result changes for a specific fold each time I hit run

import ntpath

nb_classes = 9

import torch.nn.functional as F

confusion_matrix = torch.zeros(nb_classes, nb_classes)

_classes = []
_preds = []
predicted_labels = []

class_probs = torch.Tensor()

im_paths = []
with torch.no_grad():
    for i, (inputs, classes, im_path) in enumerate(dataloaders['test']):


        im_paths.append(im_path)
        inputs = inputs.to(device)
        tmp_labels = model_ft(inputs)

        classes = classes.to(device)
        classes_list = classes.cpu().detach().numpy().tolist()
        _classes[:]=[i+1 for i in classes_list]
        outputs = model_ft(inputs)



        class_probs = class_probs.cuda()

        class_probs = torch.cat((class_probs, F.softmax(outputs, 1)))



        _, preds = torch.max(outputs, 1)
        preds_list = preds.cpu().detach().numpy().tolist()
        _preds[:]=[i+1 for i in preds_list]

        predicted_labels.append(preds.cpu().detach().numpy().tolist())
        for t, p in zip(classes.view(-1), preds.view(-1)):
                confusion_matrix[t.long(), p.long()] += 1

print(confusion_matrix)
per_class_accuracies = (confusion_matrix.diag()/confusion_matrix.sum(1)).cpu().detach().numpy().tolist()

print(','.join("{:2.04f}".format(x) for x in per_class_accuracies))
for i in range(nb_classes):
    print("class {:d} --> accuracy: {:.2f}, correct predictions: {:d}, all: {:d}".format(i+1, (confusion_matrix.diag()/confusion_matrix.sum(1))[i]*100, int(confusion_matrix[i][i].numpy()), int(confusion_matrix.sum(dim=1)[i].numpy())))




flattened_im_paths = flattened = [item for sublist in im_paths for item in sublist]

for i in range(len(flattened_im_paths)):
    class_p = class_probs[i].cpu().detach().numpy().tolist()

    print('{}, {}'