You can now find us at CVPR 2020. Our live Q&A sessions are on June 18, 2020 @ 5pm - 7pm PDT (click here to join) and June 19, 2020 @ 5am - 7am PDT (click here to join). We are looking forward to seeing you at CVPR!

CVPR 2020 is now over, and we thank you for all the interesting discussions! Our presentation video is available on YouTube. We will continue the development of the code and models in this repository, so stay tuned!

---

Blueprint Separable Convolutions (BSConv)

This repository provides code and trained models for the CVPR 2020 paper (official, arXiv):

Rethinking Depthwise Separable Convolutions: How Intra-Kernel Correlations Lead to Improved MobileNets

Daniel Haase*, Manuel Amthor*

Table of Contents

1. Overview
2. Results
   1. CIFAR10 - MobileNetV1
   2. CIFAR10 - MobileNetV2
   3. CIFAR10 - MobileNetV3-small
   4. CIFAR10 - MobileNetV3-large
   5. CIFAR10 - WRN-16
   6. CIFAR10 - WRN-28
   7. CIFAR10 - WRN-40
   8. CIFAR100 - MobileNetV1
   9. CIFAR100 - MobileNetV2
   10. CIFAR100 - MobileNetV3-small
   11. CIFAR100 - MobileNetV3-large
   12. CIFAR100 - ResNets
   13. CIFAR100 - WRN-16
   14. CIFAR100 - WRN-28
   15. CIFAR100 - WRN-40
3. Requirements
4. Installation
5. Usage
6. Change Log
7. Citation

Results

CIFAR10 - MobileNetV1

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_mobilenetv1_w1	93.57	3.22	179.34
cifar_mobilenetv1_w3d4	93.51	1.82	102.66
cifar_mobilenetv1_w1d2	92.44	0.82	47.21
cifar_mobilenetv1_w1d4	91.17	0.22	12.99
cifar_mobilenetv1_w1_bsconvu	94.48	3.22	254.64
cifar_mobilenetv1_w3d4_bsconvu	94.38	1.82	144.98
cifar_mobilenetv1_w1d2_bsconvu	93.45	0.82	65.98
cifar_mobilenetv1_w1d4_bsconvu	92.13	0.22	17.66

CIFAR10 - MobileNetV2

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_mobilenetv2_w1	93.91	2.24	92.40
cifar_mobilenetv2_w3d4	93.76	1.36	55.13
cifar_mobilenetv2_w1d2	92.55	0.70	27.32
cifar_mobilenetv2_w1d4	89.93	0.25	8.97
cifar_mobilenetv2_w1_bsconvs_p1d6	94.47	2.24	92.40
cifar_mobilenetv2_w3d4_bsconvs_p1d6	94.16	1.36	55.13
cifar_mobilenetv2_w1d2_bsconvs_p1d6	93.30	0.70	27.32
cifar_mobilenetv2_w1d4_bsconvs_p1d6	90.60	0.25	8.97

CIFAR10 - MobileNetV3-small

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_mobilenetv3_small_w1	92.57	1.09	18.48
cifar_mobilenetv3_small_w3d4	91.46	0.72	11.40
cifar_mobilenetv3_small_w1d2	90.33	0.44	6.00
cifar_mobilenetv3_small_w7d20	88.75	0.31	3.45
cifar_mobilenetv3_small_w1_bsconvs_p1d6	93.06	1.09	18.48
cifar_mobilenetv3_small_w3d4_bsconvs_p1d6	92.10	0.72	11.40
cifar_mobilenetv3_small_w1d2_bsconvs_p1d6	90.58	0.44	6.00
cifar_mobilenetv3_small_w7d20_bsconvs_p1d6	89.04	0.31	3.45

CIFAR10 - MobileNetV3-large

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_mobilenetv3_large_w1	94.38	2.98	68.45
cifar_mobilenetv3_large_w3d4	94.00	1.73	40.67
cifar_mobilenetv3_large_w1d2	93.30	0.82	20.00
cifar_mobilenetv3_large_w7d20	92.16	0.44	10.89
cifar_mobilenetv3_large_w1_bsconvs_p1d6	94.81	2.98	68.45
cifar_mobilenetv3_large_w3d4_bsconvs_p1d6	94.34	1.73	40.67
cifar_mobilenetv3_large_w1d2_bsconvs_p1d6	93.85	0.82	20.00
cifar_mobilenetv3_large_w7d20_bsconvs_p1d6	92.45	0.44	10.89

CIFAR10 - WideResNets (WRN-16)

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_wrn16_1	91.11	0.18	27.06
cifar_wrn16_2	93.40	0.69	101.86
cifar_wrn16_4	94.29	2.75	394.06
cifar_wrn16_6	94.60	6.17	877.10
cifar_wrn16_8	95.05	10.96	1550.99
cifar_wrn16_10	95.03	17.12	2415.71
cifar_wrn16_12	95.11	24.64	3471.28
cifar_wrn16_1_bsconvu	89.09	0.03	5.57
cifar_wrn16_2_bsconvu	91.83	0.10	18.74
cifar_wrn16_4_bsconvu	93.56	0.36	66.59
cifar_wrn16_6_bsconvu	94.13	0.80	143.80
cifar_wrn16_8_bsconvu	94.46	1.41	250.38
cifar_wrn16_10_bsconvu	94.54	2.19	386.31
cifar_wrn16_12_bsconvu	94.82	3.13	551.60
cifar_wrn16_1_bsconvs_p1d4	87.34	0.02	4.01
cifar_wrn16_2_bsconvs_p1d4	91.56	0.06	11.85
cifar_wrn16_4_bsconvs_p1d4	93.31	0.21	38.00
cifar_wrn16_6_bsconvs_p1d4	94.48	0.46	78.84
cifar_wrn16_8_bsconvs_p1d4	94.93	0.80	134.35
cifar_wrn16_10_bsconvs_p1d4	95.17	1.23	204.55
cifar_wrn16_12_bsconvs_p1d4	95.28	1.75	289.42

CIFAR10 - WideResNets (WRN-28)

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_wrn28_1	92.36	0.37	55.72
cifar_wrn28_2	94.29	1.47	215.79
cifar_wrn28_3	94.99	3.29	479.94
cifar_wrn28_4	94.96	5.85	848.42
cifar_wrn28_6	95.35	13.14	1898.38
cifar_wrn28_8	95.73	23.35	3365.68
cifar_wrn28_10	95.72	36.48	5250.31
cifar_wrn28_12	95.54	52.52	7552.27
cifar_wrn28_1_bsconvu	91.28	0.05	10.09
cifar_wrn28_2_bsconvu	93.39	0.19	34.08
cifar_wrn28_3_bsconvu	93.77	0.42	71.44
cifar_wrn28_4_bsconvu	94.59	0.73	122.43
cifar_wrn28_6_bsconvu	95.16	1.61	265.31
cifar_wrn28_8_bsconvu	95.21	2.82	462.71
cifar_wrn28_10_bsconvu	95.36	4.39	714.64
cifar_wrn28_12_bsconvu	95.46	6.29	1021.10
cifar_wrn28_1_bsconvs_p1d4	90.22	0.04	7.25
cifar_wrn28_2_bsconvs_p1d4	93.13	0.12	21.47
cifar_wrn28_3_bsconvs_p1d4	94.28	0.24	42.23
cifar_wrn28_4_bsconvs_p1d4	94.81	0.41	69.82
cifar_wrn28_6_bsconvs_p1d4	95.10	0.88	145.44
cifar_wrn28_8_bsconvs_p1d4	95.44	1.53	248.32
cifar_wrn28_10_bsconvs_p1d4	96.02	2.36	378.46
cifar_wrn28_12_bsconvs_p1d4	96.29	3.37	535.87

CIFAR10 - WideResNets (WRN-40)

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_wrn40_1	93.30	0.56	84.37
cifar_wrn40_2	94.44	2.24	329.73
cifar_wrn40_3	95.03	5.04	735.78
cifar_wrn40_4	95.36	8.95	1302.78
cifar_wrn40_6	95.63	20.12	2919.66
cifar_wrn40_8	95.58	35.75	5180.37
cifar_wrn40_10	95.66	55.84	8084.90
cifar_wrn40_1_bsconvu	92.07	0.08	14.61
cifar_wrn40_2_bsconvu	93.91	0.29	49.41
cifar_wrn40_3_bsconvu	94.65	0.63	103.88
cifar_wrn40_4_bsconvu	94.80	1.09	178.27
cifar_wrn40_6_bsconvu	95.20	2.41	386.81
cifar_wrn40_8_bsconvu	95.54	4.24	675.05
cifar_wrn40_10_bsconvu	95.83	6.59	1042.98
cifar_wrn40_1_bsconvs_p1d4	91.24	0.05	10.49
cifar_wrn40_2_bsconvs_p1d4	93.55	0.17	31.08
cifar_wrn40_3_bsconvs_p1d4	94.64	0.36	61.38
cifar_wrn40_4_bsconvs_p1d4	94.98	0.61	101.64
cifar_wrn40_6_bsconvs_p1d4	95.66	1.31	212.04
cifar_wrn40_8_bsconvs_p1d4	95.74	2.27	362.29
cifar_wrn40_10_bsconvs_p1d4	96.00	3.50	552.38

CIFAR100 - MobileNetV1

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_mobilenetv1_w1	74.58	3.31	179.43
cifar_mobilenetv1_w3d4	73.48	1.89	102.72
cifar_mobilenetv1_w1d2	71.61	0.87	47.25
cifar_mobilenetv1_w1d4	68.23	0.24	13.01
cifar_mobilenetv1_w1_bsconvu	75.80	3.31	254.73
cifar_mobilenetv1_w3d4_bsconvu	75.27	1.89	145.04
cifar_mobilenetv1_w1d2_bsconvu	73.59	0.87	66.03
cifar_mobilenetv1_w1d4_bsconvu	70.37	0.24	17.68

CIFAR100 - MobileNetV2

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_mobilenetv2_w1	74.67	2.35	92.51
cifar_mobilenetv2_w3d4	74.57	1.48	55.24
cifar_mobilenetv2_w1d2	73.03	0.81	27.43
cifar_mobilenetv2_w1d4	67.89	0.36	9.08
cifar_mobilenetv2_w1_bsconvs_p1d6	76.91	2.35	92.51
cifar_mobilenetv2_w3d4_bsconvs_p1d6	75.45	1.48	55.24
cifar_mobilenetv2_w1d2_bsconvs_p1d6	73.43	0.81	27.43
cifar_mobilenetv2_w1d4_bsconvs_p1d6	69.06	0.36	9.08

CIFAR100 - MobileNetV3-small

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_mobilenetv3_small_w1	72.93	1.15	18.54
cifar_mobilenetv3_small_w3d4	70.87	0.77	11.46
cifar_mobilenetv3_small_w1d2	66.83	0.49	6.05
cifar_mobilenetv3_small_w7d20	63.16	0.37	3.50
cifar_mobilenetv3_small_w1_bsconvs_p1d6	73.93	1.15	18.54
cifar_mobilenetv3_small_w3d4_bsconvs_p1d6	72.28	0.77	11.46
cifar_mobilenetv3_small_w1d2_bsconvs_p1d6	68.92	0.49	6.05
cifar_mobilenetv3_small_w7d20_bsconvs_p1d6	65.90	0.37	3.50

CIFAR100 - MobileNetV3-large

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_mobilenetv3_large_w1	75.09	3.07	68.54
cifar_mobilenetv3_large_w3d4	74.42	1.82	40.75
cifar_mobilenetv3_large_w1d2	71.83	0.91	20.09
cifar_mobilenetv3_large_w7d20	70.34	0.52	10.98
cifar_mobilenetv3_large_w1_bsconvs_p1d6	78.11	3.07	68.54
cifar_mobilenetv3_large_w3d4_bsconvs_p1d6	76.41	1.82	40.75
cifar_mobilenetv3_large_w1d2_bsconvs_p1d6	75.22	0.91	20.09
cifar_mobilenetv3_large_w7d20_bsconvs_p1d6	72.31	0.52	10.98

CIFAR100 - ResNets

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_resnet20	68.59	0.28	41.42
cifar_resnet56	71.31	0.86	127.39
cifar_resnet110	71.29	1.74	256.34
cifar_resnet302	72.22	4.85	714.83
cifar_resnet602	71.22	9.71	1431.22
cifar_resnet20_bsconvu	64.41	0.05	7.86
cifar_resnet56_bsconvu	69.43	0.13	21.42
cifar_resnet110_bsconvu	71.16	0.24	41.77
cifar_resnet302_bsconvu	72.67	0.67	114.12
cifar_resnet602_bsconvu	73.48	1.33	227.17
cifar_resnet20_bsconvs_p1d4	62.03	0.03	5.66
cifar_resnet56_bsconvs_p1d4	68.72	0.08	15.37
cifar_resnet110_bsconvs_p1d4	71.15	0.16	29.93
cifar_resnet302_bsconvs_p1d4	72.53	0.43	81.70
cifar_resnet602_bsconvs_p1d4	73.05	0.85	162.60

CIFAR100 - WideResNets (WRN-16)

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_wrn16_1	66.81	0.18	27.07
cifar_wrn16_2	71.29	0.70	101.87
cifar_wrn16_4	75.07	2.77	394.08
cifar_wrn16_6	76.50	6.21	877.14
cifar_wrn16_8	77.30	11.01	1551.03
cifar_wrn16_10	77.28	17.17	2415.77
cifar_wrn16_12	78.02	24.71	3471.34
cifar_wrn16_1_bsconvu	62.79	0.03	5.57
cifar_wrn16_2_bsconvu	68.33	0.11	18.75
cifar_wrn16_4_bsconvu	72.51	0.39	66.62
cifar_wrn16_6_bsconvu	74.02	0.84	143.84
cifar_wrn16_8_bsconvu	75.61	1.45	250.42
cifar_wrn16_10_bsconvu	76.23	2.24	386.36
cifar_wrn16_12_bsconvu	76.48	3.20	551.67
cifar_wrn16_1_bsconvs_p1d4	58.48	0.02	4.02
cifar_wrn16_2_bsconvs_p1d4	68.62	0.07	11.86
cifar_wrn16_4_bsconvs_p1d4	73.01	0.24	38.03
cifar_wrn16_6_bsconvs_p1d4	75.46	0.49	78.87
cifar_wrn16_8_bsconvs_p1d4	77.18	0.84	134.40
cifar_wrn16_10_bsconvs_p1d4	77.64	1.29	204.60
cifar_wrn16_12_bsconvs_p1d4	78.39	1.82	289.49

CIFAR100 - WideResNets (WRN-28)

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_wrn28_1	69.00	0.38	55.72
cifar_wrn28_2	73.38	1.48	215.80
cifar_wrn28_3	75.25	3.31	479.96
cifar_wrn28_4	76.85	5.87	848.44
cifar_wrn28_6	78.18	13.18	1898.42
cifar_wrn28_8	78.07	23.40	3365.72
cifar_wrn28_10	78.58	36.54	5250.36
cifar_wrn28_12	79.04	52.59	7552.34
cifar_wrn28_1_bsconvu	66.21	0.06	10.09
cifar_wrn28_2_bsconvu	71.78	0.20	34.09
cifar_wrn28_3_bsconvu	73.79	0.44	71.46
cifar_wrn28_4_bsconvu	75.29	0.75	122.45
cifar_wrn28_6_bsconvu	76.67	1.64	265.34
cifar_wrn28_8_bsconvu	77.15	2.87	462.76
cifar_wrn28_10_bsconvu	78.04	4.44	714.70
cifar_wrn28_12_bsconvu	78.30	6.36	1021.17
cifar_wrn28_1_bsconvs_p1d4	64.65	0.04	7.26
cifar_wrn28_2_bsconvs_p1d4	71.55	0.13	21.48
cifar_wrn28_3_bsconvs_p1d4	74.42	0.26	42.25
cifar_wrn28_4_bsconvs_p1d4	76.22	0.43	69.84
cifar_wrn28_6_bsconvs_p1d4	78.18	0.92	145.47
cifar_wrn28_8_bsconvs_p1d4	79.49	1.58	248.36
cifar_wrn28_10_bsconvs_p1d4	80.09	2.42	378.52
cifar_wrn28_12_bsconvs_p1d4	80.26	3.44	535.94

CIFAR100 - WideResNets (WRN-40)

Model	Accuracy (top-1)	Params [M]	FLOPs [M]
cifar_wrn40_1	70.34	0.57	84.38
cifar_wrn40_2	74.13	2.26	329.74
cifar_wrn40_3	75.70	5.06	735.79
cifar_wrn40_4	77.55	8.97	1302.81
cifar_wrn40_6	77.41	20.15	2919.70
cifar_wrn40_8	78.33	35.79	5180.42
cifar_wrn40_10	78.49	55.90	8084.96
cifar_wrn40_1_bsconvu	68.98	0.09	14.61
cifar_wrn40_2_bsconvu	72.41	0.30	49.42
cifar_wrn40_3_bsconvu	74.91	0.64	103.90
cifar_wrn40_4_bsconvu	76.42	1.12	178.29
cifar_wrn40_6_bsconvu	77.12	2.44	386.85
cifar_wrn40_8_bsconvu	78.01	4.29	675.09
cifar_wrn40_10_bsconvu	78.45	6.64	1043.03
cifar_wrn40_1_bsconvs_p1d4	67.66	0.06	10.49
cifar_wrn40_2_bsconvs_p1d4	73.19	0.18	31.09
cifar_wrn40_3_bsconvs_p1d4	75.83	0.37	61.40
cifar_wrn40_4_bsconvs_p1d4	76.97	0.63	101.66
cifar_wrn40_6_bsconvs_p1d4	78.42	1.34	212.07
cifar_wrn40_8_bsconvs_p1d4	79.51	2.32	362.33
cifar_wrn40_10_bsconvs_p1d4	80.21	3.56	552.44

Requirements

• Python>=3.6
• PyTorch>=1.0.0 (support for other frameworks will be added later)

Installation

pip install --upgrade bsconv

Usage

See here for PyTorch usage details.

Support for other frameworks will be added later.

Please note that the code provided here is work-in-progress. Therefore, some features may be missing or may change between versions.

Change Log

0.4.0 (2020-09-22)

• BSConv for PyTorch:
  • added support for more model architectures (see bsconv.pytorch.get_model)
  • added result tables and plots for ResNets, WRNs, MobileNets on CIFAR datasets
  • removed script bin/bsconv_pytorch_list_architectures.py, because bsconv.pytorch.get_model is more flexible now (see the BSConv PyTorch usage readme for available architectures)

0.3.0 (2020-06-16)

• BSConv for PyTorch:
  • added ready-to-use model definitions (MobileNetV1, MobileNetV2, MobileNetsV3, ResNets and WRNs and their BSConv variants for CIFAR and ImageNet/fine-grained datasets)
  • added training script for CIFAR and ImageNet/fine-grained datasets
  • added class for the StanfordDogs dataset

0.2.0 (2020-04-16)

• BSConv for PyTorch:
  • removed activation and added option for normalization of PW layers in BSConv-S (issue #1) (API change)
  • added option for normalization of PW layers in BSConv-U (API change)
  • ensure that BSConv-S never uses more mid channels (= M') than input channels (M) and added parameter min_mid_channels (= M'_min) (API change)
  • added model profiler for parameter and FLOP counting
  • replacer now shows number of old and new model parameters

0.1.0 (2020-04-08)

• first public version
• BSConv for PyTorch:
  • modules BSConvU and BSConvS
  • replacers BSConvU_Replacer and BSConvS_Replacer

Citation

If you find this work useful in your own research, please cite the paper as:

@InProceedings{Haase_2020_CVPR,
    author = {Haase, Daniel and Amthor, Manuel},
    title = {Rethinking Depthwise Separable Convolutions: How Intra-Kernel Correlations Lead to Improved {MobileNets}},
    booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month = {June},
    year = {2020}
}