TensorRT-Demo/utils/yolo_with_plugins.py

"""yolo_with_plugins.py

Implementation of TrtYOLO class with the yolo_layer plugins.
"""


from __future__ import print_function

import ctypes

import numpy as np
import cv2
import tensorrt as trt
import pycuda.driver as cuda


try:
    ctypes.cdll.LoadLibrary('./plugins/libyolo_layer.so')
except OSError as e:
    raise SystemExit('ERROR: failed to load ./plugins/libyolo_layer.so.  '
                     'Did you forget to do a "make" in the "./plugins/" '
                     'subdirectory?') from e


def _preprocess_yolo(img, input_shape, letter_box=False):
    """Preprocess an image before TRT YOLO inferencing.

    # Args
        img: int8 numpy array of shape (img_h, img_w, 3)
        input_shape: a tuple of (H, W)
        letter_box: boolean, specifies whether to keep aspect ratio and
                    create a "letterboxed" image for inference

    # Returns
        preprocessed img: float32 numpy array of shape (3, H, W)
    """
    if letter_box:
        img_h, img_w, _ = img.shape
        new_h, new_w = input_shape[0], input_shape[1]
        offset_h, offset_w = 0, 0
        if (new_w / img_w) <= (new_h / img_h):
            new_h = int(img_h * new_w / img_w)
            offset_h = (input_shape[0] - new_h) // 2
        else:
            new_w = int(img_w * new_h / img_h)
            offset_w = (input_shape[1] - new_w) // 2
        resized = cv2.resize(img, (new_w, new_h))
        img = np.full((input_shape[0], input_shape[1], 3), 127, dtype=np.uint8)
        img[offset_h:(offset_h + new_h), offset_w:(offset_w + new_w), :] = resized
    else:
        img = cv2.resize(img, (input_shape[1], input_shape[0]))

    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = img.transpose((2, 0, 1)).astype(np.float32)
    img /= 255.0
    return img


def _nms_boxes(detections, nms_threshold):
    """Apply the Non-Maximum Suppression (NMS) algorithm on the bounding
    boxes with their confidence scores and return an array with the
    indexes of the bounding boxes we want to keep.

    # Args
        detections: Nx7 numpy arrays of
                    [[x, y, w, h, box_confidence, class_id, class_prob],
                     ......]
    """
    x_coord = detections[:, 0]
    y_coord = detections[:, 1]
    width = detections[:, 2]
    height = detections[:, 3]
    box_confidences = detections[:, 4] * detections[:, 6]

    areas = width * height
    ordered = box_confidences.argsort()[::-1]

    keep = list()
    while ordered.size > 0:
        # Index of the current element:
        i = ordered[0]
        keep.append(i)
        xx1 = np.maximum(x_coord[i], x_coord[ordered[1:]])
        yy1 = np.maximum(y_coord[i], y_coord[ordered[1:]])
        xx2 = np.minimum(x_coord[i] + width[i], x_coord[ordered[1:]] + width[ordered[1:]])
        yy2 = np.minimum(y_coord[i] + height[i], y_coord[ordered[1:]] + height[ordered[1:]])

        width1 = np.maximum(0.0, xx2 - xx1 + 1)
        height1 = np.maximum(0.0, yy2 - yy1 + 1)
        intersection = width1 * height1
        union = (areas[i] + areas[ordered[1:]] - intersection)
        iou = intersection / union
        indexes = np.where(iou <= nms_threshold)[0]
        ordered = ordered[indexes + 1]

    keep = np.array(keep)
    return keep


def _postprocess_yolo(trt_outputs, img_w, img_h, conf_th, nms_threshold,
                      input_shape, letter_box=False):
    """Postprocess TensorRT outputs.

    # Args
        trt_outputs: a list of 2 or 3 tensors, where each tensor
                    contains a multiple of 7 float32 numbers in
                    the order of [x, y, w, h, box_confidence, class_id, class_prob]
        conf_th: confidence threshold
        letter_box: boolean, referring to _preprocess_yolo()

    # Returns
        boxes, scores, classes (after NMS)
    """
    # filter low-conf detections and concatenate results of all yolo layers
    # 输出应该是[3*(80 + c + w + h + x + y)),(13+26+52),(13+26+52)]
    # 猜测是应该将80分类换算成了1个id号和1个置信度，这可能就是yolo更换输出的原因，确实是在插件中更换的输出类型
    detections = []
    for o in trt_outputs:
        # x, y, w, h , c , id , score
        dets = o.reshape((-1, 7))
        dets = dets[dets[:, 4] * dets[:, 6] >= conf_th]
        detections.append(dets)
    detections = np.concatenate(detections, axis=0)

    if len(detections) == 0:
        boxes = np.zeros((0, 4), dtype=np.int)
        scores = np.zeros((0,), dtype=np.float32)
        classes = np.zeros((0,), dtype=np.float32)
    else:
        box_scores = detections[:, 4] * detections[:, 6]

        # scale x, y, w, h from [0, 1] to pixel values
        old_h, old_w = img_h, img_w
        offset_h, offset_w = 0, 0
        if letter_box:
            if (img_w / input_shape[1]) >= (img_h / input_shape[0]):
                old_h = int(input_shape[0] * img_w / input_shape[1])
                offset_h = (old_h - img_h) // 2
            else:
                old_w = int(input_shape[1] * img_h / input_shape[0])
                offset_w = (old_w - img_w) // 2
        detections[:, 0:4] *= np.array(
            [old_w, old_h, old_w, old_h], dtype=np.float32)

        # NMS
        nms_detections = np.zeros((0, 7), dtype=detections.dtype)
        for class_id in set(detections[:, 5]):
            idxs = np.where(detections[:, 5] == class_id)
            cls_detections = detections[idxs]
            keep = _nms_boxes(cls_detections, nms_threshold)
            nms_detections = np.concatenate(
                [nms_detections, cls_detections[keep]], axis=0)

        xx = nms_detections[:, 0].reshape(-1, 1)
        yy = nms_detections[:, 1].reshape(-1, 1)
        if letter_box:
            xx = xx - offset_w
            yy = yy - offset_h
        ww = nms_detections[:, 2].reshape(-1, 1)
        hh = nms_detections[:, 3].reshape(-1, 1)
        boxes = np.concatenate([xx, yy, xx+ww, yy+hh], axis=1) + 0.5
        boxes = boxes.astype(np.int)
        scores = nms_detections[:, 4] * nms_detections[:, 6]
        classes = nms_detections[:, 5]
    return boxes, scores, classes


class HostDeviceMem(object):
    """Simple helper data class that's a little nicer to use than a 2-tuple."""
    def __init__(self, host_mem, device_mem):
        self.host = host_mem
        self.device = device_mem

    def __str__(self):
        return "Host:\n" + str(self.host) + "\nDevice:\n" + str(self.device)

    def __repr__(self):
        return self.__str__()

    def __del__(self):
        del self.device
        del self.host        


def get_input_shape(engine):
    """Get input shape of the TensorRT YOLO engine."""
    binding = engine[0]
    assert engine.binding_is_input(binding)
    binding_dims = engine.get_binding_shape(binding)
    if len(binding_dims) == 4:
        return tuple(binding_dims[2:])
    elif len(binding_dims) == 3:
        return tuple(binding_dims[1:])
    else:
        raise ValueError('bad dims of binding %s: %s' % (binding, str(binding_dims)))


def allocate_buffers(engine):
    """Allocates all host/device in/out buffers required for an engine."""
    inputs = []
    outputs = []
    bindings = []
    output_idx = 0
    stream = cuda.Stream()
    for binding in engine:
        binding_dims = engine.get_binding_shape(binding)
        if len(binding_dims) == 4:
            # explicit batch case (TensorRT 7+)
            size = trt.volume(binding_dims)
        elif len(binding_dims) == 3:
            # implicit batch case (TensorRT 6 or older)
            size = trt.volume(binding_dims) * engine.max_batch_size
        else:
            raise ValueError('bad dims of binding %s: %s' % (binding, str(binding_dims)))
        dtype = trt.nptype(engine.get_binding_dtype(binding))
        # Allocate host and device buffers
        host_mem = cuda.pagelocked_empty(size, dtype)
        device_mem = cuda.mem_alloc(host_mem.nbytes)
        # Append the device buffer to device bindings.
        bindings.append(int(device_mem))
        # Append to the appropriate list.
        if engine.binding_is_input(binding):
            inputs.append(HostDeviceMem(host_mem, device_mem))
        else:
            # each grid has 3 anchors, each anchor generates a detection
            # output of 7 float32 values
            assert size % 7 == 0
            outputs.append(HostDeviceMem(host_mem, device_mem))
            output_idx += 1
    assert len(inputs) == 1
    assert len(outputs) == 1
    return inputs, outputs, bindings, stream


def do_inference(context, bindings, inputs, outputs, stream, batch_size=1):
    """do_inference (for TensorRT 6.x or lower)

    This function is generalized for multiple inputs/outputs.
    Inputs and outputs are expected to be lists of HostDeviceMem objects.
    """
    # Transfer input data to the GPU.
    [cuda.memcpy_htod_async(inp.device, inp.host, stream) for inp in inputs]
    # Run inference.
    context.execute_async(batch_size=batch_size,
                          bindings=bindings,
                          stream_handle=stream.handle)
    # Transfer predictions back from the GPU.
    [cuda.memcpy_dtoh_async(out.host, out.device, stream) for out in outputs]
    # Synchronize the stream
    stream.synchronize()
    # Return only the host outputs.
    return [out.host for out in outputs]


def do_inference_v2(context, bindings, inputs, outputs, stream):
    """do_inference_v2 (for TensorRT 7.0+)

    This function is generalized for multiple inputs/outputs for full
    dimension networks.
    Inputs and outputs are expected to be lists of HostDeviceMem objects.
    """
    # Transfer input data to the GPU.
    [cuda.memcpy_htod_async(inp.device, inp.host, stream) for inp in inputs]
    # Run inference.
    context.execute_async_v2(bindings=bindings, stream_handle=stream.handle)
    # Transfer predictions back from the GPU.
    [cuda.memcpy_dtoh_async(out.host, out.device, stream) for out in outputs]
    # Synchronize the stream
    stream.synchronize()
    # Return only the host outputs.
    return [out.host for out in outputs]


class TrtYOLO(object):
    """TrtYOLO class encapsulates things needed to run TRT YOLO."""

    def _load_engine(self):
        TRTbin = 'yolo/%s.trt' % self.model
        with open(TRTbin, 'rb') as f, trt.Runtime(self.trt_logger) as runtime:
            return runtime.deserialize_cuda_engine(f.read())

    def __init__(self, model, category_num=80, letter_box=False, cuda_ctx=None):
        """Initialize TensorRT plugins, engine and conetxt."""
        # 保存engine模型
        self.model = model
        # 保存分类数
        self.category_num = category_num
        # 统一输入大小到letterbox
        self.letter_box = letter_box
        # 默认CUDA上下文只能从创建它的CPU线程访问，其他线程访问需push/pop从创建它的线程中弹出它，这样context可以被推送到任何其他CPU线程的当前上下文栈，并且随后的CUDA调用将引用该上下文。
        self.cuda_ctx = cuda_ctx
        if self.cuda_ctx:
            self.cuda_ctx.push()
        # 设置推理函数
        self.inference_fn = do_inference if trt.__version__[0] < '7' \
                                         else do_inference_v2
        # 打印日志，启动一个logging界面，抑制warning和errors，仅报告informational messages。
        self.trt_logger = trt.Logger(trt.Logger.INFO)
        # 加载模型deserialize
        self.engine = self._load_engine()
        # 从模型中获取输入大小
        self.input_shape = get_input_shape(self.engine)

        try:
            # 创建一个上下文，储存中间值，因为engine包含network定义和训练参数，因此需要额外的空间。
            self.context = self.engine.create_execution_context() # create_execution_context是写在ICudaEngine.py的一个闭源方法，这个方法是创建立一个IExecutionContext类型的对象。
            self.inputs, self.outputs, self.bindings, self.stream = \
                allocate_buffers(self.engine)   # 为输入输出分配host和device的buffers。host指的是CPU内存，device指的是GPU显存
        except Exception as e:
            raise RuntimeError('fail to allocate CUDA resources') from e
        finally:
            if self.cuda_ctx:
                self.cuda_ctx.pop()

    def __del__(self):
        """Free CUDA memories."""
        del self.outputs
        del self.inputs
        del self.stream

    def detect(self, img, conf_th=0.3, letter_box=None):
        """Detect objects in the input image."""
        letter_box = self.letter_box if letter_box is None else letter_box
        # 保证输入源统一大小，符合推理模型使用
        img_resized = _preprocess_yolo(img, self.input_shape, letter_box)

        # Set host input to the image. The do_inference() function
        # will copy the input to the GPU before executing.
        # 开辟一块内存空间，用于放入输入图像
        self.inputs[0].host = np.ascontiguousarray(img_resized)
        if self.cuda_ctx:
            self.cuda_ctx.push()
        # 开始推理
        trt_outputs = self.inference_fn(
            context=self.context,   # 制定GPU的Context，可以理解为上下文，{}
            bindings=self.bindings, # 大概指的是内存到显存之间的绑定关系
            inputs=self.inputs,     # 输入数据
            outputs=self.outputs,   # 输出数据
            stream=self.stream)     # cuda的操作顺序流
        if self.cuda_ctx:
            self.cuda_ctx.pop()

        # 后处理GPU返回的输出结果
        boxes, scores, classes = _postprocess_yolo(
            trt_outputs, img.shape[1], img.shape[0], conf_th,
            nms_threshold=0.5, input_shape=self.input_shape,
            letter_box=letter_box)

        # clip x1, y1, x2, y2 within original image
        boxes[:, [0, 2]] = np.clip(boxes[:, [0, 2]], 0, img.shape[1]-1)
        boxes[:, [1, 3]] = np.clip(boxes[:, [1, 3]], 0, img.shape[0]-1)
        return boxes, scores, classes