Common Signatures for Images

This page describes common signatures that should be implemented by modules for image-related tasks.

Some modules can be used for more than one task (e.g., image classification modules tend do to some feature extraction on the way). Therefore, each module provides (1) named signatures for all the tasks anticipated by the publisher, and (2) a default signature output = m(images) for its designated primary task.

Image Feature Vector

Usage summary

An image feature vector is a dense 1-D tensor that represents a whole image, typically for classification by the consumer model. (Unlike the intermediate activations of CNNs, it does not offer a spatial breakdown. Unlike image classification, it discards the classification learned by the publisher model.)

A module for image feature extraction has a default signature that maps a batch of images to a batch of feature vectors. It can be used like so:

  module_spec = hub.load_module_spec("path/to/module")
  height, width = hub.get_expected_image_size(module_spec)
  images = ...  # A batch of images with shape [batch_size, height, width, 3].
  module = hub.Module(module_spec)
  features = module(images)   # A batch with shape [batch_size, num_features].

It also defines the corresponding named signature.

Signature specification

The named signature for extracting image feature vectors is invoked as

  outputs = module(dict(images=images), signature="image_feature_vector",
                   as_dict=True)
  features = outputs["default"]

The input follows the general convention for input of images.

The outputs dictionary contains a "default" output of dtype float32 and shape [batch_size, num_features]. The batch_size is the same as in the input, but not known at graph construction time. num_features is a known, module-specific constant independent of input size.

These feature vectors are meant to be usable for classification with a simple feed-forward classifier (like the pooled features from the topmost convolutional layer in a typical CNN for image classification).

Applying dropout to the output features (or not) should be left to the module consumer. The module itself should not perform dropout on the actual outputs (even if it uses dropout internally in other places).

The outputs dictionary may provide further outputs, for example, the activations of hidden layers inside the module. Their keys and values are module-dependent. It is recommended to prefix architecture-dependent keys with an architecture name (e.g., to avoid confusing the intermediate layer "InceptionV3/Mixed_5c" with the topmost convolutional layer "InceptionV2/Mixed_5c").

Image Classification

Usage summary

Image classification maps the pixels of an image to linear scores (logits) for membership in the classes of a taxonomy selected by the module publisher. This allows consumers to draw conclusions from the particular classification learned by the publisher module, and not just its underlying features (cf. Image Feature Vector).

A module for image feature extraction has a default signature that maps a batch of images to a batch of logits. It can be used like so:

  module_spec = hub.load_module_spec("path/to/module")
  height, width = hub.get_expected_image_size(module_spec)
  images = ...  # A batch of images with shape [batch_size, height, width, 3].
  module = hub.Module(module_spec)
  logits = module(images)   # A batch with shape [batch_size, num_classes].

It also defines the corresponding named signature.

Signature specification

The named signature for extracting image feature vectors is invoked as

  outputs = module(dict(images=images), signature="image_classification",
                   as_dict=True)
  logits = outputs["default"]

The input follows the general convention for input of images.

The outputs dictionary contains a "default" output of dtype float32 and shape [batch_size, num_classes]. The batch_size is the same as in the input, but not known at graph construction time. num_classes is the number of classes in the classification, which is a known constant independent of input size.

Evaluating outputs["default"][i, c] yields a score predicting the membership of example i in the class with index c.

It depends on the underlying classification whether these scores are meant to be used with softmax (for mutually exclusive classes), sigmoid (for orthogonal classes), or something else. The module documentation should describe this, and refer to a definition of the class indices.

The outputs dictionary may provide further outputs, for example, the activations of hidden layers inside the module. Their keys and values are module-dependent. It is recommended to prefix architecture-dependent keys with an architecture name (e.g., to avoid confusing the intermediate layer "InceptionV3/Mixed_5c" with the topmost convolutional layer "InceptionV2/Mixed_5c").

Image input

This is common to all types of image modules and image signatures.

A signature that takes a batch of images as input accepts them as a dense 4-D tensor of dtype float32 and shape [batch_size, height, width, 3] whose elements are RGB color values of pixels normalized to the range [0, 1]. This is what you get from tf.image.decode_*() followed by tf.image.convert_image_dtype(..., tf.float32).

A module with exactly one (or one principal) input of images uses the name "images" for this input.

The module accepts any batch_size, and correspondingly sets the first dimension of TensorInfo.tensor_shape to "unknown". The last dimension is fixed to the number 3 of RGB channels. The height and width dimensions are fixed to the expected size of input images. (Future work may remove that restriction for fully convolutional modules.)

Consumers of the module should not inspect the shape directly, but obtain the size information by calling hub.get_expected_image_size() on the module or module spec, and are expected to resize input images accordingly (typically before/during batching).

For simplicity, TF-Hub modules use the channels_last (or NHWC) layout of Tensors, and leave it to TensorFlow's graph optimizer to rewrite to channels_first (or NCHW) if needed. It has been doing that by default since TensorFlow version 1.7.