resnet-50-tf

Use Case and High-Level Description

resnet-50-tf is a TensorFlow* implementation of ResNet-50 - an image classification model pretrained on the ImageNet dataset. Originally redistributed in Saved model format, converted to frozen graph using tf.graph_util module. For details see paper, repository.

Steps to Reproduce Conversion to Frozen Graph

  1. Install TensorFlow*, version 1.14.0.
  2. Download pretrained weights
  3. Run example conversion code, available at freeze_saved_model.py
    python3 freeze_saved_model.py --saved_model_dir path/to/downloaded/saved_model --save_file path/to/resulting/frozen_graph.pb

Specification

Metric Value
Type Classification
GFLOPs 8.2164
MParams 25.53
Source framework TensorFlow*

Accuracy

Metric Original model Converted model
Top 1 76.45% 76.17%
Top 5 93.05% 92.98%

Input

Original Model

Image, name: map/TensorArrayStack/TensorArrayGatherV3, shape: 1,224,224,3, format is B,H,W,C where:

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

Channel order is RGB. Mean values: [123.68,116.78,103.94].

Converted Model

Image, name: map/TensorArrayStack/TensorArrayGatherV3, shape: 1,224,224,3, format is B,H,W,C where:

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

Channel order is BGR.

Output

Original Model

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

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

Converted Model

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

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

Legal Information

The original model is distributed under the Apache License, Version 2.0. A copy of the license is provided in APACHE-2.0-TensorFlow.txt.