problem inference with onnx model

[Handsome-cp]

after I training swinT classification model with my own dataset (10 classes),

I transfer it to onnx model using torch.onnx.export tool, codes below:

import torch
from models import build_model
from main import parse_option


if __name__ == '__main__':
    args, config = parse_option()
    model = build_model(config)
    checkpoint = torch.load('./output/swin_tiny_patch4_window7_224/default/ckpt_epoch_60.pth', map_location='cpu')
    print('start eval')
    model.eval()

    dummy_input = torch.randn(1, 3, 224, 224, device='cpu')
    input_names = ['input']
    output_names = ['output']

    print('start export')
    torch.onnx.export(
        model, dummy_input, './output/swin_tiny_patch4_window7_224/default/swin_tiny_epoch_60.onnx', verbose=True,  
        opset_version=11, input_names=input_names, output_names=output_names
    )

but when testing this onnx on training dataset, the result is quit different from the pth， which almost to be one class.

[hua1024]

@Handsome-cp Hi, I met same question with you , when i export the onnx model like you，but the result is different from the pth , Have you found a solution ?