	5	,	,	PUNCT
fcis-27197	127	6	b.	b.	PROPN
fcis-27197	127	7	,	,	PUNCT
fcis-27197	127	8	zhu	zhu	PROPN
fcis-27197	127	9	,	,	PUNCT
fcis-27197	127	10	c.	c.	PROPN
fcis-27197	127	11	,	,	PUNCT
fcis-27197	127	12	guo	guo	PROPN
fcis-27197	127	13	,	,	PUNCT
fcis-27197	127	14	b.	b.	PROPN
fcis-27197	127	15	,	,	PUNCT
fcis-27197	127	16	&	&	CCONJ
fcis-27197	127	17	wipf	wipf	PROPN
fcis-27197	127	18	,	,	PUNCT
fcis-27197	127	19	d.p	d.p	PROPN
fcis-27197	127	20	.	.	PROPN
fcis-27197	128	1	(	(	PUNCT
fcis-27197	128	2	2018	2018	NUM
fcis-27197	128	3	)	)	PUNCT
fcis-27197	128	4	an	an	DET
fcis-27197	128	5	adaptive	adaptive	ADJ
fcis-27197	128	6	joint	joint	ADJ
fcis-27197	128	7	optimization	optimization	NOUN
fcis-27197	128	8	framework	framework	NOUN
fcis-27197	128	9	for	for	ADP
fcis-27197	128	10	pruning	prune	VERB
fcis-27197	128	11	and	and	CCONJ
fcis-27197	128	12	quantization	quantization	NOUN
fcis-27197	128	13	.	.	PUNCT
fcis-27197	129	1	in	in	ADP
fcis-27197	129	2	:	:	PUNCT
fcis-27197	129	3	35th	35th	ADJ
fcis-27197	129	4	international	international	ADJ
fcis-27197	129	5	conference	conference	NOUN
fcis-27197	129	6	on	on	ADP
fcis-27197	129	7	machine	machine	NOUN
fcis-27197	129	8	learning	learning	NOUN
fcis-27197	129	9	.	.	PUNCT
fcis-27197	130	1	stockholm	stockholm	PROPN
fcis-27197	130	2	.	.	PUNCT
fcis-27197	131	1	pp	pp	ADJ
fcis-27197	131	2	.	.	PUNCT
fcis-27197	132	1	1143	1143	NUM
fcis-27197	132	2	-	-	SYM
fcis-27197	132	3	1152	1152	NUM
fcis-27197	132	4	.	.	PUNCT
fcis-27197	133	1	[	[	X
fcis-27197	133	2	4	4	NUM
fcis-27197	133	3	]	]	SYM
fcis-27197	133	4	li	li	PROPN
fcis-27197	133	5	,	,	PUNCT
fcis-27197	133	6	z.	z.	PROPN
fcis-27197	133	7	,	,	PUNCT
fcis-27197	133	8	li	li	PROPN
fcis-27197	133	9	,	,	PUNCT
fcis-27197	133	10	h.	h.	PROPN
fcis-27197	133	11	,	,	PUNCT
fcis-27197	133	12	&	&	CCONJ
fcis-27197	133	13	meng	meng	PROPN
fcis-27197	133	14	,	,	PUNCT
fcis-27197	133	15	l.	l.	PROPN
fcis-27197	133	16	(	(	PUNCT
fcis-27197	133	17	2023	2023	NUM
fcis-27197	133	18	)	)	PUNCT
fcis-27197	133	19	.	.	PUNCT
fcis-27197	134	1	a	a	DET
fcis-27197	134	2	survey	survey	NOUN
fcis-27197	134	3	on	on	ADP
fcis-27197	134	4	deep	deep	ADJ
fcis-27197	134	5	neural	neural	ADJ
fcis-27197	134	6	network	network	NOUN
fcis-27197	134	7	pruning	pruning	NOUN
fcis-27197	134	8	.	.	PUNCT
fcis-27197	135	1	computers	computer	NOUN
fcis-27197	135	2	,	,	PUNCT
fcis-27197	135	3	12(3	12(3	NUM
fcis-27197	135	4	)	)	PUNCT
fcis-27197	135	5	,	,	PUNCT
fcis-27197	135	6	60	60	NUM
fcis-27197	135	7	.	.	PUNCT
fcis-27197	136	1	[	[	X
fcis-27197	136	2	5	5	NUM
fcis-27197	136	3	]	]	PUNCT
fcis-27197	136	4	zacchigna	zacchigna	PROPN
fcis-27197	136	5	,	,	PUNCT
fcis-27197	136	6	g.	g.	PROPN
fcis-27197	136	7	f.	f.	PROPN
fcis-27197	136	8	,	,	PUNCT
fcis-27197	136	9	lew	lew	PROPN
fcis-27197	136	10	,	,	PUNCT
fcis-27197	136	11	s.	s.	PROPN
fcis-27197	136	12	,	,	PUNCT
fcis-27197	136	13	&	&	CCONJ
fcis-27197	136	14	lutenberg	lutenberg	PROPN
fcis-27197	136	15	,	,	PUNCT
fcis-27197	136	16	a.	a.	NOUN
fcis-27197	136	17	(	(	PUNCT
fcis-27197	136	18	2024	2024	NUM
fcis-27197	136	19	)	)	PUNCT
fcis-27197	136	20	.	.	PUNCT
fcis-27197	137	1	flexible	flexible	ADJ
fcis-27197	137	2	quantization	quantization	NOUN
fcis-27197	137	3	for	for	ADP
fcis-27197	137	4	efficient	efficient	ADJ
fcis-27197	137	5	convolutional	convolutional	ADJ
fcis-27197	137	6	neural	neural	ADJ
fcis-27197	137	7	networks	network	NOUN
fcis-27197	137	8	.	.	PUNCT
fcis-27197	138	1	electronics	electronic	NOUN
fcis-27197	138	2	,	,	PUNCT
fcis-27197	138	3	13(10	13(10	NUM
fcis-27197	138	4	)	)	PUNCT
fcis-27197	138	5	,	,	PUNCT
fcis-27197	138	6	1923	1923	NUM
fcis-27197	138	7	.	.	PUNCT