search

doonny / PipeCNN

An OpenCL-based FPGA Accelerator for Convolutional Neural Networks
Apache License 2.0
1.26k stars 369 forks source link

准确率的问题 #70

Closed HulkSun closed 6 years ago

HulkSun commented 6 years ago

您好， @doonny ： 我在ImageNet的ILSVRC2012_img_val数据集上运行pipeCNN，得到的Top1和Top5 accuracy分别是64%和85%，而VGG的是75%和93%。请问得到的结果本身就有这么大的误差还是说我使用不当造成的？ 谢谢！

laski007 commented 6 years ago

I think the accuracy on AlexNet or VGGNet could not be that much high. You may not test all the images on ILSVRC image test dataset. The more images, the accuracy will be decreased. E.g., the official accuracy of AlexNet is 56.9% and a top-5 accuracy 80.1%. As for VGG16, which is 70.6% and 89.9%.

HulkSun commented 6 years ago

@laski007 Thanks for the reply! I found the accuracy data at https://github.com/BVLC/caffe/wiki/Models-accuracy-on-ImageNet-2012-val which means you are right.