资源分类
》 资源 》算法》NASNet-A-Mobile. Ported weights

NASNet-A-Mobile. Ported weights

2019-09-18 | |  62 |   0 |   0

Pretrained models for Pytorch (Work in progress)

The goal of this repo is:

  • to help to reproduce research papers results (transfer learning setups for instance),

  • to access pretrained ConvNets with a unique interface/API inspired by torchvision.

News: - 21/03/2018: NASNet Mobile - 25/01/2018: DualPathNetworks thanks to Ross Wightman, Xception thanks to T Standley, improved TransformImage API - 13/01/2018: pip install pretrainedmodels, pretrainedmodels.model_names, pretrainedmodels.pretrained_settings - 12/01/2018: python setup.py install - 08/12/2017: update data url (/! git pull is needed) - 30/11/2017: improve API (model.features(input), model.logits(features), model.forward(input), model.last_linear) - 16/11/2017: nasnet-a-large pretrained model ported by T. Durand and R. Cadene - 22/07/2017: torchvision pretrained models - 22/07/2017: momentum in inceptionv4 and inceptionresnetv2 to 0.1 - 17/07/2017: model.input_range attribut - 17/07/2017: BNInception pretrained on Imagenet

Summary

  • Inception*

    • Installation

    1. python3 with anaconda

    2. pytorch with/out CUDA

    Install from pip

    1. pip install pretrainedmodels

    Install from repo

    1. git clone https://github.com/Cadene/pretrained-models.pytorch.git

    2. cd pretrained-models.pytorch

    3. python setup.py install

    Quick examples

    • To import pretrainedmodels:

    import pretrainedmodels

    • To print the available pretrained models:

    print(pretrainedmodels.model_names)> ['fbresnet152', 'bninception', 'resnext101_32x4d', 'resnext101_64x4d', 'inceptionv4', 'inceptionresnetv2', 'alexnet', 'densenet121', 'densenet169', 'densenet201', 'densenet161', 'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152', 'inceptionv3', 'squeezenet1_0', 'squeezenet1_1', 'vgg11', 'vgg11_bn', 'vgg13', 'vgg13_bn', 'vgg16', 'vgg16_bn', 'vgg19_bn', 'vgg19', 'nasnetalarge']

    • To print the available pretrained settings for a chosen model:

    print(pretrainedmodels.pretrained_settings['nasnetalarge'])> {'imagenet': {'url': 'http://data.lip6.fr/cadene/pretrainedmodels/nasnetalarge-a1897284.pth', 'input_space': 'RGB', 'input_size': [3, 331, 331], 'input_range': [0, 1], 'mean': [0.5, 0.5, 0.5], 'std': [0.5, 0.5, 0.5], 'num_classes': 1000}, 'imagenet+background': {'url': 'http://data.lip6.fr/cadene/pretrainedmodels/nasnetalarge-a1897284.pth', 'input_space': 'RGB', 'input_size': [3, 331, 331], 'input_range': [0, 1], 'mean': [0.5, 0.5, 0.5], 'std': [0.5, 0.5, 0.5], 'num_classes': 1001}}

    • To load a pretrained models from imagenet:

    model_name = 'nasnetalarge' # could be fbresnet152 or inceptionresnetv2model = pretrainedmodels.__dict__[model_name](num_classes=1000, pretrained='imagenet')model.eval()

    Note: By default, models will be downloaded to your $HOME/.torch folder. You can modify this behavior using the $TORCH_MODEL_ZOO variable as follow: export TORCH_MODEL_ZOO="/local/pretrainedmodels

    • To load an image and do a complete forward pass:

    import torchimport pretrainedmodels.utils as utilsload_img = utils.LoadImage()# transformations depending on the model#rescale, center crop, normalize, and others (ex: ToBGR, ToRange255)tf_img = utils.TransformImage(model) path_img = 'data/cat.jpg'input_img = load_img(path_img)input_tensor = tf_img(input_img)         # 3x400x225 -> 3x299x299 size may differinput_tensor = input_tensor.unsqueeze(0) # 3x299x299 -> 1x3x299x299input = torch.autograd.Variable(input_tensor,
    requires_grad=False)output_logits = model(input) # 1x1000

    • To extract features (beware this API is not available for all networks):

    output_features = model.features(input) # 1x14x14x2048 size may differoutput_logits = model.logits(output_features) # 1x1000

    Few use cases

    Compute imagenet logits

    $ python examples/imagenet_logits.py -h
> nasnetalarge, resnet152, inceptionresnetv2, inceptionv4, ...
    $ python examples/imagenet_logits.py -a nasnetalarge --path_img data/cat.png
> 'nasnetalarge': data/cat.png' is a 'tiger cat'

    Compute imagenet evaluation metrics

    $ python examples/imagenet_eval.py /local/common-data/imagenet_2012/images -a nasnetalarge -b 20 -e
> * Acc@1 92.693, Acc@5 96.13

    Evaluation on imagenet

    Accuracy on validation set (single model)

    Results were obtained using (center cropped) images of the same size than during the training process.

    Model | Version | Acc@1 | Acc@5 --- | --- | --- | --- NASNet-A-Mobile | Tensorflow | 74.0 | 91.6 NASNet-A-Mobile | Our porting | 74.08 | 91.74 NASNet-A-Large | Tensorflow | 82.693 | 96.163 NASNet-A-Large | Our porting | 82.566 | 96.086 InceptionResNetV2 | Tensorflow | 80.4 | 95.3 InceptionV4 | Tensorflow | 80.2 | 95.3 InceptionResNetV2 | Our porting | 80.170 | 95.234 InceptionV4 | Our porting | 80.062 | 94.926 DualPathNet107_5k | Our porting | 79.746 | 94.684 ResNeXt101_64x4d | Torch7 | 79.6 | 94.7 DualPathNet131 | Our porting | 79.432 | 94.574 DualPathNet92_5k | Our porting | 79.400 | 94.620 DualPathNet98 | Our porting | 79.224 | 94.488 Xception | Keras | 79.000 | 94.500 ResNeXt101_64x4d | Our porting | 78.956 | 94.252 Xception | Our porting | 78.888 | 94.292 ResNeXt101_32x4d | Torch7 | 78.8 | 94.4 ResNet152 | Pytorch | 78.428 | 94.110 ResNeXt101_32x4d | Our porting | 78.188 | 93.886 FBResNet152 | Torch7 | 77.84 | 93.84 DenseNet161 | Pytorch | 77.560 | 93.798 ResNet101 | Pytorch | 77.438 | 93.672 FBResNet152 | Our porting | 77.386 | 93.594 InceptionV3 | Pytorch | 77.294 | 93.454 DenseNet201 | Pytorch | 77.152 | 93.548 DualPathNet68b_5k | Our porting | 77.034 | 93.590 DenseNet169 | Pytorch | 76.026 | 92.992 ResNet50 | Pytorch | 76.002 | 92.980 DualPathNet68 | Our porting | 75.868 | 92.774 DenseNet121 | Pytorch | 74.646 | 92.136 VGG19_BN | Pytorch | 74.266 | 92.066 ResNet34 | Pytorch | 73.554 | 91.456 BNInception | Our porting | 73.522 | 91.560 VGG16_BN | Pytorch | 73.518 | 91.608 VGG19 | Pytorch | 72.080 | 90.822 VGG16 | Pytorch | 71.636 | 90.354 VGG13_BN | Pytorch | 71.508 | 90.494 VGG11_BN | Pytorch | 70.452 | 89.818 ResNet18 | Pytorch | 70.142 | 89.274 VGG13 | Pytorch | 69.662 | 89.264 VGG11 | Pytorch | 68.970 | 88.746 SqueezeNet1_1 | Pytorch | 58.250 | 80.800 SqueezeNet1_0 | Pytorch | 58.108 | 80.428 Alexnet | Pytorch | 56.432 | 79.194

    Note: the Pytorch version of ResNet152 is not a porting of the Torch7 but has been retrained by facebook.

    Beware, the accuracy reported here is not always representative of the transferable capacity of the network on other tasks and datasets. You must try them all! :P

    Reproducing results

    Please see Compute imagenet validation metrics

    Documentation

    Available models

    NASNet*

    Source: TensorFlow Slim repo

    • nasnetalarge(num_classes=1000, pretrained='imagenet')

    • nasnetalarge(num_classes=1001, pretrained='imagenet+background')

    FaceBook ResNet*

    Source: Torch7 repo of FaceBook

    There are a bit different from the ResNet* of torchvision. ResNet152 is currently the only one available.

    • fbresnet152(num_classes=1000, pretrained='imagenet')

    Inception*

    Source: TensorFlow Slim repo and Pytorch/Vision repo for inceptionv3

    • inceptionresnetv2(num_classes=1000, pretrained='imagenet')

    • inceptionresnetv2(num_classes=1001, pretrained='imagenet+background')

    • inceptionv4(num_classes=1000, pretrained='imagenet')

    • inceptionv4(num_classes=1001, pretrained='imagenet+background')

    • inceptionv3(num_classes=1000, pretrained='imagenet')

    BNInception

    Source: Trained with Caffe by Xiong Yuanjun

    • bninception(num_classes=1000, pretrained='imagenet')

    ResNeXt*

    Source: ResNeXt repo of FaceBook

    • resnext101_32x4d(num_classes=1000, pretrained='imagenet')

    • resnext101_62x4d(num_classes=1000, pretrained='imagenet')

    DualPathNetworks

    Source: MXNET repo of Chen Yunpeng

    The porting has been made possible by Ross Wightman in his PyTorch repo.

    As you can see here DualPathNetworks allows you to try different scales. The default one in this repo is 0.875 meaning that the original input size is 256 before croping to 224.

    • dpn68(num_classes=1000, pretrained='imagenet')

    • dpn98(num_classes=1000, pretrained='imagenet')

    • dpn131(num_classes=1000, pretrained='imagenet')

    • dpn68b(num_classes=1000, pretrained='imagenet+5k')

    • dpn92(num_classes=1000, pretrained='imagenet+5k')

    • dpn107(num_classes=1000, pretrained='imagenet+5k')

    'imagenet+5k' means that the network has been pretrained on imagenet5k before being finetuned on imagenet1k.

    Xception

    Source: Keras repo

    The porting has been made possible by T Standley.

    • xception(num_classes=1000, pretrained='imagenet')

    TorchVision

    Source: Pytorch/Vision repo

    (inceptionv3 included in Inception*)

    • resnet18(num_classes=1000, pretrained='imagenet')

    • resnet34(num_classes=1000, pretrained='imagenet')

    • resnet50(num_classes=1000, pretrained='imagenet')

    • resnet101(num_classes=1000, pretrained='imagenet')

    • resnet152(num_classes=1000, pretrained='imagenet')

    • densenet121(num_classes=1000, pretrained='imagenet')

    • densenet161(num_classes=1000, pretrained='imagenet')

    • densenet169(num_classes=1000, pretrained='imagenet')

    • densenet201(num_classes=1000, pretrained='imagenet')

    • squeezenet1_0(num_classes=1000, pretrained='imagenet')

    • squeezenet1_1(num_classes=1000, pretrained='imagenet')

    • alexnet(num_classes=1000, pretrained='imagenet')

    • vgg11(num_classes=1000, pretrained='imagenet')

    • vgg13(num_classes=1000, pretrained='imagenet')

    • vgg16(num_classes=1000, pretrained='imagenet')

    • vgg19(num_classes=1000, pretrained='imagenet')

    • vgg11_bn(num_classes=1000, pretrained='imagenet')

    • vgg13_bn(num_classes=1000, pretrained='imagenet')

    • vgg16_bn(num_classes=1000, pretrained='imagenet')

    • vgg19_bn(num_classes=1000, pretrained='imagenet')

    Model API

    Once a pretrained model has been loaded, you can use it that way.

    Important note: All image must be loaded using PIL which scales the pixel values between 0 and 1.

    model.input_size

    Attribut of type list composed of 3 numbers:

    • number of color channels,

    • height of the input image,

    • width of the input image.

    Example:

    • [3, 299, 299] for inception* networks,

    • [3, 224, 224] for resnet* networks.

    model.input_space

    Attribut of type str representating the color space of the image. Can be RGB or BGR.

    model.input_range

    Attribut of type list composed of 2 numbers:

    • min pixel value,

    • max pixel value.

    Example:

    • [0, 1] for resnet* and inception* networks,

    • [0, 255] for bninception network.

    model.mean

    Attribut of type list composed of 3 numbers which are used to normalize the input image (substract "color-channel-wise").

    Example:

    • [0.5, 0.5, 0.5] for inception* networks,

    • [0.485, 0.456, 0.406] for resnet* networks.

    model.std

    Attribut of type list composed of 3 numbers which are used to normalize the input image (divide "color-channel-wise").

    Example:

    • [0.5, 0.5, 0.5] for inception* networks,

    • [0.229, 0.224, 0.225] for resnet* networks.

    model.features

    /! work in progress (may not be available)

    Method which is used to extract the features from the image.

    Example when the model is loaded using fbresnet152:

    print(input_224.size())            # (1,3,224,224)output = model.features(input_224) print(output.size())               # (1,2048,1,1)# print(input_448.size())          # (1,3,448,448)output = model.features(input_448)# print(output.size())             # (1,2048,7,7)

    model.logits

    /! work in progress (may not be available)

    Method which is used to classify the features from the image.

    Example when the model is loaded using fbresnet152:

    output = model.features(input_224) print(output.size())               # (1,2048, 1, 1)output = model.logits(output)print(output.size())               # (1,1000)

    model.forward

    Method used to call model.features and model.logits. It can be overwritten as desired.

    Note: A good practice is to use model.__call__ as your function of choice to forward an input to your model. See the example bellow.

    # Without model.__call__output = model.forward(input_224)print(output.size())      # (1,1000)# With model.__call__output = model(input_224)print(output.size())      # (1,1000)

    model.last_linear

    Attribut of type nn.Linear. This module is the last one to be called during the forward pass.

    • Can be replaced by an adapted nn.Linear for fine tuning.

    • Can be replaced by pretrained.utils.Identity for features extraction.

    Example when the model is loaded using fbresnet152:

    print(input_224.size())            # (1,3,224,224)output = model.features(input_224) print(output.size())               # (1,2048,1,1)output = model.logits(output)print(output.size())               # (1,1000)# fine tuningdim_feats = model.last_linear.in_features # =2048nb_classes = 4model.last_linear = nn.Linear(dim_feats, nb_classes)output = model(input_224)print(output.size())               # (1,4)# features extractionmodel.last_linear = pretrained.utils.Identity()output = model(input_224)print(output.size())               # (1,2048)

    Reproducing

    Hand porting of ResNet152

    th pretrainedmodels/fbresnet/resnet152_dump.lua
python pretrainedmodels/fbresnet/resnet152_load.py

    Automatic porting of ResNeXt

    https://github.com/clcarwin/convert_torch_to_pytorch

    Hand porting of NASNet, InceptionV4 and InceptionResNetV2

    https://github.com/Cadene/tensorflow-model-zoo.torch

    Acknowledgement

    Thanks to the deep learning community and especially to the contributers of the pytorch ecosystem.


上一篇：deep-dream-in-pytorch

下一篇：FFTNet

用户评价
全部评价

热门资源

  • Keras-ResNeXt

    Keras ResNeXt Implementation of ResNeXt models...

  • seetafaceJNI

    项目介绍 基于中科院seetaface2进行封装的JAVA...

  • spark-corenlp

    This package wraps Stanford CoreNLP annotators ...

  • capsnet-with-caps...

    CapsNet with capsule-wise convolution Project ...

  • inferno-boilerplate

    This is a very basic boilerplate example for pe...