CODEWITHSUNDEEP
machine-learningcaffe
Wanttobepro
Wanttobepro

Reputation: 77

Do I need to do retraining if I want to change input dimension during testing on caffe model

I have recently been assigned with a project that used Caffe model. So far I manage to run the test code and get good result. But since it only output a small resolution image (128x128), I was asked to get higher resolution of it for example 512x512. Since I just started learning caffe on this project, I still have some problem understanding the flow process of caffe and I stumble upon this questions.

Here is the deploy.prototxt and here is the train.prototxt

So is it actually possible to just modify the deploy.prototxt for example changing the input dimension and some number of output to the resolution that I want without retraining a new model or should I create a new model with an extra layer to enable downsampling and upsampling of higher resolution and do retraining? It would be great if someone can give me some explanation about it. Thank you for sparing your time reading this.

Upvotes: 0

Views: 236

Answers (1)

Hong-333
Hong-333

Reputation: 81

The network will be optimized for 128x128. So you can't put the input in 512x512 or see the output.

To use as input in 512x512 size, you must get a dataset in the 512x512 size and retrain it. It's going to be hard and time consuming to get a dataset.

So there's also a way to change the output resolution using Python.

Upvotes: 0

Related Questions