Segformer output not determinsitic

[migperfer]

Hi! Thanks for the library.

I'm using a Segformer model, and the same input through the same model is giving me always two different outputs.
I leave one example here.

# Imports
import segmentation_models_pytorch as smp
from PIL.Image import open as open_image
from PIL.Image import FLIP_LEFT_RIGHT
import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
import random, torch
from pathlib import Path
import torch

random.seed(0)
np.random.seed(0)
torch.manual_seed(0)
torch.cuda.manual_seed_all(0)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch.use_deterministic_algorithms(True)

# ... Some other code
device = "cpu"
model = smp.Segformer(**model_params)
model.load_state_dict("my/checkpoint.ckpt").to(device)

image = open_image("some/path.png")
images = [image, image]  # The same image

with torch.inference_mode():        
        images_tensor =  torch.utils.data.default_collate([inference_transform(img) for img in images])
        assert torch.allclose(images_tensor[0], images_tensor[1])  # 1st assert
        mask_predictions = torch.sigmoid(self(images_tensor.to(model.device)))
        assert torch.allclose(mask_predictions[0], mask_predictions[1])  # 2nd assert

Assert # 1 does not raise any errors (the two elements being input to the network are really the same)
However, assert # 2 raises an error, as the outputs are not the same.

Any idea of what could be happening?
Thanks.

[migperfer]

Sorry for the issue, it was a very dumb detail. It seems that torch.inference_mode() does not disable dropout, change the behavior of any batch normalization etc. After doing model.eval() everything worked fine.