Large-scale image classification models on Keras with MXNet backend

This is a collection of large-scale image classification models. Many of them are pretrained on ImageNet-1K dataset and loaded automatically during use. All pretrained models require the same ordinary normalization. Scripts for training/evaluating/converting models are in the imgclsmob repo.

List of implemented models

AlexNet ('One weird trick for parallelizing convolutional neural networks')
VGG/BN-VGG ('Very Deep Convolutional Networks for Large-Scale Image Recognition')
ResNet ('Deep Residual Learning for Image Recognition')
PreResNet ('Identity Mappings in Deep Residual Networks')
ResNeXt ('Aggregated Residual Transformations for Deep Neural Networks')
SENet/SE-ResNet/SE-PreResNet/SE-ResNeXt ('Squeeze-and-Excitation Networks')
DenseNet ('Densely Connected Convolutional Networks')
DarkNet Ref/Tiny/19 ('Darknet: Open source neural networks in c')
DarkNet-53 ('YOLOv3: An Incremental Improvement')
SqueezeNet/SqueezeResNet ('SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size')
SqueezeNext ('SqueezeNext: Hardware-Aware Neural Network Design')
ShuffleNet ('ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices')
ShuffleNetV2 ('ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design')
MENet ('Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications')
MobileNet ('MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications')
FD-MobileNet ('FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy')
MobileNetV2 ('MobileNetV2: Inverted Residuals and Linear Bottlenecks')
IGCV3 ('IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural Networks')
MnasNet ('MnasNet: Platform-Aware Neural Architecture Search for Mobile')

Installation

To use the models in your project, simply install the kerascv package with mxnet:

pip install kerascv mxnet>=1.2.1

To enable different hardware supports such as GPUs, check out MXNet variants. For example, you can install with CUDA-9.2 supported MXNet:

pip install kerascv mxnet-cu92>=1.2.1

After installation change the value of the image_data_format field to channels_first in the file ~/.keras/keras.json. Also check that the backend field is set to mxnet.

Usage

Example of using the pretrained ResNet-18 model:

from kerascv.model_provider import get_model as kecv_get_model
import numpy as np

net = kecv_get_model("resnet18", pretrained=True)
x = np.zeros((1, 3, 224, 224), np.float32)
y = net.predict(x)

Pretrained models

Some remarks:

Top1/Top5 are the standard 1-crop Top-1/Top-5 errors (in percents) on the validation subset of the ImageNet-1K dataset.
FLOPs/2 is the number of FLOPs divided by two to be similar to the number of MACs.
Remark Converted from GL model means that the model was trained on MXNet/Gluon and then converted to Keras.

Model	Top1	Top5	Params	FLOPs/2	Remarks
AlexNet	44.10	21.26	61,100,840	714.83M	From dmlc/gluon-cv (log)
VGG-11	31.90	11.75	132,863,336	7,615.87M	From dmlc/gluon-cv (log)
VGG-13	31.06	11.12	133,047,848	11,317.65M	From dmlc/gluon-cv (log)
VGG-16	26.78	8.69	138,357,544	15,507.20M	From dmlc/gluon-cv (log)
VGG-19	25.87	8.23	143,667,240	19,642.55M	From dmlc/gluon-cv (log)
BN-VGG-11b	30.34	10.57	132,868,840	7,630.72M	From dmlc/gluon-cv (log)
BN-VGG-13b	29.48	10.16	133,053,736	11,342.14M	From dmlc/gluon-cv (log)
BN-VGG-16b	26.88	8.65	138,365,992	15,507.20M	From dmlc/gluon-cv (log)
BN-VGG-19b	25.65	8.14	143,678,248	19,672.26M	From dmlc/gluon-cv (log)
ResNet-10	37.09	15.54	5,418,792	894.04M	Converted from GL model (log)
ResNet-12	35.86	14.45	5,492,776	1,126.25M	Converted from GL model (log)
ResNet-14	32.85	12.42	5,788,200	1,357.94M	Converted from GL model (log)
ResNet-16	30.67	11.09	6,968,872	1,589.34M	Converted from GL model (log)
ResNet-18 x0.25	49.14	24.45	831,096	137.32M	Converted from GL model (log)
ResNet-18 x0.5	36.54	14.96	3,055,880	486.49M	Converted from GL model (log)
ResNet-18 x0.75	33.24	12.54	6,675,352	1,047.53M	Converted from GL model (log)
ResNet-18	28.08	9.52	11,689,512	1,820.41M	Converted from GL model (log)
ResNet-34	25.32	7.92	21,797,672	3,672.68M	From dmlc/gluon-cv (log)
ResNet-50	22.63	6.41	25,557,032	3,877.95M	From dmlc/gluon-cv (log)
ResNet-50b	22.31	6.18	25,557,032	4,110.48M	From dmlc/gluon-cv (log)
ResNet-101	21.64	5.99	44,549,160	7,597.95M	From dmlc/gluon-cv (log)
ResNet-101b	20.78	5.39	44,549,160	7,830.48M	From dmlc/gluon-cv (log)
ResNet-152	20.74	5.35	60,192,808	11,321.85M	From dmlc/gluon-cv (log)
ResNet-152b	20.30	5.25	60,192,808	11,554.38M	From dmlc/gluon-cv (log)
PreResNet-18	28.16	9.52	11,687,848	1,820.56M	Converted from GL model (log)
PreResNet-34	25.86	8.11	21,796,008	3,672.83M	From dmlc/gluon-cv (log)
PreResNet-50	23.38	6.68	25,549,480	3,875.44M	From dmlc/gluon-cv (log)
PreResNet-50b	23.14	6.63	25,549,480	4,107.97M	From dmlc/gluon-cv (log)
PreResNet-101	21.43	5.75	44,541,608	7,595.44M	From dmlc/gluon-cv (log)
PreResNet-101b	21.71	5.88	44,541,608	7,827.97M	From dmlc/gluon-cv (log)
PreResNet-152	20.69	5.31	60,185,256	11,319.34M	From dmlc/gluon-cv (log)
PreResNet-152b	20.99	5.76	60,185,256	11,551.87M	From dmlc/gluon-cv (log)
PreResNet-200b	21.09	5.64	64,666,280	15,068.63M	From tornadomeet/ResNet (log)
ResNeXt-101 (32x4d)	21.30	5.78	44,177,704	8,003.45M	From Cadene/pretrained...pytorch (log)
ResNeXt-101 (64x4d)	20.59	5.41	83,455,272	15,500.27M	From Cadene/pretrained...pytorch (log)
SE-ResNet-50	22.50	6.43	28,088,024	3,880.49M	From Cadene/pretrained...pytorch (log)
SE-ResNet-101	21.92	5.88	49,326,872	7,602.76M	From Cadene/pretrained...pytorch (log)
SE-ResNet-152	21.46	5.77	66,821,848	11,328.52M	From Cadene/pretrained...pytorch (log)
SE-ResNeXt-50 (32x4d)	21.05	5.57	27,559,896	4,258.40M	From Cadene/pretrained...pytorch (log)
SE-ResNeXt-101 (32x4d)	19.98	4.99	48,955,416	8,008.26M	From Cadene/pretrained...pytorch (log)
SENet-154	18.83	4.65	115,088,984	20,745.78M	From Cadene/pretrained...pytorch (log)
DenseNet-121	25.09	7.80	7,978,856	2,872.13M	From dmlc/gluon-cv (log)
DenseNet-161	22.39	6.18	28,681,000	7,793.16M	From dmlc/gluon-cv (log)
DenseNet-169	23.88	6.89	14,149,480	3,403.89M	From dmlc/gluon-cv (log)
DenseNet-201	22.69	6.35	20,013,928	4,347.15M	From dmlc/gluon-cv (log)
DarkNet Tiny	40.31	17.46	1,042,104	500.85M	Converted from GL model (log)
DarkNet Ref	37.99	16.68	7,319,416	367.59M	Converted from GL model (log)
DarkNet-53	21.43	5.56	41,609,928	7,133.86M	From dmlc/gluon-cv (log)
SqueezeNet v1.0	39.17	17.56	1,248,424	823.67M	Converted from GL model (log)
SqueezeNet v1.1	39.08	17.39	1,235,496	352.02M	Converted from GL model (log)
SqueezeResNet v1.1	39.82	17.84	1,235,496	352.02M	Converted from GL model (log)
1.0-SqNxt-23	42.28	18.62	724,056	287.28M	Converted from GL model (log)
1.0-SqNxt-23v5	40.38	17.57	921,816	285.82M	Converted from GL model (log)
ShuffleNet x0.25 (g=1)	62.00	36.76	209,746	12.35M	Converted from GL model (log)
ShuffleNet x0.25 (g=3)	61.32	36.15	305,902	13.09M	Converted from GL model (log)
ShuffleNet x0.5 (g=3)	43.82	20.60	718,324	41.70M	Converted from GL model (log)
ShuffleNetV2 x0.5	40.76	18.40	1,366,792	43.31M	Converted from GL model (log)
ShuffleNetV2 x1.0	31.02	11.33	2,278,604	149.72M	Converted from GL model (log)
ShuffleNetV2 x1.5	32.46	12.47	4,406,098	320.77M	Converted from GL model (log)
ShuffleNetV2 x2.0	31.91	12.23	7,601,686	595.84M	Converted from GL model (log)
108-MENet-8x1 (g=3)	43.61	20.31	654,516	42.68M	Converted from GL model (log)
128-MENet-8x1 (g=4)	42.08	19.14	750,796	45.98M	Converted from GL model (log)
160-MENet-8x1 (g=8)	43.47	20.28	850,120	45.63M	Converted from GL model (log)
228-MENet-12x1 (g=3)	33.85	12.88	1,806,568	152.93M	Converted from GL model (log)
256-MENet-12x1 (g=4)	32.22	12.17	1,888,240	150.65M	Converted from GL model (log)
348-MENet-12x1 (g=3)	31.17	11.42	3,368,128	312.00M	From clavichord93/MENet (log)
352-MENet-12x1 (g=8)	34.69	13.75	2,272,872	157.35M	From clavichord93/MENet (log)
456-MENet-24x1 (g=3)	29.55	10.44	5,304,784	567.90M	From clavichord93/MENet (log)
MobileNet x0.25	45.80	22.17	470,072	44.09M	Converted from GL model (log)
MobileNet x0.5	33.94	13.30	1,331,592	155.42M	Converted from GL model (log)
MobileNet x0.75	29.85	10.51	2,585,560	333.99M	Converted from GL model (log)
MobileNet x1.0	26.43	8.66	4,231,976	579.80M	Converted from GL model (log)
FD-MobileNet x0.25	56.17	31.37	383,160	12.95M	Converted from GL model (log)
FD-MobileNet x0.5	42.61	19.69	993,928	41.84M	Converted from GL model (log)
FD-MobileNet x0.75	37.90	16.01	1,833,304	86.68M	Converted from GL model (log)
FD-MobileNet x1.0	33.80	13.12	2,901,288	147.46M	Converted from GL model (log)
MobileNetV2 x0.25	48.06	24.12	1,516,392	34.24M	Converted from GL model (log)
MobileNetV2 x0.5	35.63	14.43	1,964,736	100.13M	Converted from GL model (log)
MobileNetV2 x0.75	30.81	11.26	2,627,592	198.50M	From dmlc/gluon-cv (log)
MobileNetV2 x1.0	28.50	9.90	3,504,960	329.36M	From dmlc/gluon-cv (log)
IGCV3 x0.25	53.41	28.29	1,534,020	41.29M	Converted from GL model (log)
IGCV3 x0.5	39.39	17.04	1,985,528	111.12M	Converted from GL model (log)
IGCV3 x1.0	28.21	9.55	3,491,688	340.79M	From homles11/IGCV3 (log)
MnasNet	31.30	11.45	4,308,816	317.67M	From zeusees/Mnasnet...Model (log)

Development Status
- 3 - Alpha
Intended Audience
- Science/Research
License
- OSI Approved :: MIT License
Operating System
- OS Independent
Programming Language
- Python
Topic
- Scientific/Engineering :: Image Recognition

Release files for kerascv 0.0.18

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for kerascv 0.0.18
File	Size	Uploaded
kerascv-0.0.18.tar.gz	40.5 kB	Jan 20, 2019	Details

Built distribution (wheel)

Table of built distributions (wheels) for kerascv 0.0.18
File	Interpreter	ABI	Platform	Reset
kerascv-0.0.18-py2.py3-none-any.whl 75.3 kB Jan 20, 2019	Python 3, Python 2	none	any	Details

Total release size:115.8 kB

Release files / kerascv-0.0.18.tar.gz

Download URL	kerascv-0.0.18.tar.gz
Size	40.5 kB
Tags	Source
SHA-256 checksum How to use checksums	`ce09e8861fda2c8657bff8681d6e64293b25fdbab6bcab36b63bddecb33aa7b2`
BLAKE2b-256 checksum How to use checksums	`1b51081faae4d8c24c3bfb240cb49bd6b17c899ca97ba0cc90bde5bea364ef30`
Upload date	Jan 20, 2019
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/1.12.1 pkginfo/1.4.2 requests/2.18.4 setuptools/40.6.3 requests-toolbelt/0.8.0 tqdm/4.23.4 CPython/3.6.5`

Release files / kerascv-0.0.18-py2.py3-none-any.whl

Download URL	kerascv-0.0.18-py2.py3-none-any.whl
Size	75.3 kB
Tags	Python 2 Python 3
SHA-256 checksum How to use checksums	`dbedf93b8cdf83ca15a2137d4ac38b3a5080b2aec0c4533e0b9ee57d16e1c6fa`
BLAKE2b-256 checksum How to use checksums	`055643e8e91a1cfc0aee622d51d8b86e8b79f0f93a9c546656d258ac7dea8268`
Upload date	Jan 20, 2019
Uploaded using Trusted Publishing? What is trusted publishing?	No
Uploaded via	`twine/1.12.1 pkginfo/1.4.2 requests/2.18.4 setuptools/40.6.3 requests-toolbelt/0.8.0 tqdm/4.23.4 CPython/3.6.5`