Deploy pre-trained Caffe Model

[Valentin Stefanov]

Dear Group,

as I am performing some ML/DL tasks in our university my Professor gave me the responsibility to test pre-trained caffe networks/models on our University internal data. 

He found many pre-trained networks for our problem (text detection) on the web and would like me to verify their accuracy on our data without tuning them (kind of "out-off-the-shelf solution"). Some of them, however, include only the .caffemodel and .prototxt file. 

I wanted to ask if those two files are sufficient to test the network's performance on our dataset. So far I have written the networks myself, mostly in keras, and I could validate them on the data we have pre-processed (80/20) but I have never used them on totally novel data. 

Best Regards!

Valentin

[Kağan İncetan]

Hi Valentin,

These two files are definitely enough. .caffemodel represents the weights of the trained model and .prototxt is to define the network. So once you have these two you can easily define your net then you can test it by getting the prediction values from the corresponding layers. Here is the starting point how to define your NET.
net = caffe.Net(model,
                weights,
                caffe.TEST)
where model is the prototxt file, weights is the caffemodel file. caffe.TEST is showing that you are going to test the model not train it as it is already trained.

Best regards

18 Ekim 2017 Çarşamba 22:55:01 UTC+2 tarihinde Valentin Stefanov yazdı:

[Valentin Stefanov]

Dear Kagan,

thank you very much for your answer! Do you know any official documentation or some guide how to use these pre-trained models? I do not really find something on the web about this. Seems to be either self-explanatory or just not on the web yet?

Best Regards

Valentin

[6459...@qq.com]

Hi Valentin,

Caffe's web(http://caffe.berkeleyvision.org/) has some examples(Notebook Examples,Command Line Examples); you may check it, i am not sure if  these samples helpful to you... 

Best Regards,

Tom

在 2017年10月20日星期五 UTC+8上午4:28:45，Valentin Stefanov写道：