search

lucbettaieb / thesis

Probably some kind of CNN localization stuff
2 stars 0 forks source link

Inefficient runtime for running images #3

Closed lucbettaieb closed 7 years ago

lucbettaieb commented 7 years ago

When running a graph, the C++ code runs very slowly. This must be due to inefficient code somewhere in the graph running area.

lucbettaieb commented 7 years ago

Copying the images, while inefficient, is not to blame for the holdup.

lucbettaieb commented 7 years ago

tensorflow::Status run_status = g_tf_session_ptr_->Run({{InputName, input_image}}, {OutputName}, {}, &finalOutput); is the culprit

lucbettaieb commented 7 years ago

Seems to be that running BIG tensors through the graph built using BIG tensors (1x307200) from a 640x480 video frame is bad.

It works, but is just slow!

I should now get reduced size training data and then use the image downsize node to run things nicely.

Also, seems as if I can't run images through the network that aren't of the same size...?

ZhiangChen commented 7 years ago

No, if you use the published weights of the Inception model, you can't change the size of the input image.

ZhiangChen commented 7 years ago

You could try tensorboard to check the running graph.

lucbettaieb commented 7 years ago

I'm going to test my hypothesis because I'm getting errors that seem to be indicating what I'm thinking o be correct. After retraining the network on 640x480 images, it will error if I try to run a different size image through the graph.

lucbettaieb commented 7 years ago

Stopped using TF