models Package

Python representation of models accelerated with the Azure ML Hardware Accelerated Models Service.

Modules

accel_model	Module with abstract base class of HW accelerated models.
doesnotexisterror	Does not exist error.
utils	Utilities for models - mostly preprocessing related.

Classes

Densenet121	Float-32 Version of Densenet. This model is in RGB format, and has a scaling factor of 0.017 Create a Float-32 version of Densenet. This model is in RGB format.
QuantizedDensenet121	Quantized version of Densenet. This model is in RGB format. Create a version of Densenet quantized for the Azure ML Hardware Accelerated Models Service. This model is in RGB format.
QuantizedResnet152	Quantized version of Renset-152. Create a version of resnet 50 quantized for the Azure ML Hardware Accelerated Models Service.
QuantizedResnet50	Quantized version of Renset-50. Create a version of resnet 50 quantized for the Azure ML Hardware Accelerated Models Service.
QuantizedSsdVgg	Quantized version of SSD-VGG. This model is in RGB format. Create a version of SSD VGG quantized for the Azure ML Hardware Accelerated Models Service. This model is in RGB format. :param model_base_path: Path to download the model into. Used as a cache locally. :param is_frozen: Should the weights of the model be frozen when it is imported. Freezing the weights can lead to faster training time, but may cause your model to perform worse overall. Defaults to false.
QuantizedVgg16	Quantized version of VGG-16. This model is in RGB format. Create a version of VGG 16 quantized for the Azure ML Hardware Accelerated Models Service. This model is in RGB format.
Resnet152	Float-32 Version of Resnet-152. Create a Float-32 version of resnet 152.
Resnet50	Float-32 Version of Resnet-50. Create a Float-32 version of resnet 50.
SsdVgg	Float-32 Version of SSD-VGG. This model is in RGB format. Create a Float-32 version of SSD-VGG.
Vgg16	Float-32 Version of VGG-16. This model is in RGB format. Create a Float-32 version of VGG 16.

Functions

preprocess_array

Create a tensorflow op that takes an array of image bytes and returns regularized images.

preprocess_array(in_images, order='RGB', scaling_factor=1.0, output_height=224, output_width=224, preserve_aspect_ratio=True)

Parameters

in_images

Required

[?] dim tensor of image bytes. (Typically a placeholder)

order

default value: RGB

order of channels - either 'BGR' or 'RGB'

scaling_factor

default value: 1.0

multiplier for channel values

output_height

default value: 224

output image height

output_width

default value: 224

output image width

preserve_aspect_ratio

default value: True

if True, preserve image aspect ratio while scaling

Returns

[?, output_height, output_width, 3] dim tensor of float32 pixel values of the image.