Help with DeepLab. Runing problem. Nan output

I am really new here but I have managed to train (i think) my own DeepLab model with my custom dataset in coco format:
(Epoch 3, training loss 0.165: 100%|█████████████████████████████| 70/70 [12:10<00:00, 10.44s/it]
Epoch 3, validation pixAcc: 0.872, mIoU: 0.436: 100%|█████████████████████████████| 5/5 [01:04<00:00, 12.89s/it]
Epoch 3 validation pixAcc: 0.872, mIoU: 0.436)

Howewer I cannot run it successfully
I’ve not been able to find tutorial about training and using my own DeepLab model so I have adapdet some found logic to get it processing my images on cpu without exeptions
But now I am stuck
Every time a get the same empty result
Maybe the problem is the same as here:

About running on CPU and stuff, but I have managed to use GPUs only in training process, not in the running. And I don’t know how to do that
Maybe I have done all wrong
Could you get me some advice?
Thank you

Running script:

import mxnet as mx
from mxnet import image
from import transforms
import gluoncv
# using cpu
ctx = mx.cpu(0)


filepath = r"F:\mxnet-cu100\gebrei(2)-1-2_06.jpg"


img = image.imread(filepath)

from matplotlib import pyplot as plt


from import test_transform
img = test_transform(img, ctx)


CAT_LIST = [0, 1]
CLASSES = ("background", "building")

model = gluoncv.model_zoo.get_model(
    'deeplab_resnet50_coco', pretrained=False, num_class=NUM_CLASS, classes=CLASSES)


output = model.predict(img)
predict = mx.nd.squeeze(mx.nd.argmax(output, 1)).asnumpy()


from gluoncv.utils.viz import get_color_pallete
import matplotlib.image as mpimg
mask = get_color_pallete(predict, 'ade20k')'output.png')


mmask = mpimg.imread('output.png')

You might need another color pallete other than ade20k. BTW you can also check the output directly to see if it’s bad

predict = mx.nd.squeeze(mx.nd.argmax(output, 1)).asnumpy()

Thanks for replying
Unfortunatly there are only zeros here:

what happens if allow_missing=True changed to allow_missing=False, I suspect that you might have missing parameters causing the model not loaded properly.