chengyangfu / pytorch-vgg-cifar10 Goto Github PK

View Code? Open in Web Editor NEW

338.0 7.0 117.0 214 KB

This is the PyTorch implementation of VGG network trained on CIFAR10 dataset

License: MIT License

Python 12.84% Jupyter Notebook 86.63% Shell 0.53%

pytorch-vgg-cifar10's Introduction

pytorch-vgg-cifar10

This is the PyTorch implementation of VGG network trained on CIFAR10 dataset

Requirements.

[PyTorch] (https://github.com/pytorch/pytorch)

[torchvision] (https://github.com/pytorch/vision)

Update.

Adding support for CPU. Add --cpu can make the training or evaluation in cpu mode.

Download the model

The trained VGG model. 92.4% Accuracy VGG

Evaluation

# CUDA
wget http://www.cs.unc.edu/~cyfu/cifar10/model_best.pth.tar
python main.py --resume=./model_best.pth.tar -e
# or use CPU version
wget http://www.cs.unc.edu/~cyfu/cifar10/model_best_cpu.pth.tar
python main.py --resume=./model_best_cpu.pth.tar -e --cpu

Train with script! (16-bit precision)

./run.sh

Using the run.sh script to generate the training log and models of different versions of VGG in 16-bit or 32-bit precision. Then use the ipython notebook plot.ipynb to view the results.

pytorch-vgg-cifar10's People

Stargazers

Watchers

Forkers

danzschulman wensiding evanmy yangbo joshuaebenezer riceric22 niluanwudidadi tangbohu sufeidechabei adolhung lemonqc huipengzhang changyong-oh waimaoer happyconan srkiyengar mingyangyi wuzhenjiexd myhub totoro97 punsach whudonggu log4me zhizhuozhang sfarkya slowbull yanjiegao 666dzy666 dutzhangke jeffjunzhang rameshwar-wiai pingaowang csyhhu ttang2322 lgsaber hotndy albertghtoun onursahin93 wuzhongwulidong shuo0108 wojiaoyanmin sallyhuangimperial cat-white guobbin yuzihu yluo39github zhoushanglin100 bmlh haha-533 zifanw peiyangl milkigit assyriacross13 yejing-lai subramanian1990 fabienvorpe cadmium-cd caomy7 dsoselia tgeller08 briardoty yuhaozh88 harmandotpy can-song lzjpaul sihyunyoo yukewang96 renyiryry thhhomas wei-ao sarthakpati phy12321 ameenali jeremiemelo yurongchen1998 debangliu stjordanis qi-n-qigt ankita-shukla wangyunjeff twinkle-jc junfenggo zhangyuqaq myunghaklee xiangyu-gao syll-star sssupertian jerry-lizj xdg2016 debesheedas jwhita14 zhangconghhh jawoos zebfff mominabbass aorvieto seanpresent dr-lingxiao jushelf vicramr

pytorch-vgg-cifar10's Issues

What is the model_best.pth file's model?

I need Vgg16 pretrained model.
but there is error like this.

RuntimeError: Error(s) in loading state_dict for DataParallel:
Missing key(s) in state_dict: "module.features.0.bias", "module.features.0.weight", "module.features.2.bias", "module.features.2.weight", "module.features.5.bias", "module.features.5.weight", "module.features.7.bias", "module.features.7.weight", "module.features.10.bias", "module.features.10.weight", "module.features.12.bias", "module.features.12.weight", "module.features.14.bias", "module.features.14.weight", "module.features.17.bias", "module.features.17.weight", "module.features.19.bias", "module.features.19.weight", "module.features.21.bias", "module.features.21.weight", "module.features.24.bias", "module.features.24.weight", "module.features.26.bias", "module.features.26.weight", "module.features.28.bias", "module.features.28.weight", "module.classifier.1.bias", "module.classifier.1.weight", "module.classifier.4.bias", "module.classifier.4.weight", "module.classifier.6.bias", "module.classifier.6.weight".
Unexpected key(s) in state_dict: "state_dict", "best_prec1", "epoch".

What is the exact model?

License?

Hi Cheng-Yang,

Thanks for the CIFAR10 training code.

I wonder whether you would be interested to please license this code under an open source license? See e.g. the licenses here:

https://opensource.org/licenses

If so, then others will be more likely to contribute to your codebase, or incorporate it into their projects.

I run your code and always nan loss, can you help me?

I just run with ./run.sh and got nan loss after a few steps. Here is the printed log:

(base) root@For-Judy-And-Ian:~/pytorchProjects/pytorch-vgg-cifar10-master# ./run.sh
python main.py --arch=vgg11 --save-dir=save_vgg11 |& tee -a log_vgg11
Files already downloaded and verified
Epoch: [0 ][ 0 /391] Time 0.831 (0.831) Data 0.190 (0.190) Loss 2.3037 (2.3037) Prec@1 10.938 (10.938)
Epoch: [0 ][20 /391] Time 0.018 (0.052) Data 0.000 (0.009) Loss 2.2982 (2.3029) Prec@1 9.375 (9.487)
Epoch: [0 ][40 /391] Time 0.012 (0.035) Data 0.000 (0.005) Loss 2.2928 (2.3018) Prec@1 12.500 (9.546)
Epoch: [0 ][60 /391] Time 0.012 (0.030) Data 0.000 (0.003) Loss 2.2685 (2.2970) Prec@1 15.625 (10.720)
Epoch: [0 ][80 /391] Time 0.012 (0.027) Data 0.000 (0.002) Loss 2.1417 (2.2787) Prec@1 21.875 (11.960)
Epoch: [0 ][100/391] Time 0.016 (0.026) Data 0.000 (0.002) Loss 2.1417 (2.2518) Prec@1 22.656 (13.134)
Epoch: [0 ][120/391] Time 0.029 (0.024) Data 0.000 (0.002) Loss 1.9975 (2.2189) Prec@1 21.094 (14.463)
Epoch: [0 ][140/391] Time 0.028 (0.024) Data 0.000 (0.002) Loss 2.0889 (2.1959) Prec@1 26.562 (15.459)
Epoch: [0 ][160/391] Time 0.018 (0.023) Data 0.000 (0.001) Loss 2.0179 (2.1856) Prec@1 21.875 (16.193)
Epoch: [0 ][180/391] Time 0.012 (0.023) Data 0.000 (0.001) Loss 1.9825 (2.1645) Prec@1 25.000 (16.894)
Epoch: [0 ][200/391] Time 0.012 (0.022) Data 0.000 (0.001) Loss 1.8724 (2.1434) Prec@1 27.344 (17.623)
Epoch: [0 ][220/391] Time 0.012 (0.022) Data 0.000 (0.001) Loss 2.0147 (2.1258) Prec@1 25.000 (18.121)
Epoch: [0 ][240/391] Time 0.012 (0.021) Data 0.000 (0.001) Loss 1.8679 (2.1128) Prec@1 22.656 (18.458)
Epoch: [0 ][260/391] Time 0.016 (0.021) Data 0.000 (0.001) Loss 1.8262 (2.0923) Prec@1 28.125 (19.202)
Epoch: [0 ][280/391] Time 0.012 (0.021) Data 0.000 (0.001) Loss 1.7779 (2.0737) Prec@1 31.250 (19.834)
Epoch: [0 ][300/391] Time 0.011 (0.020) Data 0.000 (0.001) Loss 1.7415 (2.0569) Prec@1 38.281 (20.359)
Epoch: [0 ][320/391] Time 0.012 (0.020) Data 0.000 (0.001) Loss 1.7895 (2.0431) Prec@1 26.562 (20.863)
Epoch: [0 ][340/391] Time 0.012 (0.020) Data 0.000 (0.001) Loss 1.7198 (2.0292) Prec@1 31.250 (21.355)
Epoch: [0 ][360/391] Time 0.012 (0.019) Data 0.000 (0.001) Loss 1.9042 (2.0171) Prec@1 27.344 (21.827)
Epoch: [0 ][380/391] Time 0.012 (0.019) Data 0.000 (0.001) Loss 2.6430 (2.0338) Prec@1 12.500 (21.900)
Test[0/79] Time 0.136 (0.136) Loss 2.3228 (2.3228) Prec@1 10.938 (10.938)
Test[20/79] Time 0.004 (0.013) Loss 2.3267 (2.3337) Prec@1 7.812 (8.891)
Test[40/79] Time 0.013 (0.009) Loss 2.3235 (2.3322) Prec@1 10.156 (8.670)
Test[60/79] Time 0.011 (0.009) Loss 2.3311 (2.3303) Prec@1 10.156 (8.799)
* Prec@1 8.810
Epoch: [1 ][ 0 /391] Time 0.099 (0.099) Data 0.085 (0.085) Loss 2.3538 (2.3538) Prec@1 8.594 (8.594)
Epoch: [1 ][20 /391] Time 0.028 (0.021) Data 0.000 (0.005) Loss nan (nan) Prec@1 1.562 (8.036)
Epoch: [1 ][40 /391] Time 0.018 (0.019) Data 0.000 (0.003) Loss nan (nan) Prec@1 1.562 (5.011)
Epoch: [1 ][60 /391] Time 0.012 (0.017) Data 0.000 (0.002) Loss nan (nan) Prec@1 1.562 (3.893)
Epoch: [1 ][80 /391] Time 0.012 (0.016) Data 0.000 (0.002) Loss nan (nan) Prec@1 2.344 (3.279)
Epoch: [1 ][100/391] Time 0.013 (0.016) Data 0.002 (0.001) Loss nan (nan) Prec@1 2.344 (2.908)
Epoch: [1 ][120/391] Time 0.017 (0.015) Data 0.000 (0.001) Loss nan (nan) Prec@1 3.906 (2.686)
Epoch: [1 ][140/391] Time 0.012 (0.016) Data 0.000 (0.001) Loss nan (nan) Prec@1 2.344 (2.549)
Epoch: [1 ][160/391] Time 0.012 (0.015) Data 0.000 (0.001) Loss nan (nan) Prec@1 2.344 (2.451)
Epoch: [1 ][180/391] Time 0.017 (0.016) Data 0.000 (0.001) Loss nan (nan) Prec@1 3.906 (2.348)
Epoch: [1 ][200/391] Time 0.012 (0.016) Data 0.000 (0.001) Loss nan (nan) Prec@1 2.344 (2.320)
Epoch: [1 ][220/391] Time 0.011 (0.015) Data 0.000 (0.001) Loss nan (nan) Prec@1 0.781 (2.238)
Epoch: [1 ][240/391] Time 0.012 (0.015) Data 0.000 (0.001) Loss nan (nan) Prec@1 0.781 (2.217)
Epoch: [1 ][260/391] Time 0.013 (0.015) Data 0.000 (0.001) Loss nan (nan) Prec@1 1.562 (2.176)
Epoch: [1 ][280/391] Time 0.012 (0.015) Data 0.000 (0.001) Loss nan (nan) Prec@1 2.344 (2.149)
Epoch: [1 ][300/391] Time 0.016 (0.015) Data 0.000 (0.001) Loss nan (nan) Prec@1 1.562 (2.108)
Epoch: [1 ][320/391] Time 0.018 (0.015) Data 0.007 (0.001) Loss nan (nan) Prec@1 0.781 (2.078)
Epoch: [1 ][340/391] Time 0.017 (0.015) Data 0.006 (0.001) Loss nan (nan) Prec@1 3.125 (2.067)
Epoch: [1 ][360/391] Time 0.018 (0.015) Data 0.006 (0.001) Loss nan (nan) Prec@1 3.906 (2.052)
Epoch: [1 ][380/391] Time 0.012 (0.016) Data 0.000 (0.001) Loss nan (nan) Prec@1 0.781 (2.010)
Test[0/79] Time 0.094 (0.094) Loss nan (nan) Prec@1 0.000 (0.000)
Test[20/79] Time 0.009 (0.014) Loss nan (nan) Prec@1 0.000 (0.335)
Test[40/79] Time 0.015 (0.015) Loss nan (nan) Prec@1 0.000 (0.419)
Test[60/79] Time 0.015 (0.015) Loss nan (nan) Prec@1 0.000 (0.538)
* Prec@1 0.540

The version of Pytorch and torchvision

Please pardon my poor English.
And I wonder what's the version of pytorch and torchvision.
I have met some error as follow:

RuntimeError: not allowed to set torch.backends.cudnn flags after disable_global_flags; please use flags() context manager instead

I have't figure out what's the cause ...

Should it be decay by 2?

pytorch-vgg-cifar10/main.py

Line 271 in 2a41b25

"""Sets the learning rate to the initial LR decayed by 10 every 30 epochs"""

which pytorch version did you use to train this model? i am trying to do torch. load on my CPU but even after i give map_location cpu, it still gives pytorch not compiled for CUDA enabled assertion error

How is the data divided among workers

Hello, I want to know if the data is divided equally among all workers?

Question regarding VGG cfg

Sorry I have a stupid question.
in the cfg: [64, 'M', 128, 'M', 256, 256, 'M', 512, 512, 'M', 512, 512, 'M'],

what 'M' stands for?

Why not giving the whole model to DataParallel ？

I feel confused about the code in main function as follows :
model.features = torch.nn.DataParallel(model.features)
May I ask that why not giving the whole model to DataParallel ？

I use your code the result all are 100 or 99.9?

I use the main.py to train cifar10 ,the result is incredible, the test acc is100% or 99.9%,

Is your code based on a code template?

I once saw a code on AlexNet that is very similar to your coding style（https://github.com/jiecaoyu/pytorch_imagenet）...
May I ask that is your code based on a standard code template?

Oversampling in CIFAR-10

Hi,

I need to train this model (VGG) using some oversampling technique in the Cifar-10 database. But I do not know how to do it in Pytorch.
I want to simulate a dataset of real world, since in the real world the classes are unbalanced. However, I need to adjust the network to learn the unbalanced classes. So first I need to simulate the problem of class imbalance at the dataset, because CIFAR-10 is a balanced dataset. And then apply some oversampling technique. Could you give me an example?