pretrainedmodels

# 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.

<a href=""><img src="></a>

News: - 27/10/2018: Fix compatibility issues, Add tests, Add travis - 04/06/2018: [PolyNet]() and [PNASNet-5-Large]() thanks to [Alex Parinov]() - 16/04/2018: [SE-ResNet* and SE-ResNeXt*]() thanks to [Alex Parinov]() - 09/04/2018: [SENet154]() thanks to [Alex Parinov]() - 22/03/2018: CaffeResNet101 (good for localization with FasterRCNN) - 21/03/2018: NASNet Mobile thanks to [Veronika Yurchuk]() and [Anastasiia]() - 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

• [Installation]()

• [Quick examples]()

• [Few use cases]()

  • [Compute imagenet logits]()
  • [Compute imagenet validation metrics]()

• [Evaluation on ImageNet]()

  • [Accuracy on valset]()
  • [Reproducing results]()

• [Documentation]()

  • [Available models]()

    • [AlexNet]()
    • [BNInception]()
    • [CaffeResNet101]()
    • [DenseNet121]()
    • [DenseNet161]()
    • [DenseNet169]()
    • [DenseNet201]()
    • [DenseNet201]()
    • [DualPathNet68]()
    • [DualPathNet92]()
    • [DualPathNet98]()
    • [DualPathNet107]()
    • [DualPathNet113]()
    • [FBResNet152]()
    • [InceptionResNetV2]()
    • [InceptionV3]()
    • [InceptionV4]()
    • [NASNet-A-Large]()
    • [NASNet-A-Mobile]()
    • [PNASNet-5-Large]()
    • [PolyNet]()
    • [ResNeXt101_32x4d]()
    • [ResNeXt101_64x4d]()
    • [ResNet101]()
    • [ResNet152]()
    • [ResNet18]()
    • [ResNet34]()
    • [ResNet50]()
    • [SENet154]()
    • [SE-ResNet50]()
    • [SE-ResNet101]()
    • [SE-ResNet152]()
    • [SE-ResNeXt50_32x4d]()
    • [SE-ResNeXt101_32x4d]()
    • [SqueezeNet1_0]()
    • [SqueezeNet1_1]()
    • [VGG11]()
    • [VGG13]()
    • [VGG16]()
    • [VGG19]()
    • [VGG11_BN]()
    • [VGG13_BN]()
    • [VGG16_BN]()
    • [VGG19_BN]()
    • [Xception]()

  • [Model API]()

    • [model.input_size]()
    • [model.input_space]()
    • [model.input_range]()
    • [model.mean]()
    • [model.std]()
    • [model.features]()
    • [model.logits]()
    • [model.forward]()

• [Reproducing porting]()

  • [ResNet*]()
  • [ResNeXt*]()
  • [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`:

`python import pretrainedmodels`

• To print the available pretrained models:

`python 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', 'nasnetamobile', 'cafferesnet101', 'senet154', 'se_resnet50', 'se_resnet101', 'se_resnet152', 'se_resnext50_32x4d', 'se_resnext101_32x4d', 'cafferesnet101', 'polynet', 'pnasnet5large']`

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

`python 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:

`python model_name = 'nasnetalarge' # could be fbresnet152 or inceptionresnetv2 model = 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:

`python import torch import pretrainedmodels.utils as utils load_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 differ input_tensor = input_tensor.unsqueeze(0) # 3x299x299 -> 1x3x299x299 input = 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):

`python output_features = model.features(input) # 1x14x14x2048 size may differ output_logits = model.logits(output_features) # 1x1000`

## Few use cases

### Compute imagenet logits

• See [examples/imagenet_logits.py]() to compute logits of classes appearance over a single image with a pretrained model on imagenet.

` $ 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

• See [examples/imagenet_eval.py]() to evaluate pretrained models on imagenet valset.

` $ 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 | | --- | --- | --- | ---PNASNet-5-Large | [Tensorflow]() | 82.858 | 96.182 [PNASNet-5-Large]() | Our porting | 82.736 | 95.992 NASNet-A-Large | [Tensorflow]() | 82.693 | 96.163 [NASNet-A-Large]() | Our porting | 82.566 | 96.086 SENet154 | [Caffe]() | 81.32 | 95.53 [SENet154]() | Our porting | 81.304 | 95.498 PolyNet | [Caffe]() | 81.29 | 95.75 [PolyNet]() | Our porting | 81.002 | 95.624 InceptionResNetV2 | [Tensorflow]() | 80.4 | 95.3 InceptionV4 | [Tensorflow]() | 80.2 | 95.3 [SE-ResNeXt101_32x4d]() | Our porting | 80.236 | 95.028 SE-ResNeXt101_32x4d | [Caffe]() | 80.19 | 95.04 [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 [SE-ResNeXt50_32x4d]() | Our porting | 79.076 | 94.434 SE-ResNeXt50_32x4d | [Caffe]() | 79.03 | 94.46 [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 SE-ResNet152 | [Caffe]() | 78.66 | 94.46 [SE-ResNet152]() | Our porting | 78.658 | 94.374 ResNet152 | [Pytorch]() | 78.428 | 94.110 [SE-ResNet101]() | Our porting | 78.396 | 94.258 SE-ResNet101 | [Caffe]() | 78.25 | 94.28 [ResNeXt101_32x4d]() | Our porting | 78.188 | 93.886 FBResNet152 | [Torch7]() | 77.84 | 93.84 SE-ResNet50 | [Caffe]() | 77.63 | 93.64 [SE-ResNet50]() | Our porting | 77.636 | 93.752 [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 [CaffeResnet101]() | [Caffe]() | 76.400 | 92.900 [CaffeResnet101]() | Our porting | 76.200 | 92.766 [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 NASNet-A-Mobile | [Tensorflow]() | 74.0 | 91.6 [NASNet-A-Mobile]() | Our porting | 74.080 | 91.740 [ResNet34]() | [Pytorch]() | 73.554 | 91.456 [BNInception]() | Our porting | 73.524 | 91.562 [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

Notes: - the Pytorch version of ResNet152 is not a porting of the Torch7 but has been retrained by facebook. - For the PolyNet evaluation each image was resized to 378x378 without preserving the aspect ratio and then the central 331×331 patch from the resulting image was used.

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! 😛

### 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')
• nasnetamobile(num_classes=1000, pretrained='imagenet')

#### 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')

#### Caffe ResNet*

Source: [Caffe repo of KaimingHe]()

• cafferesnet101(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')

#### SENet*

Source: [Caffe repo of Jie Hu]()

• senet154(num_classes=1000, pretrained='imagenet')
• se_resnet50(num_classes=1000, pretrained='imagenet')
• se_resnet101(num_classes=1000, pretrained='imagenet')
• se_resnet152(num_classes=1000, pretrained='imagenet')
• se_resnext50_32x4d(num_classes=1000, pretrained='imagenet')
• se_resnext101_32x4d(num_classes=1000, pretrained='imagenet')

#### PNASNet*

Source: [TensorFlow Slim repo]()

• pnasnet5large(num_classes=1000, pretrained='imagenet')
• pnasnet5large(num_classes=1001, pretrained='imagenet+background')

#### PolyNet

Source: [Caffe repo of the CUHK Multimedia Lab]()

• polynet(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`:

`python 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`:

`python 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.

`python # 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`:

`python 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 tuning dim_feats = model.last_linear.in_features # =2048 nb_classes = 4 model.last_linear = nn.Linear(dim_feats, nb_classes) output = model(input_224) print(output.size()) # (1,4) # features extraction model.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

### Hand porting of NASNet, InceptionV4 and InceptionResNetV2

## Acknowledgement

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