RT-DETR TensorRT 输出边框都是0,

[aliencaocao]

问题确认 Search before asking

• 我已经查询历史issue，没有发现相似的bug。I have searched the issues and found no similar bug report.

Bug组件 Bug Component

Inference, Export, Deploy

Bug描述 Describe the Bug

使用教程内的命令转RT-DETR-X到onnx，然后再转到TensorRT。ONNX一切正常，用TensorRTExecutionProvider也正常，但是如果用trtexec转到TensorRT engine，模型输出的class id和confidence看起来正常（实际对不对不知道），但是x1 x2 y1 y2都是0.000。在Windows 10 CUDA 11.8 + CUDNN 8.9和WSL2 Cuda 11.8 + CUDNN 8.9都可以复现。

图片是

复现环境 Environment

Windows 10 和 WSL2
CUDA 11.8
CUDNN 8.9
PaddlePaddle-gpu 2.4.2-post117
PaddleDetection develop
Python 3.9.13和3.10.6都可以复现

Bug描述确认 Bug description confirmation

• 我确认已经提供了Bug复现步骤、代码改动说明、以及环境信息，确认问题是可以复现的。I confirm that the bug replication steps, code change instructions, and environment information have been provided, and the problem can be reproduced.

是否愿意提交PR？ Are you willing to submit a PR?

• 我愿意提交PR！I'd like to help by submitting a PR!

[aliencaocao]

用了另一个版本的推理代码，现在bbox不是0了，但是整个结果都是乱的，bbox完全不着边，而且置信度都很低

我用的推理代码：

class TRTInference:
    def __init__(self, engine_path, output_names_mapping: dict = None, verbose=False):
        cuda.init()
        self.device_ctx = cuda.Device(0).make_context()
        self.engine_path = engine_path
        self.output_names_mapping = output_names_mapping or {}
        self.logger = trt.Logger(trt.Logger.VERBOSE) if verbose else trt.Logger(trt.Logger.INFO)
        self.engine = None
        self.load_engine()
        assert self.engine is not None, 'Failed to load TensorRT engine.'

        self.context = self.engine.create_execution_context()
        self.stream = cuda.Stream()

        self.bindings = self.get_bindings()
        self.bindings_addr = OrderedDict((n, v.ptr) for n, v in self.bindings.items())

        self.input_names = self.get_input_names()
        self.output_names = self.get_output_names()

    def load_engine(self):
        with open(self.engine_path, 'rb') as f, trt.Runtime(self.logger) as runtime:
            self.engine = runtime.deserialize_cuda_engine(f.read())

    def get_input_names(self):
        names = []
        for _, name in enumerate(self.engine):
            if self.engine.get_tensor_mode(name) == trt.TensorIOMode.INPUT:
                names.append(name)
        return names

    def get_output_names(self):
        names = []
        for _, name in enumerate(self.engine):
            if self.engine.get_tensor_mode(name) == trt.TensorIOMode.OUTPUT:
                names.append(name)
        return names

    def get_bindings(self):
        Binding = namedtuple('Binding', ('name', 'dtype', 'shape', 'data', 'ptr'))
        bindings = OrderedDict()

        for i, name in enumerate(self.engine):
            shape = self.engine.get_tensor_shape(name)
            dtype = trt.nptype(self.engine.get_tensor_dtype(name))
            if self.engine.get_tensor_mode(name) == trt.TensorIOMode.INPUT:
                data = np.random.randn(*shape).astype(dtype)
                ptr = cuda.mem_alloc(data.nbytes)
                bindings[name] = Binding(name, dtype, shape, data, ptr)
            else:
                data = cuda.pagelocked_empty(trt.volume(shape), dtype)
                ptr = cuda.mem_alloc(data.nbytes)
                bindings[name] = Binding(name, dtype, shape, data, ptr)

        return bindings

    def __call__(self, blob):
        blob = {n: np.ascontiguousarray(v) for n, v in blob.items()}
        for n in self.input_names:
            cuda.memcpy_htod_async(self.bindings_addr[n], blob[n], self.stream)

        bindings_addr = [int(v) for _, v in self.bindings_addr.items()]
        self.context.execute_async_v2(bindings=bindings_addr, stream_handle=self.stream.handle)

        outputs = {}
        for n in self.output_names:
            cuda.memcpy_dtoh_async(self.bindings[n].data, self.bindings[n].ptr, self.stream)
            o = self.bindings[n].data
            # reshape to correct output shape
            if o.shape != self.bindings[n].shape:
                o = o.reshape(self.bindings[n].shape)
            outputs[self.output_names_mapping.get(n, n)] = o

        self.stream.synchronize()

        return outputs

    def warmup(self, blob, n):
        for _ in range(n):
            _ = self(blob)

    def __del__(self):
        try:
            self.device_ctx.pop()
        except cuda.LogicError as _:
            pass

model = TRTInference('rtdetr_hgnetv2_x_6x_coco.trt', output_names_mapping={'tile_3.tmp_0': 'bbox_num', 'reshape2_83.tmp_0': 'bbox'}, verbose=True)

img = cv2.imread("soccer.jpg")
h, w = 640, 640
org_img = img
img = cv2.resize(img, (h, w))
im_shape = np.array([[float(img.shape[0]), float(img.shape[1])]]).astype('float32')
scale_factor = np.array([[float(h/org_img.shape[0]), float(w/org_img.shape[1])]]).astype('float32')
img = img.astype(np.float32) / 255.0
img = np.expand_dims(img, axis=0)  # add batch dimension
inputs_dict = {
    'image': img,
    'im_shape': im_shape,
    'scale_factor': scale_factor
}
result = model(inputs_dict)['bbox']

confidence_threshold = 0.1
boxes = result[result[:, 1] > confidence_threshold]
for box in boxes:
    cv2.rectangle(org_img, (int(box[2]), int(box[3])), (int(box[4]), int(box[5])), (0, 255, 0), 2)
cv2.imwrite("output/soccer_trt.jpg", org_img)

[aliencaocao]

用了 https://aistudio.baidu.com/aistudio/projectdetail/6000200的 inputs_dict 这部分推理代码后问题解决了。

img = cv2.imread("soccer.jpg")
org_img = img
im_shape = np.array([[float(img.shape[0]), float(img.shape[1])]]).astype('float32')
img = cv2.resize(img, (640, 640))
scale_factor = np.array([[float(640/img.shape[0]), float(640/img.shape[1])]]).astype('float32')
img = img.astype(np.float32) / 255.0
input_img = np.transpose(img, [2, 0, 1])
image = input_img[np.newaxis, :, :, :]
output_dict = ["reshape2_83.tmp_0", "tile_3.tmp_0"]
inputs_dict = {
    'im_shape': im_shape,
    'image': image,
    'scale_factor': scale_factor
}
result = model(inputs_dict)['bbox']

虽然并不懂为什么。我个人认为这个代码是错的，希望能有大神来指导一下。根据https://github.com/PaddlePaddle/PaddleDetection/blob/develop/deploy/EXPORT_MODEL.md，有以下几点不明白：

1. im_shape 是resize后的大小，但是这个代码里用的是resize之前的，我觉得应该是640x640。
2. scale factor 是输入图像大小比真实图像大小，但是这个代码里不管那种情况都是1，我觉得应该是640/im_shape

我原来的代码其实就是以上两点不同，但是不知道为什么。

[aliencaocao]

@lyuwenyu 希望大佬能帮忙解释一下谢谢

[lyuwenyu]

这块的逻辑可以看DETRPostProcess的逻辑, 就是im_shape和scale_factor这两者相除能得到origin_shape就行

https://github.com/PaddlePaddle/PaddleDetection/blob/develop/ppdet/modeling/post_process.py#L507

[lyuwenyu]

用trt推理的结果没啥问题，你那个结果应该就是im_shape, scale_factor 这些参数不太对

[aliencaocao]

是的，确实是这两个参数不对，谢谢解答

[luoshiyong]

大佬好，遇到了同样的问题(box全零，labels和scores正常)，使用了@aliencaocao的推理代码后box变成全是inf。。。