AAAI-challenge

[TOC]

The content of training for the contest

For the .onnx files are bigger than the upload limit, I upload the files to Beihang Cloud. Links

The structure of training is https://github.com/Fafa-DL/Awesome-Backbones. It's easier compared with the official code of RepVGG, and there are many others classic models for classification.

Now the training codes are all from the official repo, and I create a new repo here to store the modifications I made for training AAAI dataset.

In addition, show the results below

Model	Type	Accuracy	Weight	onnx
RepVGG	A0	88.23%	-	-
RepOpt	B1	90.91%	-	-
Repopt	A0-(Hyper-serach-AAAI-dataset)	89.42%	-	-
Repopt	A0-(Hyper-serach-CF100-dataset)	89.84%	-	-

Train Repopt-A0

For the scale file of A0 isn't released, so I train the scale file. Run the code in terminal below:

python3 -m torch.distributed.launch --nproc_per_node 1 --master_port 12349 main_repopt.py --data-path /data/AAAI/Awesome-Backbones/datasets --arch RepOpt-VGG-A0-hs --batch-size 32 --tag search --opts TRAIN.EPOCHS 240 TRAIN.BASE_LR 0.1 TRAIN.WEIGHT_DECAY 4e-5 TRAIN.WARMUP_EPOCHS 10 MODEL.LABEL_SMOOTHING 0.1 DATA.DATASET imagenet 

Something about the dataset

Through observing the picture number of each class, the long-tailed feature is obvious.

The statistical data:

	num
count	89.000000
mean	561.831461
std	809.292792
min	8.000000
25%	135.000000
50%	271.000000
75%	502.000000
max	5121.000000

The resolution of dataset

Not only is the number of each class unbalanced, but also the resolution of the pictures vary greatly. As the image below shows:

The channel of the picture is 3. Among all 50003 pictures, there are 11133 different resolution. Most of them are smaller than 200*200.

The accuracy of each class

if you want to reproduce the result, you can run

python3 -m torch.distributed.launch --nproc_per_node 1 --master_port 12349 main.py --arch RepVGG-A0 --data-path /data/AAAI/datasets --batch-size 32 --tag test --eval --resume /home/qinhaiyan/RepVGG/result/VGG-A0/RepVGG-A0/train_EQLv2/best_ckpt.pth  --opts DATA.DATASET imagenet DATA.IMG_SIZE 112 --out /home/qinhaiyan/RepVGG/result/VGG-A0/RepVGG-A0/train_EQLv2 

and then run this code.

Validation Dataset

This table show the accuracy of each class with ascending trend. The validation number means the number of each class in validation dataset. The sampler ratio is 0.1, except the class 81, I add all 8 picture as both validation dataset and training dataset.

Class	Accuracy	validation num
81	0.125000	8
83	0.418803	234
76	0.500000	2
63	0.558824	136
74	0.559322	118
79	0.647059	17
86	0.666667	3
73	0.686441	118
78	0.687500	16
62	0.699248	133
27	0.714286	7
52	0.800000	10
64	0.815789	38
84	0.833333	234
32	0.846154	13
82	0.850427	234
65	0.888889	36
33	0.900000	10
8	0.904255	94
39	0.904762	21
80	0.905512	127
41	0.906250	32
7	0.909091	44
14	0.909091	22
5	0.909091	11
70	0.911765	306
72	0.915254	118
53	0.920000	25
75	0.922780	259
10	0.931034	58
24	0.935484	31
66	0.937500	16
40	0.937500	16
85	0.939453	512
19	0.944444	18
38	0.945946	37
20	0.958333	24
61	0.965517	58
48	0.968750	32
18	0.975610	41
12	0.976190	42
87	0.978261	46
50	0.979167	48
6	0.982906	117
44	0.984615	65
4	0.987805	82
43	0.988372	172
67	1.000000	29
68	1.000000	1
69	1.000000	1
71	1.000000	115
60	1.000000	30
58	1.000000	10
59	1.000000	210
77	1.000000	17
57	1.000000	26
0	1.000000	22
55	1.000000	23
1	1.000000	29
2	1.000000	37
3	1.000000	44
9	1.000000	27
11	1.000000	32
13	1.000000	29
15	1.000000	3
16	1.000000	18
17	1.000000	18
21	1.000000	15
22	1.000000	19
23	1.000000	30
25	1.000000	30
26	1.000000	7
28	1.000000	11
29	1.000000	7
30	1.000000	25
31	1.000000	7
34	1.000000	8
35	1.000000	7
36	1.000000	11
37	1.000000	11
42	1.000000	25
45	1.000000	2
46	1.000000	8
47	1.000000	45
49	1.000000	14
51	1.000000	50
54	1.000000	60
56	1.000000	1
88	1.000000	15

Training Dataset

In addition, I test the model’s training performance in training dataset：

class	num in training dataset	True	ACC
81	8	1	0.125000
86	31	8	0.258065
76	27	11	0.407407
83	2111	889	0.421127
63	1224	691	0.564542
79	154	90	0.584416
74	1068	625	0.585206
62	1202	887	0.737937
52	98	73	0.744898
78	153	118	0.771242
73	1066	833	0.781426
88	135	110	0.814815
82	2112	1797	0.850852
64	347	297	0.855908
84	2110	1835	0.869668
65	327	287	0.877676
72	1071	965	0.901027
45	21	19	0.904762
26	66	60	0.909091
53	234	214	0.914530
68	12	11	0.916667
70	2763	2595	0.939197
85	4609	4342	0.942070
75	2333	2200	0.942992
27	71	67	0.943662
80	1147	1085	0.945946
39	198	191	0.964646
22	172	166	0.965116
33	92	89	0.967391
77	155	150	0.967742
15	32	31	0.968750
16	162	157	0.969136
10	524	509	0.971374
8	846	823	0.972813
18	369	361	0.978320
9	244	239	0.979508
48	293	287	0.979522
14	205	201	0.980488
12	379	372	0.981530
17	166	163	0.981928
20	225	221	0.982222
7	400	393	0.982500
35	66	65	0.984848
4	743	732	0.985195
71	1037	1024	0.987464
3	403	398	0.987593
1	262	259	0.988550
6	1053	1041	0.988604
11	290	287	0.989655
41	292	289	0.989726
61	530	525	0.990566
36	107	106	0.990654
30	225	223	0.991111
57	243	241	0.991770
32	122	121	0.991803
47	413	410	0.992736
24	279	277	0.992832
40	149	148	0.993289
38	340	338	0.994118
43	1553	1544	0.994205
54	541	538	0.994455
44	593	590	0.994941
0	207	206	0.995169
87	414	412	0.995169
13	269	268	0.996283
25	271	270	0.996310
23	277	276	0.996390
2	335	334	0.997015
50	432	431	0.997685
51	452	451	0.997788
5	102	102	1.000000
19	171	171	1.000000
21	138	138	1.000000
28	99	99	1.000000
29	71	71	1.000000
31	69	69	1.000000
59	1892	1892	1.000000
60	274	274	1.000000
42	232	232	1.000000
46	74	74	1.000000
49	131	131	1.000000
69	15	15	1.000000
67	264	264	1.000000
66	146	146	1.000000
55	211	211	1.000000
56	15	15	1.000000
58	98	98	1.000000
34	79	79	1.000000
37	100	100	1.000000

EQLv2 Trainnig Dataset

==Need to update.==

The analysis of model and performance on different hardware

Model Analysis

Type	num blocks	width multiplier	Flops	Params	Max Memory Cost
RepVGG-A0	[2,4,14,1]	[0.75,0.75,0.75,2.5]	202.337M	7.142M	2261.25M(stage4.Conv2d)
RepVGG-N0	[2,4,14,1]	[1.0,1.0,1.0,1.0]	268.466M	9.784M	1192.5M(stage4.Conv2d)
RepVGG-N1	[2,4,14,1]	[0.25,0.25,0.25,0.25]	67.117M	621.369K	245.4375M(stage0.Conv2d)
RepVGG-N2	[2,4,14,1]	[0.5,0.5,0.5,0.5]	134.233M	2.459M	490.875M(stage0.Conv2d)
RepVGG-N3	[2,4,14,1]	[0.25,0.25,0.25,0.1]	67.540M	877.113K	328.5M(stage4.Conv2d)
RepVGG-N4	[2,4,14,1]	[0.5,0.5,0.5,1.0]	134.515M	2.777M	616.5M(stage4.Conv2d)
RepVGG-M0	[2,2,2,2]	[1.0,1.0,1.0,1.0]	252.95M	4.768M	2344.5M(stage4.Conv2d)
RepVGG-M1	[2,2,2,2]	[0.75,0.75,0.75,2.5]	191.220M	17.740M	14501.25M(stage4.Conv2d)

Memory计算策略

每一次计算需要存储在片上内存的有：

1. 本层的weight和bias
2. 卷积的输出out_conv（int32）
3. 卷积输出转换为下一层输入（int8）

relu应该不用存吧？

The detail of computation can be found in [The Memory Cost computation process ](./The Memory Cost computation process.md)

	ATLAS300(int8)-Latency(ms)	RV1126(int8)-Latency(ms)	T4(int8)-Latency(ms)	Adreno 650 GPU(fp16)-Latency(ms)	Hexagon 698 DSP(uint8)-Latency(ms)	TDA4VM(int8)-Latency(ms)	STPU(int8)-Latency(ms)	Max Memory Cost
A0	0.641636		0.683414	10.06408	2.645636	1.684263		2261.25
N0	0.671535	4.930283	0.782737	13.22945	3.153485	1.815879	1.50102	1192.5
N1						0.749081		245.4375
N2	0.547707	3.260808		7.10899	1.886929	0.988788		490.875
M0	0.490182	4.128485	0.393848	7.735838	2.07997	1.215657	0.987919	2344.5
M1	0.846596	6.195939	0.465677	15.59577	Failure	5.205919	2.016444	14501.25