Skip to content

Latest commit

 

History

History
49 lines (30 loc) · 1.75 KB

inceptionv3-int8-sparse-v1-tf-0001.md

File metadata and controls

49 lines (30 loc) · 1.75 KB
Raw

inceptionv3-int8-sparse-v1-tf-0001

Use Case and High-Level Description

This is the Inception v3 model that is designed to perform image classification. The model has been pretrained on the ImageNet image database and then pruned to 30.9% of sparsity and quantized to INT8 fixed-point precision using so-called Quantization-aware training approach implemented in TensorFlow framework. The sparsity is represented by zeros inside the weights of Convolutional and Fully-conneted layers. For details about the original floating point model, check out the paper.

The model input is a blob that consists of a single image of "1x299x299x3" in BGR order.

The model output for inceptionv3-int8-sparse-v1-tf-0001 is the usual object classifier output for the 1001 different classifications matching those in the ImageNet database (the first item represents the background).

Example

Specification

Metric Value
Type Classification
GFLOPs 11.469
MParams 23.819
Source framework TensorFlow

Accuracy

The quality metrics calculated on ImageNet validation dataset is 78.65% accuracy top-1.

Metric Value
Accuracy top-1 (ImageNet) 78.65%

Performance

Input

Image, shape - 1,299,299,3, format is B,H,W,C where:

  • B - batch size
  • H - height
  • W - width
  • C - channel

Channel order is BGR

Output

Object classifier according to ImageNet classes, shape -1,1001, output data format is B,C where:

  • B - batch size
  • C - predicted probabilities for each class in [0, 1] range