def export_unet1():
        from diffusers import StableDiffusionXLPipeline
        from unet import UNet2DConditionModel
        pipe = StableDiffusionXLPipeline.from_pretrained("G:/stable-model/stable-diffusion-xl", use_safetensors=True)
        model = UNet2DConditionModel()
        model.load_state_dict(pipe.unet.state_dict())
        model.eval()
        latents = torch.randn(2, 4, 128, 128)
        timesteps = torch.tensor([77], dtype=torch.int64)
        prompt_embeds = torch.randn(2, 77, 2048)
        text_embeds = torch.randn(2, 1280)
        time_ids = torch.randn(2, 6)
        input_list = ['latents', 'timesteps', 'prompt_embeds', 'text_embeds', 'time_ids']
        output_list = ['noise_pred']
        torch.onnx.export(model, (latents, timesteps, prompt_embeds, text_embeds, time_ids), f="./onnxmodel/unet1.onnx", 
                          input_names=input_list, output_names=output_list,do_constant_folding=True)
        unet_onnx = onnx.load("./onnxmodel/unet1.onnx")
        onnx.save_model(
                unet_onnx,
                "./onnxmodel/unet.onnx",
                save_as_external_data=True,
                all_tensors_to_one_file=True,
                location="unet.pb",
                size_threshold = 0,
                convert_attribute=True,
            )

    def test_unet_onnx():
        unet = onnxruntime.InferenceSession("./onnxmodel/unet.onnx")#unet.onnx
        latents = np.ones((2, 4, 128, 128)).astype(np.float32)
        timesteps = np.array([77]).astype(np.int64)
        prompt_embeds = np.ones((2, 77, 2048)).astype(np.float32)
        text_embeds = np.ones((2, 1280)).astype(np.float32)
        time_ids = np.ones((2, 6)).astype(np.float32)
        input_list = {'latents':latents, 'timesteps':timesteps, 'prompt_embeds':prompt_embeds, 
                      'text_embeds':text_embeds, 'time_ids':time_ids}
        output_list = ['noise_pred']
        noise_pred = unet.run(output_list, input_list)[0]
        print('noise_pred', noise_pred.shape)
        print(noise_pred)# not correct


    def test_unet_onnx1():
        unet = onnxruntime.InferenceSession("./onnxmodel/unet1.onnx")#unet1.onnx
        latents = np.ones((2, 4, 128, 128)).astype(np.float32)
        timesteps = np.array([77]).astype(np.int64)
        prompt_embeds = np.ones((2, 77, 2048)).astype(np.float32)
        text_embeds = np.ones((2, 1280)).astype(np.float32)
        time_ids = np.ones((2, 6)).astype(np.float32)
        input_list = {'latents':latents, 'timesteps':timesteps, 'prompt_embeds':prompt_embeds, 
                      'text_embeds':text_embeds, 'time_ids':time_ids}
        output_list = ['noise_pred']
        noise_pred = unet.run(output_list, input_list)[0]
        print('noise_pred', noise_pred.shape)
        print(noise_pred)# correct