kuan-wang / pytorch-mobilenet-v3 Goto Github PK

View Code? Open in Web Editor NEW

758.0 12.0 187.0 52 KB

MobileNetV3 in pytorch and ImageNet pretrained models

License: Apache License 2.0

Python 100.00%

pytorch imagenet mobilenet classification mobilenetv3 mobilenetv2

pytorch-mobilenet-v3's Introduction

A PyTorch implementation of MobileNetV3

This is a PyTorch implementation of MobileNetV3 architecture as described in the paper Searching for MobileNetV3.

Some details may be different from the original paper, welcome to discuss and help me figure it out.

[NEW] The pretrained model of small version mobilenet-v3 is online, accuracy achieves the same as paper.
[NEW] The paper updated on 17 May, so I renew the codes for that, but there still are some bugs.
[NEW] I remove the se before the global avg_pool (the paper may add it in error), and now the model size is close to paper.

Training & Accuracy

training setting:

number of epochs: 150
learning rate schedule: cosine learning rate, initial lr=0.05
weight decay: 4e-5
remove dropout
batch size: 256

MobileNetV3 large

	Madds	Parameters	Top1-acc	Pretrained Model
Offical 1.0	219 M	5.4 M	75.2%	-
Offical 0.75	155 M	4 M	73.3%	-
Ours 1.0	224 M	5.48 M	72.8%	-
Ours 0.75	148 M	3.91 M	-	-

MobileNetV3 small

	Madds	Parameters	Top1-acc	Pretrained Model
Offical 1.0	66 M	2.9 M	67.4%	-
Offical 0.75	44 M	2.4 M	65.4%	-
Ours 1.0	63 M	2.94 M	67.4%	[google drive]
Ours 0.75	46 M	2.38 M	-	-

Usage

Pretrained models are still training ...

    # pytorch 1.0.1
    # large
    net_large = mobilenetv3(mode='large')
    # small
    net_small = mobilenetv3(mode='small')
    state_dict = torch.load('mobilenetv3_small_67.4.pth.tar')
    net_small.load_state_dict(state_dict)

Data Pre-processing

I used the following code for data pre-processing on ImageNet:

normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                 std=[0.229, 0.224, 0.225])

input_size = 224
train_loader = torch.utils.data.DataLoader(
    datasets.ImageFolder(
    traindir, transforms.Compose([
        transforms.RandomResizedCrop(input_size),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        normalize,
    ])),
    batch_size=batch_size, shuffle=True,
    num_workers=n_worker, pin_memory=True)

val_loader = torch.utils.data.DataLoader(
    datasets.ImageFolder(valdir, transforms.Compose([
        transforms.Resize(int(input_size/0.875)),
        transforms.CenterCrop(input_size),
        transforms.ToTensor(),
        normalize,
    ])),
    batch_size=batch_size, shuffle=False,
    num_workers=n_worker, pin_memory=True)

pytorch-mobilenet-v3's People

Contributors

Stargazers

Watchers

Forkers

hzhang57 rimu123 yekyli wintersurvival collector-m kleinxin moxiaoguai1993 jingchenustc piyush-dubey chaoso yonghoonkwon jiaquanye aiyangyang963 wzheng1983 shaohangxu fighting-jj lxxaaa donproc xiaoye77 dexception doriswzg tilmto leo-xxx jia-honghenrylee baucheng zeitgeistqian jacke121 harimyi llucid-97 kanasukita isleder lawrencewxj test-error zhenzhenxiang ankitshah009 robbie194 sjzhuo ml-lab zl2290103097 wings0820 jiyuxuan926 zyqqing igordavidyuk xiaoketongxue qianchen94 liuyanwu khmily abnerxzhe 598717026 zhangtianlun12 mafangfang9 ingeniousfrog showlo eutenacity qunxiang jiangjingz 275901397 mykameli tracy-lee1993 naah69 yanbin-wang ilyamaslo kingvergil gm19900510 silicon2006 guobinli lee-seon-woo briantliao liuwenhaha jario-jin bitbeyhub godpgf soskek tracy20180426 xytmhy jihaonew robot-ai-machinelearning zengqi0730 xingjinglu kuaikuaikim xuweizhen22 pandinosaurus ahmadprasetyaaaa hikaruzzz kaiqiao1992 sweed111 smartparrot sean-wade promyer liyantett wesley-kang ioannanti andyliu93 seryogin17 bertmoons nguyenthean liguiyuan if-only1 wangq95 hzq-zjm

pytorch-mobilenet-v3's Issues

pretrained mobilenetv3_small_67.4.pth.tar file is broken

RT. cannot open this file as a .tar file

Training Setting

Hi, could you share your training setting for Large or Small models?

Thanks!

Dropout

the paper mentioned training with dropout. How would you train with dropout with this model?

Hyperparameters for training

Would you please provide the hyperparameters used for training of the models? Thanks

Last stage

In original paper paper ,There is no activation function after avg_Pool　in fig5, but your code is added. Does this have any effect?

作者能提供一下预训练模型吗？

请问一下，关于mobilenetV3-large 1.00的预训练模型能提供一下吗？

同学您好，请问下large的预训练网络您还在计划训练吗？谢谢！！

Where in your code you have used Depthwise Conv?

Where in your code you have used Depthwise Conv? I did not find any depthwise conv.

mobilenet_v3 small last channel is 1024 instead of 1280。

mobilenet_v3 small last channel is 1024 instead of 1280 from the paper，is there any evidence that 1280 brings better results?

could you please upload your large V3 model on google drive or dropbox??

I'm trying to train mobilenetV3 on DGX 1 using imagenet 2012 dataset but I found it's quite hard to reach to the reported accuracy.

I just got Acc-1 : 70.xx %

Humm.... can I ask you to send me or upload your mobilenetV3-large model on google drive/dropbox?

Thank you!

I implemented v3 and tested with/without the modified last stage settings. Although paper claims that this modification doesn't affect accuracy, I find out accuracy actually drops (i.e. 73.45->71.84). May I know about your opinion on that? Thank you!

Can u offer the training log？

img resize in Test phase

transforms.Resize(int(input_size/0.875) used in Test phase.
Why use it instead of transforms.RandomResizedCrop() used in Train phase?
We should be consistent in Train and Test phase.

a mistake in the paper

the output channel is mismatch with the output size ,maybe it should be 112.

Duplicated avg pool?

in the init, already defined AdaptiveAvgPool2d in the feature sequence. but in the forward method, there is another mean function. I'm a little confused... 😂😂