ImageNet Models

This subpackage provides a variety of pre-trained state-of-the-art models which is trained on ImageNet dataset.

The pre-trained models can be used for both inference and training as following:

# Create ResNet-50 for inference
import nnabla as nn
import nnabla.functions as F
import nnabla.parametric_functions as PF
import numpy as np
from nnabla.models.imagenet import ResNet50
model = ResNet50()
batch_size = 1
# model.input_shape returns (3, 224, 224) when ResNet-50
x = nn.Variable((batch_size,) + model.input_shape)
y = model(x, training=False)

# Execute inference
# Load input image as uint8 array with shape of (3, 224, 224)
from nnabla.utils.image_utils import imread
img = imread('example.jpg', size=model.input_shape[1:], channel_first=True)
x.d[0] = img
y.forward()
predicted_label = np.argmax(y.d[0])
print('Predicted label:', model.category_names[predicted_label])


# Create ResNet-50 for fine-tuning
batch_size=32
x = nn.Variable((batch_size,) + model.input_shape)
# * By training=True, it sets batch normalization mode for training
#   and gives trainable attributes to parameters.
# * By use_up_to='pool', it creats a network up to the output of
#   the final global average pooling.
pool = model(x, training=True, use_up_to='pool')

# Add a classification layer for another 10 category dataset
# and loss function
num_classes = 10
y = PF.affine(pool, num_classes, name='classifier10')
t = nn.Variable((batch_size, 1))
loss = F.sum(F.softmax_cross_entropy(y, t))

# Training...

Available models are summarized in the following table. Error rates are calculated using single center crop.

Available ImageNet models

Name	Class	Top-1 error	Top-5 error	Trained by/with
ResNet-18	ResNet18	30.28	10.90	Neural Network Console
ResNet-34	ResNet34	26.72	8.89	Neural Network Console
ResNet-50	ResNet50	24.59	7.48	Neural Network Console
ResNet-101	ResNet101	23.81	7.01	Neural Network Console
ResNet-152	ResNet152	23.48	7.09	Neural Network Console
MobileNet	MobileNet	29.51	10.34	Neural Network Console
MobileNetV2	MobileNetV2	29.94	10.82	Neural Network Console
SENet-154	SENet	22.04	6.29	Neural Network Console
SqueezeNet v1.0	SqueezeNetV10	42.71	20.12	Neural Network Console
SqueezeNet v1.1	SqueezeNetV11	41.23	19.18	Neural Network Console
VGG-11	VGG11	30.85	11.38	Neural Network Console
VGG-13	VGG13	29.51	10.46	Neural Network Console
VGG-16	VGG16	29.03	10.07	Neural Network Console
NIN	NIN	42.91	20.66	Neural Network Console
DenseNet-161	DenseNet	23.82	7.02	Neural Network Console
InceptionV3	InceptionV3	21.82	5.88	Neural Network Console
Xception	Xception	23.59	6.91	Neural Network Console
GoogLeNet	GoogLeNet	31.22	11.34	Neural Network Console
ResNeXt-50	ResNeXt50	22.95	6.73	Neural Network Console
ResNeXt-101	ResNeXt101	22.80	6.74	Neural Network Console
ShuffleNet	ShuffleNet10	34.15	13.85	Neural Network Console
ShuffleNet-0.5x	ShuffleNet05	41.99	19.64	Neural Network Console
ShuffleNet-2.0x	ShuffleNet20	30.34	11.12	Neural Network Console

Common interfaces

class nnabla.models.imagenet.base.ImageNetBase

Most of ImageNet pretrained models are inherited from this class so that it provides some common interfaces.

__call__(input_var=None, use_from=None, use_up_to='classifier', training=False, force_global_pooling=False, check_global_pooling=True, returns_net=False, verbose=0)

Create a network (computation graph) from a loaded model.

Parameters

• input_var (Variable, optional) – If given, input variable is replaced with the given variable and a network is constructed on top of the variable. Otherwise, a variable with batch size as 1 and a default shape from self.input_shape.

• use_up_to (str) – Network is constructed up to a variable specified by a string. A list of string-variable correspondences in a model is described in documentation for each model class.

• training (bool) – This option enables additional training (fine-tuning, transfer learning etc.) for the constructed network. If True, the batch_stat option in batch normalization is turned True, and need_grad attribute in trainable variables (conv weights and gamma and beta of bn etc.) is turned True. The default is False.

• force_global_pooling (bool) – Regardless the input image size, the final average pooling before classification layer will be automatically transformed to a global average pooling. The default is False.

• check_global_pooling (bool) – If True, and if the stride configuration of the final average pooling is not for global pooling, it raises an exception. The default is True. Use False when user want to do the pooling with the trained stride (7, 7) regardless the input spatial size.

• returns_net (bool) – When True, it returns a NnpNetwork object. Otherwise, It only returns the last variable of the constructed network. The default is False.

• verbose (bool, or int) – Verbose level. With 0, it says nothing during network construction.

property category_names

Returns category names of 1000 ImageNet classes.

property input_shape

Should returns default image size (channel, height, width) as a tuple.

List of models

class nnabla.models.imagenet.ResNet18

An alias of ResNet (18).

class nnabla.models.imagenet.ResNet34

An alias of ResNet (34).

class nnabla.models.imagenet.ResNet50

An alias of ResNet (50).

class nnabla.models.imagenet.ResNet101

An alias of ResNet (101).

class nnabla.models.imagenet.ResNet152

An alias of ResNet (152).

class nnabla.models.imagenet.ResNet(num_layers=18)

ResNet architectures for 18, 34, 50, 101, and 152 of number of layers.

Parameters

num_layers (int) – Number of layers chosen from 18, 34, 50, 101, and 152.

The following is a list of string that can be specified to use_up_to option in __call__ method;

• 'classifier' (default): The output of the final affine layer for classification.

• 'pool': The output of the final global average pooling.

• 'lastconv': The input of the final global average pooling without ReLU activation.

• 'lastconv+relu': Network up to 'lastconv' followed by ReLU activation.

References

• He et al, Deep Residual Learning for Image Recognition.

class nnabla.models.imagenet.MobileNet

MobileNet architecture.

The following is a list of string that can be specified to use_up_to option in __call__ method;

• 'classifier' (default): The output of the final affine layer for classification.

• 'pool': The output of the final global average pooling.

• 'lastconv': The input of the final global average pooling without ReLU activation.

• 'lastconv+relu': Network up to 'lastconv' followed by ReLU activation.

References

• Howard et al., MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications.

class nnabla.models.imagenet.MobileNetV2

MobileNetV2 architecture.

The following is a list of string that can be specified to use_up_to option in __call__ method;

• 'classifier' (default): The output of the final affine layer for classification.

• 'pool': The output of the final global average pooling.

• 'lastconv': The input of the final global average pooling without ReLU activation.

• 'lastconv+relu': Network up to 'lastconv' followed by ReLU activation.

References

• Sandelr et al., MobileNetV2: Inverted Residuals and Linear Bottlenecks.

class nnabla.models.imagenet.SENet

SENet-154 model which integrates SE blocks with a modified ResNeXt architecture.

The following is a list of string that can be specified to use_up_to option in __call__ method;

• 'classifier' (default): The output of the final affine layer for classification.

• 'pool': The output of the final global average pooling.

• 'lastconv': The input of the final global average pooling without ReLU activation.

• 'lastconv+relu': Network up to 'lastconv' followed by ReLU activation.

References

• Hu et al., Squeeze-and-Excitation Networks.

class nnabla.models.imagenet.SqueezeNetV10

SquezeNetV10 An alias of SqueezeNet ('v1.0').

class nnabla.models.imagenet.SqueezeNetV11

SquezeNetV11 An alias of SqueezeNet ('v1.1').

class nnabla.models.imagenet.SqueezeNet(version='v1.1')

SqueezeNet model for architecture-v1.0 and v1.1 .

Parameters

version (str) – Version chosen from ‘v1.0’ and ‘v1.1’.

The following is a list of string that can be specified to use_up_to option in __call__ method;

• 'classifier' (default): The output of the final affine layer for classification.

• 'pool': The output of the final global average pooling.

• 'lastconv': The input of the final global average pooling without ReLU activation.

• 'lastconv+relu': Network up to 'lastconv' followed by ReLU activation.

References

• Iandola, Forrest N. et al., SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size.

• Iandola, Forrest N. et al., SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <1MB model size.

• DeepScale/SqueezeNet on GitHub

class nnabla.models.imagenet.VGG11

An alias of VGG (11).

class nnabla.models.imagenet.VGG13

An alias of VGG (13).

class nnabla.models.imagenet.VGG16

An alias of VGG (16).

class nnabla.models.imagenet.VGG(num_layers=11)

VGG architectures for 11, 13, 16 layers.

Parameters

num_layers (int) – Number of layers chosen from 11, 13, 16.

The following is a list of string that can be specified to use_up_to option in __call__ method;

• 'classifier' (default): The output of the final affine layer for classification.

• 'pool': The output of the final global average pooling.

• 'lastconv': The input of the final global average pooling without ReLU activation.

• 'lastconv+relu': Network up to 'lastconv' followed by ReLU activation.

• 'lastfeature': Network up to one layer before 'classifier', but without activation.

References

• Simonyan and Zisserman, Very Deep Convolutional Networks for Large-Scale Image Recognition.

class nnabla.models.imagenet.NIN

NIN(Network In Network) architecture.

The following is a list of string that can be specified to use_up_to option in __call__ method;

• 'classifier' (default): The output of the final affine layer for classification.

• 'pool': The output of the final global average pooling.

• 'lastconv': The input of the final global average pooling without ReLU activation.

• 'lastconv+relu': Network up to 'lastconv' followed by ReLU activation.

References

• Lin et al., Network In Network.

class nnabla.models.imagenet.DenseNet

The following is a list of string that can be specified to use_up_to option in __call__ method;

• 'classifier' (default): The output of the final affine layer for classification.

• 'pool': The output of the final global average pooling.

• 'lastconv': The output from last denseblock.

• 'lastconv+relu': Network up to 'lastconv' followed by ReLU activation.

References

• Huang et al., Densely Connected Convolutional Networks.

class nnabla.models.imagenet.InceptionV3

InceptionV3 architecture.

The following is a list of string that can be specified to use_up_to option in __call__ method;

• 'classifier' (default): The output of the final affine layer for classification.

• 'pool': The output of the final global average pooling.

• 'prepool': The input of the final global average pooling, i.e. the output of the final inception block.

References

• Szegedy et al., Rethinking the Inception Architecture for Computer Vision.

class nnabla.models.imagenet.Xception

Xception model.

The following is a list of string that can be specified to use_up_to option in __call__ method;

• 'classifier' (default): The output of the final affine layer for classification.

• 'pool': The output of the final global average pooling.

• 'lastconv': The input of the final global average pooling without ReLU activation.

• 'lastconv+relu': Network up to 'lastconv' followed by ReLU activation.

References

• Francois Chollet, Xception: Deep Learning with Depthwise Separable Convolutions.

class nnabla.models.imagenet.GoogLeNet

GoogLeNet model.

The following is a list of string that can be specified to use_up_to option in __call__ method;

• 'classifier' (default): The output of the final affine layer for classification.

• 'pool': The output of the final global average pooling.

• 'prepool': The input of the final global average pooling, i.e. the output of the final inception block.

References

• Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, Andrew Rabinovich: Going Deeper with Convolutions.

class nnabla.models.imagenet.ResNeXt50

An alias of ResNeXt (50).

class nnabla.models.imagenet.ResNeXt101

An alias of ResNeXt (101).

class nnabla.models.imagenet.ResNeXt(num_layers=50)

ResNeXt architectures for 50 and 101 of number of layers.

Parameters

num_layers (int) – Number of layers chosen from 50 and 101.

The following is a list of string that can be specified to use_up_to option in __call__ method;

• 'classifier' (default): The output of the final affine layer for classification.

• 'pool': The output of the final global average pooling.

• 'lastconv': The input of the final global average pooling without ReLU activation.

• 'lastconv+relu': Network up to 'lastconv' followed by ReLU activation.

References

• Xie, Girshick. et al., Aggregated Residual Transformations for Deep Neural Networks.

• ResNeXt: Aggregated Residual Transformations for Deep Neural Networks on GitHub

class nnabla.models.imagenet.ShuffleNet10

An alias of ShuffleNet (10).

class nnabla.models.imagenet.ShuffleNet05

An alias of ShuffleNet (5).

class nnabla.models.imagenet.ShuffleNet20

An alias of ShuffleNet (20).

class nnabla.models.imagenet.ShuffleNet(scaling_factor=10)

Model for architecture ShuffleNet, ShuffleNet-0.5x and ShufffleNet-2.0x.

Parameters

Factor (Scaling) – To customize the network to a desired complexity, we can simply apply a scale factor on the number of channnels. This can be chosen from ‘10’, ‘5’ and ‘20’.

The following is a list of string that can be specified to use_up_to option in __call__ method;

• 'classifier' (default): The output of the final affine layer for classification.

• 'pool': The output of the final global average pooling.

• 'lastconv': The input of the final global average pooling without ReLU activation.

• 'lastconv+relu': Network up to 'lastconv' followed by ReLU activation.

References

• Zhang. et al., ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices.

• Ma, Zhang. et al., ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design.