demo infer result from colab is different sometimes

[better629]

With the guide of Class Agnostic Detection with DETReg

Sometimes the infer result is different with the same input image and pretrain model. The log is

(detreg) root@[machine]machine:code/DETReg# python3 pretrain_infer.py  # pretrain_infer.py is my custom file
Working in inference only mode.
pred_boxes_  tensor([[327.5034, 381.3973, 558.4230, 499.8267],
        [595.2957, 451.0643, 725.7651, 516.2535],
        [482.6980, 443.8614, 606.5033, 507.5866]], grad_fn=<IndexBackward>)
scores_  tensor([0.4935, 0.4557, 0.4064], grad_fn=<IndexBackward>)
time cost: 6.06195592880249
(detreg) root@[machine]machine:code/DETReg# python3 pretrain_infer.py
Working in inference only mode.
pred_boxes_  tensor([[ 597.6837,  450.0903,  717.5170,  513.5975],
        [ 324.6326,  379.7110,  553.9481,  501.0389],
        [ 731.3306,  390.3464, 1094.9594,  452.3845]], grad_fn=<IndexBackward>)
scores_  tensor([0.4473, 0.4444, 0.3841], grad_fn=<IndexBackward>)
time cost: 5.960670471191406
(detreg) root@[machine]machine:code/DETReg# python3 pretrain_infer.py
Working in inference only mode.
^[[Apred_boxes_  tensor([[599.0136, 447.6256, 722.0134, 513.5686],
        [327.3527, 378.5449, 557.3934, 495.2707],
        [229.4011, 115.7684, 830.7236, 227.1658]], grad_fn=<IndexBackward>)
scores_  tensor([0.4902, 0.4759, 0.3888], grad_fn=<IndexBackward>)
time cost: 6.040181875228882

And the visual result is

Why? It's little confusing.