Transfer learning low accuracy and CUDA not working

beazt · July 13, 2023, 7:41am

Hi there,

I am running into a few problems while trying out this code:

pytorch/examples/blob/main/cpp/transfer-learning/main.cpp

//
//  main.cpp
//  transfer-learning
//
//  Created by Kushashwa Ravi Shrimali on 12/08/19.
//

#include "main.h"

torch::Tensor read_data(std::string location) {
    /*
     Function to return image read at location given as type torch::Tensor
     Resizes image to (224, 224, 3)
     Parameters
     ===========
     1. location (std::string type) - required to load image from the location
     
     Returns
     ===========
     torch::Tensor type - image read as tensor

This file has been truncated. show original

So basically I have just downloaded the git and the dog and cat samples and run the code without any changes, except for the paths I am using of course. The first problem is, that I only get accuracies of about 51-52% which does not go higher after 25 epochs. Do I need to set up something different to get higher accuracies?

While classifying, I only get about 15-30% accuracies which is not acceptable I guess.

Another problem I have is, when I try to train on CUDA, which runs just fine and I can classifying using CUDA but the accuracies are negative like so: -35940% for example.

The changes I did for training was this:

 module = torch::jit::load(weight, at::kCUDA);
 lin->to(at::kCUDA);
 target = target.to(at::kCUDA);
 input.push_back(data.to(at::kCUDA));

and for the classifying part I made these changes:

 img_tensor = img_tensor.to(at::kCUDA);
 model = torch::jit::load(model_path, at::kCUDA);
 model.eval();
 model_linear->to(at::kCUDA);
 torch::load(model_linear, model_path_linear, at::kCUDA);

Anybody can help please?

Edit: The classifying does not work at all. While using the CPU trained version, I always get about the same results, which is about 20% cat, 30% dog, it does only change a few % whatever images I am using to classify. The results always stay about the same.