Method test_run

test/tensorrt/test_export.py:240–305

(self)

        )

    def test_run(self):
        with paddle.pir_utils.IrGuard():
            input_config = Input(
                min_input_shape=(1, 1, 28, 28),
                optim_input_shape=(1, 1, 28, 28),
                max_input_shape=(1, 1, 28, 28),
            )
            input_config2 = Input(
                min_input_shape=(1, 1, 28, 28),
                optim_input_shape=(1, 1, 28, 28),
                max_input_shape=(1, 1, 28, 28),
            )
            trt_config = TensorRTConfig(inputs=[input_config, input_config2])
            trt_config.save_model_dir = os.path.join(self.temp_dir.name, 'trt')

            min_data_list = []
            max_data_list = []
            for i, input_instance in enumerate(trt_config.inputs):
                min_data, _, max_data = input_instance.generate_input_data()

                min_data_list.append(min_data)
                max_data_list.append(max_data)
                paddle.disable_static()

            x = [paddle.to_tensor(md) for md in min_data_list]
            net = LeNetMultiInput()
            out = net(*x)

            input_spec = [
                InputSpec(
                    shape=min_data_list[0].shape, dtype='float32', name='input1'
                ),
                InputSpec(
                    shape=min_data_list[1].shape, dtype='float32', name='input2'
                ),
            ]

            program_with_trt, scope = _convert_(
                net,
                input_spec=input_spec,
                config=trt_config,
                full_graph=True,
            )

            config = paddle_infer.Config(
                trt_config.save_model_dir + '.json',
                trt_config.save_model_dir + '.pdiparams',
            )

            if paddle.is_compiled_with_cuda():
                config.enable_use_gpu(100, 0)
            else:
                config.disable_gpu()

            predictor = paddle_infer.create_predictor(config)
            output_converted = predictor.run(x)
            output_converted_np = output_converted[0]
            output_expected = out.numpy()