why set keep_ratio to False in process_model_config for mmseg

[liuyunfeng2016]

mmdeploy/mmdeploy/codebase/mmseg/deploy/segmentation.py

Line 39 in 34879e6

cfg.data.test.pipeline[1]['transforms'][0]['keep_ratio'] = False

我在转换自己的训练的模型时，发现推理时这一行会将我的keep_ration=True的配置修改成False，这影响了我模型转换的推理结果！

[RunningLeon]

@liuyunfeng2016 Hi, it is used to make sure preprocessed image has same shape as input_shape for conversion a model with static shape as input_shape. There maybe be a problem if you use this process to do inference.
And, you could evaluate your model with tools/test.py: https://mmdeploy.readthedocs.io/en/latest/tutorials/how_to_evaluate_a_model.html.

BTW, we strongly suggest to make an issue in English.

[liuyunfeng2016]

@liuyunfeng2016 Hi, it is used to make sure preprocessed image has same shape as input_shape for conversion a model with static shape as input_shape. There maybe be a problem if you use this process to do inference. And, you could evaluate your model with tools/test.py: https://mmdeploy.readthedocs.io/en/latest/tutorials/how_to_evaluate_a_model.html.

BTW, we strongly suggest to make an issue in English.

OK, Now I have a new question.
I find that the inference latency of my converted model is longer than that of the original pytorch model.
My environment configuration information is as follows:
CUDA10.2，TensorRT-8.4.0.6、deeplabv3plus、torch 1.8.1、onnxruntime 1.8.1、mmcv-full 1.3.13、mmsegmentation 0.17.0
Here are a few questions:

1. I referred to https://mmdeploy.readthedocs.io/en/latest/tutorials/how_to_measure_performance_of_models.html ，The test result is only 23 ms.
   

2.But I wrote my own code that deinitializes the model and only tests the inference time.I find that tensorrt takes much more time than the pytorch model. Why?
This is the code for the test:
`

1.tensorrt test

if isinstance(backend_models, str):
    backend_models = [backend_models]
deploy_cfg, model_cfg = load_config(deploy_cfg_path, model_cfg_path)

task_processor = build_task_processor(model_cfg, deploy_cfg, device)

model = task_processor.init_backend_model(backend_models)

input_shape = get_input_shape(deploy_cfg)
model_inputs, _ = task_processor.create_input(img_path, input_shape)

start_time = datetime.datetime.now()
with torch.no_grad():
    result = task_processor.run_inference(model, model_inputs)

end_time = datetime.datetime.now()
took_up_time = (end_time - start_time).total_seconds() * 1000
print(f"tensorrt FP32 execution took up time: {took_up_time} ms ")

2.pytorch test

deploy_cfg, model_cfg = load_config(deploy_cfg_path, model_cfg_path)

task_processor = build_task_processor(model_cfg, deploy_cfg, device)

model = task_processor.init_pytorch_model(model_path)

model = MMDataParallel(model, device_ids=[device])

if hasattr(model.module, 'CLASSES'):
    model.CLASSES = model.module.CLASSES

input_shape = get_input_shape(deploy_cfg)
model_inputs, _ = task_processor.create_input(img_path, input_shape)

start_time = datetime.datetime.now()
with torch.no_grad():
    result = task_processor.run_inference(model, model_inputs)

end_time = datetime.datetime.now()
took_up_time = (end_time - start_time).total_seconds() * 1000
print(f"pytorch execution took up time: {took_up_time} ms ")

`
This is the test result. Tensorrt takes about 600 ms and pytorch takes about 180 ms.

3.Is there a tutorial on int8 model conversion and how do I configure calibration?I reported the following error when converting

4.When I convert the FP16 model, the following warning is displayed. Will this cause the problem?

[RunningLeon]

@liuyunfeng2016 Hi,

1. speed test should be done with certain number of iterations and should exclude first N iterations.
2. For int8 conversion, you could refer to how_to_convert_model and use a deployment config with int8 such as configs/mmdet/detection/detection_tensorrt-int8_dynamic-320x320-1344x1344.py
3. Warnings of FP16 could be ignored since there are some int64 tensor in ONNX and there is no int64 in TensorRT, thus it would be converted to int32 automatically.

[liuyunfeng2016]

@liuyunfeng2016你好，

1. 速度测试应该用一定数量的迭代来完成，并且应该排除前 N 次迭代。
2. 对于 int8 转换，您可以参考how_to_convert_model并使用带有 int8 的部署配置，例如configs/mmdet/detection/detection_tensorrt-int8_dynamic-320x320-1344x1344.py
3. FP16 的警告可以忽略，因为 ONNX 中有一些 int64 张量，而 TensorRT 中没有 int64，因此它会自动转换为 int32。

Okay, you're right. Thanks for your help.

[liuyunfeng2016]

1

[RunningLeon]

fixed in #199