BEATS reproducibility

[poonehmousavi]

Describe the bug
I am using BEATs as a feature extractor. and using BEATs_iter3.pt .. however passing the same input, I got a different output every time.. I wonder what could be the cause . Since the model is pretrained and frozen, for the same input, the output should be the same. It would be great if you could help me understand it better.

tensor([[ 1.9269,  1.4873,  0.9007,  ...,  0.8227, -0.8546,  0.7941]])
tensor([[[ 0.4125,  0.1860,  0.2770,  ...,  0.0746,  0.0759, -0.1040],
         [-0.1389, -0.0395, -0.1223,  ..., -0.1694, -0.2940,  0.0465],
         [ 0.0531, -0.0998,  0.4792,  ..., -0.5941, -0.4111,  0.0983],
         ...,
         [ 0.7151,  0.0939,  0.0118,  ...,  0.6868, -0.3090,  0.2154],
         [ 0.2418,  0.2485, -0.0050,  ...,  0.5315, -0.0149, -0.1740],
         [ 0.3563,  0.4537,  0.0398,  ...,  0.5003,  0.0146, -0.0805]]],
       grad_fn=<TransposeBackward0>)

tensor([[ 1.9269,  1.4873,  0.9007,  ...,  0.8227, -0.8546,  0.7941]])
tensor([[[ 0.1959,  0.0528,  0.4861,  ...,  0.0678, -0.1611, -0.0386],
         [-0.1310, -0.3422,  0.3333,  ..., -0.0915, -0.2356, -0.0009],
         [ 0.2121,  0.0532,  0.4134,  ..., -0.6164, -0.5548, -0.0760],
         ...,
         [ 0.7319, -0.0581,  0.1427,  ...,  0.5444, -0.5717,  0.0932],
         [ 0.3550,  0.1968,  0.1670,  ...,  0.4279, -0.4536, -0.1726],
         [ 0.3206,  0.2454,  0.2433,  ...,  0.6233, -0.3299, -0.2126]]],

The problem arises when using:

• the official example scripts: (give details below)
• my own modified scripts: (give details below)

A clear and concise description of what the bug is.

To Reproduce
Steps to reproduce the behavior:
1.
2.
3.

Expected behavior
A clear and concise description of what you expected to happen.

• Platform:
• Python version:
• PyTorch version (GPU?):