GluonCV - Inference gives me different results every round

Hi, I have used the code from https://gluon-cv.mxnet.io/build/examples_action_recognition/finetune_custom.html to train a custom model, I’m using this:

net = get_model(name='i3d_resnet50_v1_custom', nclass=2)

to load the model and train it, I only want to have two classes, Violence or not Violence, I’m using a dataset with two classes…

then I use the params produced for the training:

net = get_model(name=‘i3d_resnet50_v1_custom’, pretrained=False, pretrained_base=False,nclass=2) #important to put the same number of classes
net.load_parameters(’/dbfs/mnt/xxxx/mxnetvideos/modelViolence.params’)

but when I make the inference for a set of four videos, every time I run the code I got different results, please see below the results in two different executions of the Inference code:

pred:
[[ 2.0384998 -1.9758421 ]
[-1.112703 1.1304163 ]
[-1.122371 1.1812717 ]
[ 0.3237142 -0.30243284]]
<NDArray 4x2 @cpu(0)>
label:[
[0 1 1 0]
<NDArray 4 @cpu(0)>]

pred:
[[-1.1259147 1.1834983 ]
[-0.40505636 0.42660624]
[-1.088927 1.1552093 ]
[ 0.8953121 -0.8409692 ]]
<NDArray 4x2 @cpu(0)>
label:[
[1 0 1 0]
<NDArray 4 @cpu(0)>]

I have found that I had a parameter in the VideoClsCustom function that shuffled the data, so each time it the data was shuffled