1.前言

RKNN出最新版本了，测试了一下，rk在transformer方面做了很多的工作，至少之前不能转的模型，现在可以在fp16上面运行了，在测试int8的时候还是有误差，以往后面优化吧，这一篇是DETR模型转rknn的fp16模型的过程。

2.准备工作

PC: ubuntu 18.04、rknntoolkit2-1.5

开发板:rk3588

模型链接: onnx模型提取码: yciw

关于onnx模型怎样来的，请参考博文DERT(DEtection TRansformer) ONNX直接推理！！

这里模型链接中onnx模型做了一点修改，将模型最后的两个gather算子删除了，这样转化才不出错(有心的同学可以对比一下参考博文的onnx模型和本文中的onnx模型最后的输出)

3.开始转模型

import numpy as np
import cv2
from rknn.api import RKNN

ONNX_MODEL = 'modified_models.onnx'
RKNN_MODEL = 'detr_fp16.rknn'
DATASET = './dataset.txt'
QUANTIZE_ON = True
QUANTIZE_OFF = False

if __name__ == '__main__':

    # Create RKNN object
    rknn = RKNN(verbose=True)

    # pre-process config
    print('--> Config model')
    rknn.config(mean_values=[[0, 0, 0]], std_values=[[1, 1, 1]], target_platform='rk3588')
    print('done')

    # Load ONNX model
    print('--> Loading model')
    ret = rknn.load_onnx(model=ONNX_MODEL)
    if ret != 0:
        print('Load model failed!')
        exit(ret)
    print('done')

    # Build model
    print('--> Building model')
    ret = rknn.build(do_quantization=QUANTIZE_OFF, dataset=DATASET)
    if ret != 0:
        print('Build model failed!')
        exit(ret)
    print('done')
    # Export RKNN model
    print('--> Export rknn model')
    ret = rknn.export_rknn(RKNN_MODEL)
    if ret != 0:
        print('Export rknn model failed!')
        exit(ret)
    print('done')

准换后就有了detr_fp16的模型了

4.测试代码

import numpy as np
from PIL import Image
from PIL import ImageDraw, ImageFont
import colorsys
from rknnlite.api import RKNNLite


def get_classes(classes_path):
    with open(classes_path, encoding='utf-8') as f:
        class_names = f.readlines()
    class_names = [c.strip() for c in class_names]
    return class_names, len(class_names)


def get_new_img_size(height, width, min_length=600):
    if width <= height:
        f = float(min_length) / width
        resized_height = int(f * height)
        resized_width = int(min_length)
    else:
        f = float(min_length) / height
        resized_width = int(f * width)
        resized_height = int(min_length)

    return resized_height, resized_width


def resize_image(image, min_length):
    iw, ih = image.size
    h, w = get_new_img_size(ih, iw, min_length=min_length)
    new_image = image.resize((w, h), Image.BICUBIC)
    return new_image


def cvtColor(image):
    if len(np.shape(image)) == 3 and np.shape(image)[2] == 3:
        return image
    else:
        image = image.convert('RGB')
        return image


class DecodeBox:
    """ This module converts the model's output into the format expected by the coco api"""

    def box_cxcywh_to_xyxy(self, x):
        x_c, y_c, w, h = x[..., 0], x[..., 1], x[..., 2], x[..., 3]
        b = [(x_c - 0.5 * w), (y_c - 0.5 * h),
             (x_c + 0.5 * w), (y_c + 0.5 * h)]
        return np.stack(b, axis=-1)

    def forward(self, outputs, target_sizes, confidence):
        out_logits, out_bbox = outputs["pred_logits"], outputs["pred_boxes"]

        assert len(out_logits) == len(target_sizes)
        assert target_sizes.shape[1] == 2

        prob = np.exp(out_logits) / np.exp(out_logits).sum(-1, keepdims=True)
        scores = np.max(prob[..., :-1], axis=-1)
        labels = np.argmax(prob[..., :-1], axis=-1)  # 加1来转换为类别标签（背景类别为0）


        # convert to [x0, y0, x1, y1] format
        boxes = self.box_cxcywh_to_xyxy(out_bbox)

        # and from relative [0, 1] to absolute [0, height] coordinates
        img_h, img_w = np.split(target_sizes, target_sizes.shape[1], axis=1)[0], np.split(target_sizes, target_sizes.shape[1], axis=1)[1]
        img_h = img_h.astype(float)
        img_w = img_w.astype(float)
        scale_fct = np.hstack([img_w, img_h, img_w, img_h])
        boxes = boxes * scale_fct[:, None, :]

        outputs = np.concatenate([
            np.expand_dims(boxes[:, :, 1], -1),
            np.expand_dims(boxes[:, :, 0], -1),
            np.expand_dims(boxes[:, :, 3], -1),
            np.expand_dims(boxes[:, :, 2], -1),
            np.expand_dims(scores, -1),
            np.expand_dims(labels.astype(float), -1),
        ], -1)

        results = []
        for output in outputs:
            results.append(output[output[:, 4] > confidence])
        # results = [{'scores': s, 'labels': l, 'boxes': b} for s, l, b in zip(scores, labels, boxes)]
        return results


def preprocess_input(image):
    image /= 255.0
    image -= np.array([0.485, 0.456, 0.406])
    image /= np.array([0.229, 0.224, 0.225])
    return image


if __name__ == "__main__":

    count = True
    confidence = 0.5
    min_length = 512
    image = Image.open('1.jpg')
    image = image.resize((512, 512))
    image_shape = np.array([np.shape(image)[0:2]])
    image = cvtColor(image)
    image_data = resize_image(image, min_length)
    # image_data = np.expand_dims(np.transpose(preprocess_input(np.array(image_data, dtype='float32')), (2, 0, 1)), 0)
    image_data = np.expand_dims(preprocess_input(np.array(image_data, dtype='float32')), 0)
    print(image_data.shape)
    model_name = "./detr_fp16.rknn"
    rknn_lite = RKNNLite()

    # load RKNN model
    print('--> Load RKNN model')
    ret = rknn_lite.load_rknn(model_name)
    if ret != 0:
        print('Load RKNN model failed')
        exit(ret)
    print('done')
    ret = rknn_lite.init_runtime(core_mask=RKNNLite.NPU_CORE_0)
    # Inference
    print('--> Running model')
    net_outputs = rknn_lite.inference(inputs=[image_data])
    net_outs = {"pred_logits": net_outputs[0][-1], "pred_boxes": net_outputs[1][-1]}


  
    bbox_util = DecodeBox()
    results = bbox_util.forward(net_outs, image_shape, confidence)

    if results[0] is None:
        print('NO OBJECT')
    else:
        _results = results[0]
        top_label = np.array(_results[:, 5], dtype='int32')
        top_conf = _results[:, 4]
        top_boxes = _results[:, :4]
        font = ImageFont.truetype(font='model_data/simhei.ttf', size=np.floor(3e-2 * image.size[1] + 0.5).astype('int32'))
        thickness = int(max((image.size[0] + image.size[1]) // min_length, 1))
        classes_path = 'model_data/coco_classes.txt'
        class_names, num_classes = get_classes(classes_path)
        hsv_tuples = [(x / num_classes, 1., 1.) for x in range(num_classes)]
        colors = list(map(lambda x: colorsys.hsv_to_rgb(*x), hsv_tuples))
        colors = list(map(lambda x: (int(x[0] * 255), int(x[1] * 255), int(x[2] * 255)), colors))

        for i, c in list(enumerate(top_label)):
            predicted_class = class_names[int(c)]
            box = top_boxes[i]
            score = top_conf[i]
            top, left, bottom, right = box
            top = max(0, np.floor(top).astype('int32'))
            left = max(0, np.floor(left).astype('int32'))
            bottom = min(image.size[1], np.floor(bottom).astype('int32'))
            right = min(image.size[0], np.floor(right).astype('int32'))

            label = '{} {:.2f}'.format(predicted_class, score)
            draw = ImageDraw.Draw(image)
            label_size = draw.textsize(label, font)
            label = label.encode('utf-8')
            print(label, top, left, bottom, right)

            if top - label_size[1] >= 0:
                text_origin = np.array([left, top - label_size[1]])
            else:
                text_origin = np.array([left, top + 1])

            for i in range(thickness):
                draw.rectangle([left + i, top + i, right - i, bottom - i], outline=colors[c])
            draw.rectangle([tuple(text_origin), tuple(text_origin + label_size)], fill=colors[c])
            draw.text(text_origin, str(label, 'UTF-8'), fill=(0, 0, 0), font=font)
            del draw
        image.save('output.png')

测试的结果如下，还是不错的。