extremely low probability inference on pretrained resnet50-v1-12.onnx model

[Apisteftos]

I am trying actually to quantize the pretrained resnet50 model and I am getting a low prediction, I have my own preprocessing model but I tried to compared with the provided one.

onnx preprocessing resnet50
onnx preprocessing efficientnet-lite

here is the outcome with kitten.jpgand panda.jpg according to provided preprocessing adjustments.

class=n02123159 tiger cat ; probability=12.294151
class=n02510455 giant panda, panda, panda bear, coon bear, Ailuropoda melanoleuca ; probability=11.516657

and here is my predictions according to my configuration preprocess

282 tiger cat 12.07%
281 tabby, tabby cat 11.91%
285 Egyptian cat 10.97%
287 lynx, catamount 9.93%
289 snow leopard, ounce, Panthera uncia 8.35%

388 giant panda, panda, panda bear, coon bear, Ailuropoda melanoleuca 13.74%
341 hog, pig, grunter, squealer, Sus scrofa 5.79%
812 space shuttle 5.73%
294 brown bear, bruin, Ursus arctos 5.73%
805 soccer ball 5.64%

[xadupre]

Is it possible to know how you quantized your model? (which command line).

[Apisteftos]

Oh I did that exporting from pytorch to onnx, I am not sure, because it's a long time, but it seems it didn't had any good training or quantization.