使用tensorflow来复现yolo-v3

yolo-v3是yolo系列的最新版本，用于物体检测任务中，效果较好，能同时满足精度和实时性的要求。
但是yolo-v3使用的是作者自己写的darknet框架，虽然darknet框架使用纯c和cuda编写而成，短小精悍，简单而不失性能，但是由于其简单性，意味着耦合性较高，使用者想要自己作出一些改动就需要修改源码。
使用tensorflow将其复现，可以方便作出一些调整。

为了保障性能，预测阶段前向传播的算法都是用tensorflow的操作来实现，没有用python原生的循环、计算来完成，这样前向传播的耗时为20ms，做到和yolo-v3原版的darknet框架差不多。

github地址：https://github.com/zhmc/yolov3_tensorflow

下面是yolo-v3预测阶段的代码。

from __future__ import print_function

from __future__ import division

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "1"

import numpy as np
import tensorflow as tf

import time

import matplotlib.pyplot as plt

%matplotlib inline

import cv2

input_dim = 416

bn_variance_epsilon = 1e-6


batch_size = 1

confidence_threshold = 0.55

iou_threshold=0.45

max_output_size = 1000

def print_tensor_info(t):
      print(t.op.name, ' ', t.get_shape().as_list())

# !python -m wget https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/yolov3.cfg 

blockList = []
current_block = {}
with open('cfg/yolov3.cfg') as cfg:
    line = cfg.readline()
    while line:
        if line.startswith("#"):
            line = cfg.readline()
            continue
        if line.startswith("["):
            if current_block != {}:
                blockList.append(current_block)
            current_block = {}
            current_block["type"] = line[1:-2]
        else:
            if "=" in line:
    #             print(line)
                key, value = line.split("=")
                key = key.strip()
                value = value.strip()
    #             print(key, value)
                current_block[key] = value

        line = cfg.readline()
    if current_block != {}:
        blockList.append(current_block)

yolo_blocks = []
for block in blockList:
    if block["type"] == "yolo":
        yolo_blocks.append(block)

yolo_blocks[0]

{'type': 'yolo',
 'mask': '6,7,8',
 'anchors': '10,13,  16,30,  33,23,  30,61,  62,45,  59,119,  116,90,  156,198,  373,326',
 'classes': '80',
 'num': '9',
 'jitter': '.3',
 'ignore_thresh': '.7',
 'truth_thresh': '1',
 'random': '1'}

def yolo_layer(forward_input, block_info):
    shape = forward_input.get_shape()
    
    if len(shape.as_list()) == 4:
        shape_list = [int(shape[0]), int(shape[1]), int(shape[2]), 3, int(shape[3]) // 3]
        yolo_tensor = tf.reshape(forward_input, shape_list)
    else:
        yolo_tensor = forward_input
    # yolo_tensor的形状为(1, 13, 13, 3, 85)
    # 85 = 1 （objectness) + 4 (bbox: x_center_in_gird, y_center_in_gird, width, height) + 80 (80 classes confidence)
    
    # stride from input image to feature map
    # 数据输入为416 * 416， 那么第一个yolo层是13 * 13， stride是32.
    # 第二个yolo层的stride为16，第三个为8
    stride = input_dim // shape.as_list()[1]

    
    # 总共9个prior box大小，每个yolo层使用属于它的3个
    mask_list = [ int(x) for x in block_info["mask"].split(",")]
    anchors_list = [ int(x.strip()) for x in block_info["anchors"].split(",")]
    prior_width_height = [ tuple([anchors_list[2*i], anchors_list[2*i + 1] ]) for i in mask_list]
    
    assert(len(prior_width_height) == 3)
    
    # 三个prior box 输出的bbox

    box_width_height1 = tf.exp(yolo_tensor[:, :, :, 0:1, 2:4]) * prior_width_height[0] 
    box_width_height2 = tf.exp(yolo_tensor[:, :, :, 1:2, 2:4]) * prior_width_height[1] 
    box_width_height3 = tf.exp(yolo_tensor[:, :, :, 2:3, 2:4]) * prior_width_height[2] 
    

    box_width_height = tf.concat([box_width_height1, box_width_height2, box_width_height3], axis=3)

    
    confidence_tensor = tf.sigmoid(yolo_tensor[:, :, :, :, 4:])

    
    # 过滤出objectness大于threshold的物体索引
    # result_indexes的形状为( N, H, W, anchor)
    result_indexes = tf.where(tf.greater_equal(tf.sigmoid(confidence_tensor[:, :, :, :, 0]), confidence_threshold))
    
    
    # TODO 此处若无+1 操作，则bbox坐标不对
#     gird_x_y_raw = tf.cast(result_indexes[:,1:3],  tf.float32)
    gird_x_y_raw = tf.cast(result_indexes[:,1:3],  tf.float32) + 1
    
    # indexes 中坐标为（N, H, W, anchor) 需要将H,W 互换
    gird_x_y = tf.concat([gird_x_y_raw[:,1:], gird_x_y_raw[:,0:1]] , axis=1)
    
    
    #过滤检测结果
    result_boxes = tf.gather_nd(yolo_tensor, result_indexes)
    
    #x, y 坐标是在这个网格中的相对坐标
    center_x_y = tf.sigmoid(result_boxes[:, 0:2])
    # 转化为全图绝对坐标
    output_center_x_y = (gird_x_y + center_x_y) * stride
    
    output_width_height =  tf.gather_nd(box_width_height, result_indexes)
    
    x_min = tf.maximum(output_center_x_y[:, 0:1] - 0.5 * output_width_height[:, 0:1], 0)
    x_max = tf.minimum(output_center_x_y[:, 0:1] + 0.5 * output_width_height[:, 0:1], input_dim)
    y_min = tf.maximum(output_center_x_y[:, 1:2] - 0.5 * output_width_height[:, 1:2], 0)
    y_max = tf.minimum(output_center_x_y[:, 1:2] + 0.5 * output_width_height[:, 1:2], input_dim)

    output_xmin_ymin_xmax_y_max = tf.concat([x_min, y_min, x_max, y_max], axis=1)
    
    # 将绝对的坐标值转为相对坐标值
    output_xmin_ymin_xmax_y_max_float = output_xmin_ymin_xmax_y_max / input_dim
        
    classes_confidence = tf.gather_nd(confidence_tensor[:, :, :, :,1:], result_indexes)
    index_int = tf.argmax(classes_confidence ,axis=1)[ :, tf.newaxis]
    max_index = tf.cast(index_int, dtype=tf.float32)
    max_confidence = tf.reduce_max(classes_confidence ,axis=1)[ :, tf.newaxis]
    
    # 输出的张量形状为 (None, 6)
    # 6 = 4 (bbox),  1 (class_index), 1 (class_confidence)
    yolo_output = tf.concat([output_xmin_ymin_xmax_y_max_float,  max_index, max_confidence], axis=1)
    return yolo_output

def make_one_layer(forward_input, block, layersList, np_weights,  weights_ptr):
    # 根据block信息，从上一层构建下一层
    
    forward_output = None
    
    # 读取yolov3.weights权重文件的指针
    next_weights_ptr = 0
    
    if block['type'] == "net":
#         forward_output = tf.placeholder(dtype=tf.float32, shape = 
#                                         [batch_size, int(block["height"]), int(block["width"]), int(block["channels"])])
        forward_output = tf.placeholder(dtype=tf.float32, shape = 
                                        [batch_size, input_dim, input_dim, int(block["channels"])])
        with tf.name_scope("input") as scope:
            return tf.identity(forward_output, name="x" ), weights_ptr
    if block["type"] == "convolutional":
        input_filters = int(forward_input.get_shape()[3])
        output_filters =  int(block["filters"])
        kernel_size = int(block["size"])
        kernel_shape = [int(block["size"]),  int(block["size"]), input_filters, output_filters]
        stride = int(block["stride"])
        strides = [1, stride, stride, 1]
        pad = (kernel_size - 1) / 2 
        # batch norm load more weights
        try:
            batch_norm = block["batch_normalize"]
        except:
            batch_norm = 0
        if batch_norm:
            
            #batch norm weights
            bn_offset = np_weights[weights_ptr:weights_ptr+output_filters]
            weights_ptr += output_filters
            
            bn_scale = np_weights[weights_ptr:weights_ptr+output_filters]
            weights_ptr += output_filters
            
            bn_mean = np_weights[weights_ptr:weights_ptr+output_filters]
            weights_ptr += output_filters
            
            bn_var = np_weights[weights_ptr:weights_ptr+output_filters]
            weights_ptr += output_filters
            
            #conv filter weights
            conv_weights_num = output_filters * input_filters * kernel_size * kernel_size
            np_conv_weights = np_weights[weights_ptr:weights_ptr+conv_weights_num]
            weights_ptr += conv_weights_num
            np_conv_weights = np_conv_weights.reshape( output_filters, input_filters, kernel_size, kernel_size)
#             print(np_conv_weights.shape)
#             print(tf.constant(np_conv_weights, dtype=tf.float32))

#             conv_kernel =  (tf.transpose(tf.constant(np_conv_weights, dtype=tf.float32), [2, 3, 1, 0]))
            # 此处之前使用tensorflow来做权重张量转置操作，很费时。不需要重复转置，所以使用numpy一次转置就够了
            np_conv_weights = np.transpose(np_conv_weights, [2, 3, 1, 0])
            conv_kernel =  tf.constant(np_conv_weights, dtype=tf.float32)
    
            conv_output = tf.nn.conv2d(forward_input, conv_kernel, strides, "SAME")
            
            bn_output = tf.nn.batch_normalization(conv_output, tf.constant(bn_mean, dtype=tf.float32), 
                                                 tf.constant(bn_var, dtype=tf.float32), tf.constant(bn_offset, dtype=tf.float32),
                                                 tf.constant(bn_scale, dtype=tf.float32), bn_variance_epsilon)
            
            if block["activation"] == "leaky":
                leaky_relu_outpu = tf.nn.leaky_relu(bn_output, alpha=0.1)
                # layersList.append(leaky_relu_outpu)
                with tf.name_scope("conv"+str(len(layersList))) as scope:
                    return tf.identity(leaky_relu_outpu, name=scope ), weights_ptr
            
            # layersList.append(bn_output)
            raise Exception("unhandled condition")
            return bn_output, weights_ptr
        else:
            # no batch_norm
            
            #conv bias
            conv_bias = np_weights[weights_ptr:weights_ptr+output_filters]
            weights_ptr += output_filters
            
            #conv filter weights
            conv_weights_num = output_filters * input_filters * kernel_size * kernel_size
            np_conv_weights = np_weights[weights_ptr:weights_ptr+conv_weights_num]
            weights_ptr += conv_weights_num
            np_conv_weights = np_conv_weights.reshape( output_filters, input_filters, kernel_size, kernel_size)
            
#             conv_kernel =  (tf.transpose(tf.constant(np_conv_weights, dtype=tf.float32), [2, 3, 1, 0]))
            
            np_conv_weights = np.transpose(np_conv_weights, [2, 3, 1, 0])
            conv_kernel =  tf.constant(np_conv_weights, dtype=tf.float32)
            
            
            conv_output = tf.nn.conv2d(forward_input, conv_kernel, strides, "SAME")
            
            conv_bias_output = tf.nn.bias_add(conv_output, tf.constant(conv_bias, dtype=tf.float32))
            
            # layersList.append(conv_bias_output)
            with tf.name_scope("conv"+str(len(layersList))) as scope:
                return tf.identity(conv_bias_output, name=scope), weights_ptr
    
    if  block["type"] == "route":
        layers = []
        if "," not in block["layers"]:
            layers.append(int(block["layers"]))
        else:
            for l in block["layers"].split(","):
                layers.append(int(l.strip()))
        
#         for l in layers:
#             if l <0:
#                 l = len(layersList) + l


        # 此处，因为layerlist开头多了一个input_data的layer，所以concat的位置应该向后一位
        layers_new=[]
        for l in layers:
            if l<0:
                layers_new.append(len(layersList) + l)
            else:
                layers_new.append( l +1 )
        layers = layers_new
        
        if len(layers) == 1:
            with tf.name_scope("route"+str(len(layersList))) as scope:
                return tf.identity(layersList[layers[0]], name= scope), weights_ptr
        else:
            
            to_concat_list = [layersList[i] for i in layers]
            concat_result =tf.concat( to_concat_list, axis=3)
            
            with tf.name_scope("route"+str(len(layersList))) as scope:
                return tf.identity(concat_result, name=scope), weights_ptr
        

    if  block["type"] == "shortcut":
        prior_index = int(block["from"])
#         print("prior_index = ", prior_index)
#         print("layersList[prior_index]", layersList[prior_index])
        residual_layer = forward_input + layersList[prior_index]
    
        with tf.name_scope("shortcut"+str(len(layersList))) as scope:
            return tf.identity(residual_layer, name= scope), weights_ptr
        
    
    if  block["type"] == "upsample":
        new_height = int(block["stride"]) * int(forward_input.get_shape()[1])
        new_width = int(block["stride"]) * int(forward_input.get_shape()[2])
        
        upsample_layer = tf.image.resize_bilinear(forward_input, [new_height, new_width])
        
        with tf.name_scope("upsample"+str(len(layersList))) as scope:
            return tf.identity(upsample_layer, name= scope), weights_ptr
        
    
    if  block["type"] == "yolo":
        
        with tf.name_scope("yolo"+str(len(layersList))) as scope:
            return tf.identity(yolo_layer(forward_input, block_info=block), name=scope), weights_ptr
        
    
    raise Exception('unkown block type')

# !python -m wget https://pjreddie.com/media/files/yolov3.weights

fp = open("cfg/yolov3.weights", "rb")
#The first 5 values are header information 
# 1. Major version number
# 2. Minor Version Number
# 3. Subversion number 
# 4,5. Images seen by the network (during training)
header = np.fromfile(fp, dtype = np.int32, count = 5)
print(header)

[       0        2        0 32013312        0]

yolo_weights = np.fromfile(fp, dtype = np.float32)

yolo_weights.shape

(62001757,)

layers_list = []
ptr = 0
temp_layer =None
yolo_out_list = []

for block in blockList:
#     print(block["type"])
    pre_ptr = ptr
    temp_layer, ptr = make_one_layer(forward_input=temp_layer, block=block, 
                         layersList=layers_list, np_weights=yolo_weights, weights_ptr=ptr)
    print_tensor_info(temp_layer)
    layers_list.append(temp_layer)
 
    if block["type"] == "yolo":
        yolo_out_list.append(temp_layer)
    if block["type"] == "convolutional":
        try:
            batch_norm = block["batch_normalize"]
        except:
            batch_norm = 0
        if batch_norm:
            print("convolution layer with batch_norm weights float num: %d" % (ptr - pre_ptr))
        else:
            print("convolution layer without batch_norm weights float num: %d" % (ptr - pre_ptr))
 

input/x   [1, 416, 416, 3]
conv1   [1, 416, 416, 32]
convolution layer with batch_norm weights float num: 992
conv2   [1, 208, 208, 64]
convolution layer with batch_norm weights float num: 18688
conv3   [1, 208, 208, 32]
convolution layer with batch_norm weights float num: 2176
conv4   [1, 208, 208, 64]
convolution layer with batch_norm weights float num: 18688
shortcut5   [1, 208, 208, 64]
conv6   [1, 104, 104, 128]
convolution layer with batch_norm weights float num: 74240
conv7   [1, 104, 104, 64]
convolution layer with batch_norm weights float num: 8448
conv8   [1, 104, 104, 128]
convolution layer with batch_norm weights float num: 74240
shortcut9   [1, 104, 104, 128]
conv10   [1, 104, 104, 64]
convolution layer with batch_norm weights float num: 8448
conv11   [1, 104, 104, 128]
convolution layer with batch_norm weights float num: 74240
shortcut12   [1, 104, 104, 128]
conv13   [1, 52, 52, 256]
convolution layer with batch_norm weights float num: 295936
conv14   [1, 52, 52, 128]
convolution layer with batch_norm weights float num: 33280
conv15   [1, 52, 52, 256]
convolution layer with batch_norm weights float num: 295936
shortcut16   [1, 52, 52, 256]
conv17   [1, 52, 52, 128]
convolution layer with batch_norm weights float num: 33280
conv18   [1, 52, 52, 256]
convolution layer with batch_norm weights float num: 295936
shortcut19   [1, 52, 52, 256]
conv20   [1, 52, 52, 128]
convolution layer with batch_norm weights float num: 33280
conv21   [1, 52, 52, 256]
convolution layer with batch_norm weights float num: 295936
shortcut22   [1, 52, 52, 256]
conv23   [1, 52, 52, 128]
convolution layer with batch_norm weights float num: 33280
conv24   [1, 52, 52, 256]
convolution layer with batch_norm weights float num: 295936
shortcut25   [1, 52, 52, 256]
conv26   [1, 52, 52, 128]
convolution layer with batch_norm weights float num: 33280
conv27   [1, 52, 52, 256]
convolution layer with batch_norm weights float num: 295936
shortcut28   [1, 52, 52, 256]
conv29   [1, 52, 52, 128]
convolution layer with batch_norm weights float num: 33280
conv30   [1, 52, 52, 256]
convolution layer with batch_norm weights float num: 295936
shortcut31   [1, 52, 52, 256]
conv32   [1, 52, 52, 128]
convolution layer with batch_norm weights float num: 33280
conv33   [1, 52, 52, 256]
convolution layer with batch_norm weights float num: 295936
shortcut34   [1, 52, 52, 256]
conv35   [1, 52, 52, 128]
convolution layer with batch_norm weights float num: 33280
conv36   [1, 52, 52, 256]
convolution layer with batch_norm weights float num: 295936
shortcut37   [1, 52, 52, 256]
conv38   [1, 26, 26, 512]
convolution layer with batch_norm weights float num: 1181696
conv39   [1, 26, 26, 256]
convolution layer with batch_norm weights float num: 132096
conv40   [1, 26, 26, 512]
convolution layer with batch_norm weights float num: 1181696
shortcut41   [1, 26, 26, 512]
conv42   [1, 26, 26, 256]
convolution layer with batch_norm weights float num: 132096
conv43   [1, 26, 26, 512]
convolution layer with batch_norm weights float num: 1181696
shortcut44   [1, 26, 26, 512]
conv45   [1, 26, 26, 256]
convolution layer with batch_norm weights float num: 132096
conv46   [1, 26, 26, 512]
convolution layer with batch_norm weights float num: 1181696
shortcut47   [1, 26, 26, 512]
conv48   [1, 26, 26, 256]
convolution layer with batch_norm weights float num: 132096
conv49   [1, 26, 26, 512]
convolution layer with batch_norm weights float num: 1181696
shortcut50   [1, 26, 26, 512]
conv51   [1, 26, 26, 256]
convolution layer with batch_norm weights float num: 132096
conv52   [1, 26, 26, 512]
convolution layer with batch_norm weights float num: 1181696
shortcut53   [1, 26, 26, 512]
conv54   [1, 26, 26, 256]
convolution layer with batch_norm weights float num: 132096
conv55   [1, 26, 26, 512]
convolution layer with batch_norm weights float num: 1181696
shortcut56   [1, 26, 26, 512]
conv57   [1, 26, 26, 256]
convolution layer with batch_norm weights float num: 132096
conv58   [1, 26, 26, 512]
convolution layer with batch_norm weights float num: 1181696
shortcut59   [1, 26, 26, 512]
conv60   [1, 26, 26, 256]
convolution layer with batch_norm weights float num: 132096
conv61   [1, 26, 26, 512]
convolution layer with batch_norm weights float num: 1181696
shortcut62   [1, 26, 26, 512]
conv63   [1, 13, 13, 1024]
convolution layer with batch_norm weights float num: 4722688
conv64   [1, 13, 13, 512]
convolution layer with batch_norm weights float num: 526336
conv65   [1, 13, 13, 1024]
convolution layer with batch_norm weights float num: 4722688
shortcut66   [1, 13, 13, 1024]
conv67   [1, 13, 13, 512]
convolution layer with batch_norm weights float num: 526336
conv68   [1, 13, 13, 1024]
convolution layer with batch_norm weights float num: 4722688
shortcut69   [1, 13, 13, 1024]
conv70   [1, 13, 13, 512]
convolution layer with batch_norm weights float num: 526336
conv71   [1, 13, 13, 1024]
convolution layer with batch_norm weights float num: 4722688
shortcut72   [1, 13, 13, 1024]
conv73   [1, 13, 13, 512]
convolution layer with batch_norm weights float num: 526336
conv74   [1, 13, 13, 1024]
convolution layer with batch_norm weights float num: 4722688
shortcut75   [1, 13, 13, 1024]
conv76   [1, 13, 13, 512]
convolution layer with batch_norm weights float num: 526336
conv77   [1, 13, 13, 1024]
convolution layer with batch_norm weights float num: 4722688
conv78   [1, 13, 13, 512]
convolution layer with batch_norm weights float num: 526336
conv79   [1, 13, 13, 1024]
convolution layer with batch_norm weights float num: 4722688
conv80   [1, 13, 13, 512]
convolution layer with batch_norm weights float num: 526336
conv81   [1, 13, 13, 1024]
convolution layer with batch_norm weights float num: 4722688
conv82   [1, 13, 13, 255]
convolution layer without batch_norm weights float num: 261375
yolo83   [None, 6]
route84   [1, 13, 13, 512]
conv85   [1, 13, 13, 256]
convolution layer with batch_norm weights float num: 132096
upsample86   [1, 26, 26, 256]
route87   [1, 26, 26, 768]
conv88   [1, 26, 26, 256]
convolution layer with batch_norm weights float num: 197632
conv89   [1, 26, 26, 512]
convolution layer with batch_norm weights float num: 1181696
conv90   [1, 26, 26, 256]
convolution layer with batch_norm weights float num: 132096
conv91   [1, 26, 26, 512]
convolution layer with batch_norm weights float num: 1181696
conv92   [1, 26, 26, 256]
convolution layer with batch_norm weights float num: 132096
conv93   [1, 26, 26, 512]
convolution layer with batch_norm weights float num: 1181696
conv94   [1, 26, 26, 255]
convolution layer without batch_norm weights float num: 130815
yolo95   [None, 6]
route96   [1, 26, 26, 256]
conv97   [1, 26, 26, 128]
convolution layer with batch_norm weights float num: 33280
upsample98   [1, 52, 52, 128]
route99   [1, 52, 52, 384]
conv100   [1, 52, 52, 128]
convolution layer with batch_norm weights float num: 49664
conv101   [1, 52, 52, 256]
convolution layer with batch_norm weights float num: 295936
conv102   [1, 52, 52, 128]
convolution layer with batch_norm weights float num: 33280
conv103   [1, 52, 52, 256]
convolution layer with batch_norm weights float num: 295936
conv104   [1, 52, 52, 128]
convolution layer with batch_norm weights float num: 33280
conv105   [1, 52, 52, 256]
convolution layer with batch_norm weights float num: 295936
conv106   [1, 52, 52, 255]
convolution layer without batch_norm weights float num: 65535
yolo107   [None, 6]

assert(len(yolo_weights) - ptr == 0)

config = tf.ConfigProto()
config.gpu_options.allocator_type = 'BFC'
config.gpu_options.allow_growth=True
sess = tf.Session(config=config) 
init_op = tf.global_variables_initializer()
sess.run(init_op)

def nms(yolo_output_list):
    yolo_output_concat = tf.concat(yolo_output_list, axis=0)
    
    output_xmin_ymin_xmax_y_max = yolo_output_concat[:, 0:4]
    output_score = yolo_output_concat[:, 4]
    
    selected_index = tf.image.non_max_suppression(output_xmin_ymin_xmax_y_max, output_score,  max_output_size, iou_threshold)
    
    final_result = tf.gather(yolo_output_concat, selected_index)
    
    return final_result

y = nms(yolo_out_list)
y

<tf.Tensor 'GatherV2:0' shape=(?, 6) dtype=float32>

# !python -m wget https://raw.githubusercontent.com/pjreddie/darknet/master/data/dog.jpg

# !python -m wget https://raw.githubusercontent.com/pjreddie/darknet/master/data/horses.jpg

#  !python -m wget https://images2018.cnblogs.com/blog/1103587/201806/1103587-20180609102500112-1316429265.jpg -o car.jpg

plt.rcParams['figure.figsize'] = (20, 10)  

img =cv2.imread("data/car.jpg")
# img = cv2.resize(img,(input_dim,input_dim))
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

img_float = img.astype(np.float32) / 256
img_float = cv2.resize(img_float,(input_dim,input_dim))

img_batch =  img_float.reshape(1,input_dim,input_dim,3)
plt.imshow(img)

<matplotlib.image.AxesImage at 0x7f892014f588>

init_op = tf.global_variables_initializer()
# sess.run(test_y)
r = (sess.run(y, feed_dict={layers_list[0].name: img_batch}))

start_time = time.time()
rounds =20
result_list =[]
for i in range(rounds):
    _ = (sess.run(y, feed_dict={layers_list[0].name: img_batch}))
duration = time.time() - start_time
ms_count = ( float(duration)/rounds) * 1000
print("cost time per round: %.2f ms" % ms_count)


# print(r)

cost time per round: 22.38 ms

# !python -m wget https://raw.githubusercontent.com/pjreddie/darknet/master/data/coco.names -o cfg/coco.names

coco_names = []
with open("cfg/coco.names") as f:
    name = f.readline()
    while name:
        coco_names.append(name.strip())
        name = f.readline()

print(coco_names)

['person', 'bicycle', 'car', 'motorbike', 'aeroplane', 'bus', 'train', 'truck', 'boat', 'traffic light', 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'sofa', 'pottedplant', 'bed', 'diningtable', 'toilet', 'tvmonitor', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush']

def draw_box(img, dets):
    img2 = img.copy()
    width = img2.shape[1]
    height =  img2.shape[0]
    for i in range(dets.shape[0]):

        top_left = (int(dets[i,0] * width), int(dets[i,1] * height))
        bottom_right = (int(dets[i,2] * width), int(dets[i,3] * height))
        
        img2 = cv2.rectangle(img2, top_left , bottom_right, (255,0,0),3,1,0)

        img2 = cv2.putText(img2, coco_names[int(dets[i,4])], top_left, cv2.FONT_HERSHEY_COMPLEX_SMALL,1.5, (0 ,125,0), 2)
    
    result = cv2.cvtColor(img2, cv2.COLOR_BGR2RGB)
    cv2.imwrite("data/result.jpg", result)
    plt.imshow(img2)

draw_box(img, r)