image.md

Image Modules

For image classification, two kinds of modules are available:

• Modules to do image classification with the particular classes that the module has been trained for.
• Modules to extract image feature vectors, (a.k.a. "bottleneck values") for use in custom image classifiers. (This is elaborated in the image retraining tutorial.)

Click on a module to view its documentation, or reference the URL from the TensorFlow Hub library like so:

m = hub.Module("https://tfhub.dev/...")

Modules trained on ImageNet (ILSVRC-2012-CLS)

Inception and Inception-ResNet

• Inception V1: classification, feature_vector.
• Inception V2: classification, feature_vector.
• Inception V3: classification, feature_vector.
• Inception-ResNet V2: classification, feature_vector.

MobileNet

MobileNets come in various sizes controlled by a multiplier for the depth (number of features), and trained for various sizes of input images. See the module documentation for details.

• MobileNet V1

  	224x224	192x192	160x160	128x128
  100%	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector
  75%	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector
  50%	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector
  25%	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector

• MobileNet V1 instrumented for quantization with TF-Lite ("/quantops")

  	224x224	192x192	160x160	128x128
  100%	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector
  75%	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector
  50%	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector
  25%	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector

• MobileNet V2

  	224x224	192x192	160x160	128x128	96x96
  140%	classification feature_vector
  130%	classification feature_vector
  100%	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector
  75%	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector
  50%	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector
  35%	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector	classification feature_vector

NASNet and PNASNet

• NASNet-A large: classification, feature_vector.
• NASNet-A mobile: classification, feature_vector.
• PNASNet-5 large: classification, feature_vector.

ResNet

• ResNet V1

  50 layers	101 layers	152 layers
  classification feature_vector	classification feature_vector	classification feature_vector

• ResNet V2

  50 layers	101 layers	152 layers
  classification feature_vector	classification feature_vector	classification feature_vector

Modules trained on domain-specific datasets

iNaturalist (iNat) 2017

The iNat2017 dataset consists of 579,184 training images and 95,986 validation images from 5,089 species, taken from www.inaturalist.org.

• Inception V3: feature_vector.