v0.1-attn-weights

版本发布时间: 2021-09-04 08:25:26

huggingface/pytorch-image-models最新发布版本:v1.0.9(2024-08-24 07:42:07)

A collection of weights I've trained comparing various types of SE-like (SE, ECA, GC, etc), self-attention (bottleneck, halo, lambda) blocks, and related non-attn baselines.

ResNet-26-T series

[2, 2, 2, 2] repeat Bottlneck block ResNet architecture
ReLU activations
3 layer stem with 24, 32, 64 chs, max-pool
avg pool in shortcut downsample
self-attn blocks replace 3x3 in both blocks for last stage, and second block of penultimate stage

model	top1	top1_err	top5	top5_err	param_count	img_size	cropt_pct	interpolation
botnet26t_256	79.246	20.754	94.53	5.47	12.49	256	0.95	bicubic
halonet26t	79.13	20.87	94.314	5.686	12.48	256	0.95	bicubic
lambda_resnet26t	79.112	20.888	94.59	5.41	10.96	256	0.94	bicubic
lambda_resnet26rpt_256	78.964	21.036	94.428	5.572	10.99	256	0.94	bicubic
resnet26t	77.872	22.128	93.834	6.166	16.01	256	0.94	bicubic

Details:

HaloNet - 8 pixel block size, 2 pixel halo (overlap), relative position embedding
BotNet - relative position embedding
Lambda-ResNet-26-T - 3d lambda conv, kernel = 9
Lambda-ResNet-26-RPT - relative position embedding

Benchmark - RTX 3090 - AMP - NCHW - NGC 21.09

model	infer_samples_per_sec	infer_step_time	infer_batch_size	infer_img_size	train_samples_per_sec	train_step_time	train_batch_size	train_img_size	param_count
resnet26t	2967.55	86.252	256	256	857.62	297.984	256	256	16.01
botnet26t_256	2642.08	96.879	256	256	809.41	315.706	256	256	12.49
halonet26t	2601.91	98.375	256	256	783.92	325.976	256	256	12.48
lambda_resnet26t	2354.1	108.732	256	256	697.28	366.521	256	256	10.96
lambda_resnet26rpt_256	1847.34	138.563	256	256	644.84	197.892	128	256	10.99

Benchmark - RTX 3090 - AMP - NHWC - NGC 21.09

model	infer_samples_per_sec	infer_step_time	infer_batch_size	infer_img_size	train_samples_per_sec	train_step_time	train_batch_size	train_img_size	param_count
resnet26t	3691.94	69.327	256	256	1188.17	214.96	256	256	16.01
botnet26t_256	3291.63	77.76	256	256	1126.68	226.653	256	256	12.49
halonet26t	3230.5	79.232	256	256	1077.82	236.934	256	256	12.48
lambda_resnet26rpt_256	2324.15	110.133	256	256	864.42	147.485	128	256	10.99
lambda_resnet26t	Not Supported

ResNeXT-26-T series

[2, 2, 2, 2] repeat Bottlneck block ResNeXt architectures
SiLU activations
grouped 3x3 convolutions in bottleneck, 32 channels per group
3 layer stem with 24, 32, 64 chs, max-pool
avg pool in shortcut downsample
channel attn (active in non self-attn blocks) between 3x3 and last 1x1 conv
when active, self-attn blocks replace 3x3 conv in both blocks for last stage, and second block of penultimate stage

model	top1	top1_err	top5	top5_err	param_count	img_size	cropt_pct	interpolation
eca_halonext26ts	79.484	20.516	94.600	5.400	10.76	256	0.94	bicubic
eca_botnext26ts_256	79.270	20.730	94.594	5.406	10.59	256	0.95	bicubic
bat_resnext26ts	78.268	21.732	94.1	5.9	10.73	256	0.9	bicubic
seresnext26ts	77.852	22.148	93.784	6.216	10.39	256	0.9	bicubic
gcresnext26ts	77.804	22.196	93.824	6.176	10.48	256	0.9	bicubic
eca_resnext26ts	77.446	22.554	93.57	6.43	10.3	256	0.9	bicubic
resnext26ts	76.764	23.236	93.136	6.864	10.3	256	0.9	bicubic

Benchmark - RTX 3090 - AMP - NCHW - NGC 21.09

model	infer_samples_per_sec	infer_step_time	infer_batch_size	infer_img_size	train_samples_per_sec	train_step_time	train_batch_size	train_img_size	param_count
resnext26ts	3006.57	85.134	256	256	864.4	295.646	256	256	10.3
seresnext26ts	2931.27	87.321	256	256	836.92	305.193	256	256	10.39
eca_resnext26ts	2925.47	87.495	256	256	837.78	305.003	256	256	10.3
gcresnext26ts	2870.01	89.186	256	256	818.35	311.97	256	256	10.48
eca_botnext26ts_256	2652.03	96.513	256	256	790.43	323.257	256	256	10.59
eca_halonext26ts	2593.03	98.705	256	256	766.07	333.541	256	256	10.76
bat_resnext26ts	2469.78	103.64	256	256	697.21	365.964	256	256	10.73

Benchmark - RTX 3090 - AMP - NHWC - NGC 21.09

NOTE: there are performance issues with certain grouped conv configs with channels last layout, backwards pass in particular is really slow. Also causing issues for RegNet and NFNet networks.

model	infer_samples_per_sec	infer_step_time	infer_batch_size	infer_img_size	train_samples_per_sec	train_step_time	train_batch_size	train_img_size	param_count
resnext26ts	3952.37	64.755	256	256	608.67	420.049	256	256	10.3
eca_resnext26ts	3815.77	67.074	256	256	594.35	430.146	256	256	10.3
seresnext26ts	3802.75	67.304	256	256	592.82	431.14	256	256	10.39
gcresnext26ts	3626.97	70.57	256	256	581.83	439.119	256	256	10.48
eca_botnext26ts_256	3515.84	72.8	256	256	611.71	417.862	256	256	10.59
eca_halonext26ts	3410.12	75.057	256	256	597.52	427.789	256	256	10.76
bat_resnext26ts	3053.83	83.811	256	256	533.23	478.839	256	256	10.73

ResNet-33-T series.

[2, 3, 3, 2] repeat Bottlneck block ResNet architecture
SiLU activations
3 layer stem with 24, 32, 64 chs, no max-pool, 1st and 3rd conv stride 2
avg pool in shortcut downsample
channel attn (active in non self-attn blocks) between 3x3 and last 1x1 conv
when active, self-attn blocks replace 3x3 conv last block of stage 2 and 3, and both blocks of final stage
FC 1x1 conv between last block and classifier

The 33-layer models have an extra 1x1 FC layer between last conv block and classifier. There is both a non-attenion 33 layer baseline and a 32 layer without the extra FC.

model	top1	top1_err	top5	top5_err	param_count	img_size	cropt_pct	interpolation
sehalonet33ts	80.986	19.014	95.272	4.728	13.69	256	0.94	bicubic
seresnet33ts	80.388	19.612	95.108	4.892	19.78	256	0.94	bicubic
eca_resnet33ts	80.132	19.868	95.054	4.946	19.68	256	0.94	bicubic
gcresnet33ts	79.99	20.01	94.988	5.012	19.88	256	0.94	bicubic
resnet33ts	79.352	20.648	94.596	5.404	19.68	256	0.94	bicubic
resnet32ts	79.028	20.972	94.444	5.556	17.96	256	0.94	bicubic

Benchmark - RTX 3090 - AMP - NCHW - NGC 21.09

model	infer_samples_per_sec	infer_step_time	infer_batch_size	infer_img_size	train_samples_per_sec	train_step_time	train_batch_size	train_img_size	param_count
resnet32ts	2502.96	102.266	256	256	733.27	348.507	256	256	17.96
resnet33ts	2473.92	103.466	256	256	725.34	352.309	256	256	19.68
seresnet33ts	2400.18	106.646	256	256	695.19	367.413	256	256	19.78
eca_resnet33ts	2394.77	106.886	256	256	696.93	366.637	256	256	19.68
gcresnet33ts	2342.81	109.257	256	256	678.22	376.404	256	256	19.88
sehalonet33ts	1857.65	137.794	256	256	577.34	442.545	256	256	13.69

Benchmark - RTX 3090 - AMP - NHWC - NGC 21.09

model	infer_samples_per_sec	infer_step_time	infer_batch_size	infer_img_size	train_samples_per_sec	train_step_time	train_batch_size	train_img_size	param_count
resnet32ts	3306.22	77.416	256	256	1012.82	252.158	256	256	17.96
resnet33ts	3257.59	78.573	256	256	1002.38	254.778	256	256	19.68
seresnet33ts	3128.08	81.826	256	256	950.27	268.581	256	256	19.78
eca_resnet33ts	3127.11	81.852	256	256	948.84	269.123	256	256	19.68
gcresnet33ts	2984.87	85.753	256	256	916.98	278.169	256	256	19.88
sehalonet33ts	2188.23	116.975	256	256	711.63	179.03	128	256	13.69

ResNet-50(ish) models

In Progress

RegNet"Z" series

RegNetZ inspired architecture, inverted bottleneck, SE attention, pre-classifier FC, essentially an EfficientNet w/ grouped conv instead of depthwise
b, c, and d are three different sizes I put together to cover differing flop ranges, not based on the paper (https://arxiv.org/abs/2103.06877) or a search process
for comparison to RegNetY and paper RegNetZ models, at 224x224 b,c, and d models are 1.45, 1.92, and 4.58 GMACs respectively, b, and c are trained at 256 here so higher than that (see tables)
haloregnetz_c uses halo attention for all of last stage, and interleaved every 3 (for 4) of penultimate stage
b, c variants use a stem / 1st stage like the paper, d uses a 3-deep tiered stem with 2-1-2 striding

ImageNet-1k validation at train resolution

model	top1	top1_err	top5	top5_err	param_count	img_size	cropt_pct	interpolation
regnetz_d	83.422	16.578	96.636	3.364	27.58	256	0.95	bicubic
regnetz_c	82.164	17.836	96.058	3.942	13.46	256	0.94	bicubic
haloregnetz_b	81.058	18.942	95.2	4.8	11.68	224	0.94	bicubic
regnetz_b	79.868	20.132	94.988	5.012	9.72	224	0.94	bicubic

ImageNet-1k validation at optimal test res

model	top1	top1_err	top5	top5_err	param_count	img_size	cropt_pct	interpolation
regnetz_d	84.04	15.96	96.87	3.13	27.58	320	0.95	bicubic
regnetz_c	82.516	17.484	96.356	3.644	13.46	320	0.94	bicubic
haloregnetz_b	81.058	18.942	95.2	4.8	11.68	224	0.94	bicubic
regnetz_b	80.728	19.272	95.47	4.53	9.72	288	0.94	bicubic

Benchmark - RTX 3090 - AMP - NCHW - NGC 21.09

model	infer_samples_per_sec	infer_step_time	infer_batch_size	infer_img_size	infer_GMACs	train_samples_per_sec	train_step_time	train_batch_size	train_img_size	param_count
regnetz_b	2703.42	94.68	256	224	1.45	764.85	333.348	256	224	9.72
haloregnetz_b	2086.22	122.695	256	224	1.88	620.1	411.415	256	224	11.68
regnetz_c	1653.19	154.836	256	256	2.51	459.41	277.268	128	256	13.46
regnetz_d	1060.91	241.284	256	256	5.98	296.51	430.143	128	256	27.58

Benchmark - RTX 3090 - AMP - NHWC - NGC 21.09

NOTE: channels last layout is painfully slow for backward pass here due to some sort of cuDNN issue

model	infer_samples_per_sec	infer_step_time	infer_batch_size	infer_img_size	infer_GMACs	train_samples_per_sec	train_step_time	train_batch_size	train_img_size	param_count
regnetz_b	4152.59	61.634	256	224	1.45	399.37	639.572	256	224	9.72
haloregnetz_b	2770.78	92.378	256	224	1.88	364.22	701.386	256	224	11.68
regnetz_c	2512.4	101.878	256	256	2.51	376.72	338.372	128	256	13.46
regnetz_d	1456.05	175.8	256	256	5.98	111.32	1148.279	128	256	27.58