search

doonny / PipeCNN

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

mean.data format #132

Open mAxwong1014 opened 4 years ago

mAxwong1014 commented 4 years ago

Dear @doonny prof Wang, Currently I am working on my FYP, thanks for your contribution in giving clear description. Currently I am stucked in generating the mean_data.dat. Based on my understanding, the mean_data binaryproto generated in caffe cannot be used directly, I also look through the paper several times but didnt find the answer. Your help is highly appreciated!!! Thanks in advance.

mAxwong1014 commented 4 years ago

dear prof @donny, I would appreciated it if you can help me, thank you.

doonny commented 4 years ago

Hi, the mean_data file is obtained from this repo: https://github.com/pmgysel/alexnet-forwardpath

mAxwong1014 commented 4 years ago

thanks for replying prof @doonny , if I want to train my data, how can I generate this mean data? Based on my understanding, the mean_data binaryproto generated in caffe cannot be used directly, I also look through the paper several times but didnt find the answer. Your help is highly appreciated!!! Thanks in advance.

doonny commented 4 years ago

Please refer the user's manual of Caffe. I remember they have some instruction on setting up data set for training. Here, we only focus on hardware implementation

mAxwong1014 commented 4 years ago

thanks for replying prof @doonny. based on my understanding, the hardware implementation is based on ilsvrc12 imagenet competition, I downloaded the imagenet_mean.binaryproto file and try it on pipecnn but the accuracy is poorly performed. May I know any significant change needed to be done on the imagenet_mean.binaryproto to generate the mean_data,dat? Sorry to bothering you again.

doonny commented 4 years ago

We have not made any changes to the mean file.