test.CNN-F

Test the pre-trained CNN-F in TensorFlow/Pytorch with a simple classification model using MNIST.

Codes of CNN-F and pre-trained parameters are provided in [1].

Analyses of the weight files are contained in test.cnnf.ipynb and cnnf.pytorch.ipynb.

Usage

Tensorflow 1.12

• python main.py
• tensorboard --logdir log

Pytorch

• python main_torch.py

TensorFlow 2.1.0

• python main_tf2.py

Data

MNIST, zooming into [224, 224, 3].

Result

tensorflow 1.12

• iter 0: 0.12269999995827675
• iter 450: 0.9907000076770782

pytorch

• 1 epoch: 0.98

tensorflow 2.1

• 1 epoch: 0.9742

Environment

• tensorflow 1.12.0 / 2.1.0
• pytorch 1.4.0, torchvision 0.5.0
• cuda 9.0 / 10.1

you may use docker: tyloeng/dl

Pre-trained Weights

cnnf-vggf

References

1. jiangqy/DCMH-CVPR2017
2. tensorflow加载CNN-F/VGG-F预训练参数