MLPerf Inference Reference Models
Published September 6, 2019 | Version v0 -tf1.1 + readme file
Other Open

mobilenet-v1-ssd300.tensorflow (8bit symmetrically quantized and fine-tuned)

Authors/Creators

  • 1. Habana Labs

Description

Application:  Single-stage Object Detection

Base model: MobileNet-v1

Framework: tensorflow1.1

Training Information: weights were fine-tuned using TF fake quantization nodes

Quality: The COCO mAP(IoU=0.50:0.95) on 5000 validation images is 23.4%

Precision: 8-bit precision

Is Quantized: Yes, using fake quantization with symmetric=True. - i.e., weights appear in float32 but have only 256 unique values and no zero point. 

Dataset: COCO val-2017

Notes

Fake quantization with symmetric=True. Weights appear in float32 but have only 256 unique values and no zero point. Additional information in the README file.

Files

README.txt

Files (29.2 MB)

Name Size Download all
md5:4a29c994dc9bcbf9337c742c32963d05
1.0 kB Preview Download
md5:09d4de1075f1b1ac74db505a0e9284b8
29.2 MB Download