Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/XuChunqiao/Tensorflow-Fast-MPNCOV
Fast_MPNCOV implemented by TensorFlow2.0 for training from scratch & finetuning； BCNN implemented by Tensorflow2.0 for finetuning；Compact BCNN implemented by Tensorflow2.0 for finetuning
https://github.com/XuChunqiao/Tensorflow-Fast-MPNCOV
bcnn compact-bcnn fast-mpncov python tensorflow
Last synced: 2 months ago
JSON representation
Fast_MPNCOV implemented by TensorFlow2.0 for training from scratch & finetuning； BCNN implemented by Tensorflow2.0 for finetuning；Compact BCNN implemented by Tensorflow2.0 for finetuning
- Host: GitHub
- URL: https://github.com/XuChunqiao/Tensorflow-Fast-MPNCOV
- Owner: XuChunqiao
- Created: 2019-07-27T05:39:21.000Z (over 5 years ago)
- Default Branch: master
- Last Pushed: 2019-11-22T07:55:05.000Z (about 5 years ago)
- Last Synced: 2024-08-04T01:30:18.947Z (6 months ago)
- Topics: bcnn, compact-bcnn, fast-mpncov, python, tensorflow
- Language: Python
- Homepage:
- Size: 189 KB
- Stars: 15
- Watchers: 2
- Forks: 2
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
- Awesome-Tensorflow2 - XuChunqiao/Tensorflow-Fast-MPNCOV
README
# Tensorflow Fast-MPNCOV
## Introduction
This repository contains the source code under ***TensorFlow2.0 framework*** and models trained on ImageNet 2012 dataset for the following paper:
```
@InProceedings{Li_2018_CVPR,
author = {Li, Peihua and Xie, Jiangtao and Wang, Qilong and Gao, Zilin},
title = {Towards Faster Training of Global Covariance Pooling Networks by Iterative Matrix Square Root Normalization},
booktitle = { IEEE Int. Conf. on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2018}
}
```
This paper concerns an iterative matrix square root normalization network (called fast MPN-COV), which is very efficient, fit for large-scale datasets, as opposed to its predecessor (i.e., [MPN-COV](https://github.com/jiangtaoxie/MPN-COV)) published in ICCV17) that performs matrix power normalization by Eigen-decompositon. If you use the code, please cite this [fast MPN-COV](http://peihuali.org/iSQRT-COV/iSQRT-COV_bib.htm) work and its predecessor (i.e., [MPN-COV](http://peihuali.org/iSQRT-COV/iSQRT-COV_bib.htm)).
## Classification results
#### Classification results (single crop 224x224, %) on ImageNet 2012 validation set
Network
Dim
Top1_err/Top5_err
Pre-trained models
(tensorflow)
paper
reproduce
tensorflow
pytorch
GoogleDrive
BaiduDrive
fast-MPN-COV-VGG-D
32K
26.55/8.94
23.98/7.12
23.98/7.12
650.4M
650.4M
fast-MPN-COV-ResNet50
22.14/6.22
21.57/6.14
21.71/6.13
217.3M
217.3M
fast-MPN-COV-ResNet101
21.21/5.68
20.50/5.45
20.99/5.56
289.9M
289.9M
* We convert the trained fast-MPNCOV-VGG-D model from the PyTorch framework to TensorFlow framework.
#### Fine-grained classification results (top-1 accuracy rates, %)
paper
reproduce
(tensorflow)
paper
reproduce
(tensorflow)
paper
reproduce
(tensorflow)
fast-MPNCOV-COV-VGG-D
32K
87.2
86.95
90.0
91.74
92.5
92.95
fast-MPNCOV-COV-ResNet50
88.1
87.6
90.0
90.5
92.8
93.2
fast-MPNCOV-COV-ResNet101
88.7
88.1
91.4
91.8
93.3
93.9
* Our method uses neither bounding boxes nor part annotations
* The reproduced results are obtained by simply finetuning our pre-trained fast MPN-COV-ResNet model with a small learning rate, which do not perform SVM as our paper described.
```
parameter setting
fast-MPNCOV-VGG-D: weightdecay=1e-4, batchsize=10, learningrate=3e-3 for all layers except the FC layer(which is 5×learningrate, and the learning rate is reduced to 3e-4 at epoch 20(FC: 5×3e-4)
```
## Implementation details
We implement our Fast MPN-COV (i.e., iSQRT-COV) [meta-layer](https://github.com/XuChunqiao/Tensorflow-Fast-MPNCOV/blob/master/src/representation/MPNCOV.py) under ***Tensorflow2.0*** package. We release two versions of code:
* The backpropagation of our meta-layer without using autograd package;
* The backpropagation of our meta-layer with using autograd package(**TODO**).
For making our Fast MPN-COV meta layer can be added in a network conveniently, we divide any network for three parts:
* features extractor;
* global image representation;
* classifier.
As such, we can arbitrarily combine a network with our Fast MPN-COV or some other global image representation methods (e.g.,Global average pooling, Bilinear pooling, Compact bilinear pooling, etc.)
## Installation and Usage
1. Install [Tensorflow (2.0.0b0)](https://tensorflow.google.cn/install)
2. type ```git clone https://github.com/XuChunqiao/Tensorflow-Fast-MPNCOV ```
3. prepare the dataset as follows
```
.
├── train
│ ├── class1
│ │ ├── class1_001.jpg
│ │ ├── class1_002.jpg
| | └── ...
│ ├── class2
│ ├── class3
│ ├── ...
│ ├── ...
│ └── classN
└── val
├── class1
│ ├── class1_001.jpg
│ ├── class1_002.jpg
| └── ...
├── class2
├── class3
├── ...
├── ...
└── classN
```
### train from scratch
1. ``` cp ./trainingFromScratch/imagenet/imagenet_tfrecords.py ./ ```
2. modify the dataset path and run ``` python imagenet_tfrecords.py ``` to create tfrecord files
3. modify the parameters in train.sh ```sh train.sh```
### finetune fast-MPNCOV models
1. modify the parameters in finetune.sh
2. ```sh finetune.sh```
## Other Implementations
* [MatConvNet Implementation](https://github.com/jiangtaoxie/matconvnet.fast-mpn-cov)
* [PyTorch Implementation](https://github.com/jiangtaoxie/fast-MPN-COV)