Pytorch版本yolov3源码阅读

Pytorch版本yolov3源码阅读

目录

• Pytorch版本yolov3源码阅读
  • 1. 阅读test.py
    • 1.1 参数解读
    • 1.2 data文件解析
    • 1.3 cfg文件解析
    • 1.4 根据cfg文件创建模块
    • 1.5 YOLOLayer
    • 1.6 初始化模型
    • 1.7 加载权重
    • 1.8 计算mAP
  • 2. 阅读train.py
    • 2.1 参数解读
    • 2.2 随机初始化
    • 2.3 设置优化器
    • 2.4 更新优化器
    • 2.5 loss指标
    • 2.6 checkpoint相关
  • 3. 阅读detect.py
    • 3.1 参数解读
    • 3.2 预测框的获取
    • 3.2 核心-迭代图片画出预测框

1. 阅读test.py

1.1 参数解读

parser = argparse.ArgumentParser()
parser.add_argument('-batch_size', type=int, default=32, help='size of each image batch')
parser.add_argument('-cfg', type=str, default='cfg/yolov3.cfg', help='path to model config file')
parser.add_argument('-data_config_path', type=str, default='cfg/coco.data', help='path to data config file')
parser.add_argument('-weights_path', type=str, default='checkpoints/yolov3.pt', help='path to weights file')
parser.add_argument('-class_path', type=str, default='data/coco.names', help='path to class label file')
parser.add_argument('-iou_thres', type=float, default=0.5, help='iou threshold required to qualify as detected')
parser.add_argument('-conf_thres', type=float, default=0.5, help='object confidence threshold')
parser.add_argument('-nms_thres', type=float, default=0.45, help='iou threshold for non-maximum suppression')
parser.add_argument('-n_cpu', type=int, default=0, help='number of cpu threads to use during batch generation')
parser.add_argument('-img_size', type=int, default=608, help='size of each image dimension')
opt = parser.parse_args()
print(opt)

• batch_size: 每个batch大小，跟darknet不太一样，没有subdivision
• cfg： 网络配置文件
• data_config_path: coco.data文件，存储相关信息
• weights_path: 权重文件路径
• class_path: 类别文件，注意类别的顺序，coco.names
• iou_thres: iou阈值
• conf_thres: 目标执行度阈值
• nms_thres: 非极大抑制阈值
• n_cpu: 实用多少个线程来创建batch
• img_size: 设置初始图片大小

1.2 data文件解析

def parse_data_config(path):
    """Parses the data configuration file"""
    options = dict()
    options['gpus'] = '0,1'
    options['num_workers'] = '10'
    with open(path, 'r') as fp:
        lines = fp.readlines()
    for line in lines:
        line = line.strip()
        if line == '' or line.startswith('#'):
            continue
        key, value = line.split('=')
        options[key.strip()] = value.strip()
    return options

将data文件中内容存储到options这个dict中，获取的时候就可以对这个对象通过key进行提取value。

1.3 cfg文件解析

def parse_model_config(path):
    """Parses the yolo-v3 layer configuration file and returns module definitions"""
    file = open(path, 'r')
    lines = file.read().split('\n')
    lines = [x for x in lines if x and not x.startswith('#')]
    lines = [x.rstrip().lstrip() for x in lines] # get rid of fringe whitespaces
    module_defs = []
    for line in lines:
        if line.startswith('['): # This marks the start of a new block
            module_defs.append({})
            module_defs[-1]['type'] = line[1:-1].rstrip()
            if module_defs[-1]['type'] == 'convolutional':
                module_defs[-1]['batch_normalize'] = 0
        else:
            key, value = line.split("=")
            value = value.strip()
            module_defs[-1][key.rstrip()] = value.strip()

    return module_defs

返回的module_defs存储的是所有的网络参数信息，一个list中套了很多个dict.

1.4 根据cfg文件创建模块

def create_modules(module_defs):
    """
    Constructs module list of layer blocks from module configuration in module_defs
    """
    #将第一层内容，也就是网络超参数设定
    hyperparams = module_defs.pop(0)
    
    output_filters = [int(hyperparams['channels'])]

    module_list = nn.ModuleList()
    for i, module_def in enumerate(module_defs):
        #一个时序容器。`Modules` 会以他们传入的顺序被添加到容器中。当然，也可以传入一个`OrderedDict`
        modules = nn.Sequential()
        #根据不同的层进行不同的设计
        if module_def['type'] == 'convolutional':
            bn = int(module_def['batch_normalize'])
            filters = int(module_def['filters'])
            kernel_size = int(module_def['size'])
            pad = (kernel_size - 1) // 2 if int(module_def['pad']) else 0
            #将一个 `child module` 添加到当前 `modle`。 被添加的`module`可以通过 `name`属性来获取。
            modules.add_module('conv_%d' % i, nn.Conv2d(in_channels=output_filters[-1],
                                                        out_channels=filters,
                                                        kernel_size=kernel_size,
                                                        stride=int(module_def['stride']),
                                                        padding=pad,
                                                        bias=not bn))
            if bn:
                modules.add_module('batch_norm_%d' % i, nn.BatchNorm2d(filters))
            if module_def['activation'] == 'leaky':
                modules.add_module('leaky_%d' % i, nn.LeakyReLU(0.1))

        elif module_def['type'] == 'upsample':
            # pytorch中的上采样函数
            upsample = nn.Upsample(scale_factor=int(module_def['stride']), mode='nearest')
            modules.add_module('upsample_%d' % i, upsample)

        elif module_def['type'] == 'route':
            # 对yolo cfg文件中的route层进行解析
            # eg: route -1, 14
            layers = [int(x) for x in module_def['layers'].split(',')]
            # 将多个层进行以sum的形式合并
            # 这个地方发现与darknet中不同，darknet中是以concate的方式进行的
            filters = sum([output_filters[layer_i] for layer_i in layers])
            modules.add_module('route_%d' % i, EmptyLayer())

        elif module_def['type'] == 'shortcut':
            # eg from yolov3.cfg
            # from=-3
            # activation = linear
            # 未定义activation方式？？？
            filters = output_filters[int(module_def['from'])]
            modules.add_module('shortcut_%d' % i, EmptyLayer())

        elif module_def['type'] == 'yolo':
            anchor_idxs = [int(x) for x in module_def['mask'].split(',')]
            # Extract anchors
            anchors = [float(x) for x in module_def['anchors'].split(',')]
            anchors = [(anchors[i], anchors[i + 1]) for i in range(0, len(anchors), 2)]
            anchors = [anchors[i] for i in anchor_idxs]
            num_classes = int(module_def['classes'])
            img_height = int(hyperparams['height'])
            # Define detection layer
            yolo_layer = YOLOLayer(anchors, num_classes, img_height, anchor_idxs)
            modules.add_module('yolo_%d' % i, yolo_layer)

        # Register module list and number of output filters
        # 将module添加到module_list中进行保存
        module_list.append(modules)
        output_filters.append(filters)

    return hyperparams, module_list

这里开始就涉及到pytorch部分的内容了：

• module_list = nn.ModuleList(): 创建一个list，其中存放的是module
• nn.Sequential(): 一个时序容器。Modules 会以他们传入的顺序被添加到容器中。当然，也可以传入一个OrderedDict。
• add_module(name,module):将一个 child module 添加到当前 modle。 被添加的module可以通过 name属性来获取。

1.5 YOLOLayer

class YOLOLayer(nn.Module):

    def __init__(self, anchors, nC, img_dim, anchor_idxs):
        super(YOLOLayer, self).__init__()

        anchors = [(a_w, a_h) for a_w, a_h in anchors]  # (pixels)
        nA = len(anchors)

        self.anchors = anchors
        self.nA = nA  # number of anchors (3)
        self.nC = nC  # number of classes (80)
        self.bbox_attrs = 5 + nC
        self.img_dim = img_dim  # from hyperparams in cfg file, NOT from parser

        if anchor_idxs[0] == (nA * 2):  # 6
            stride = 32
        elif anchor_idxs[0] == nA:  # 3
            stride = 16
        else:
            stride = 8

        # Build anchor grids
        nG = int(self.img_dim / stride)
        self.grid_x = torch.arange(nG).repeat(nG, 1).view([1, 1, nG, nG]).float()
        self.grid_y = torch.arange(nG).repeat(nG, 1).t().view([1, 1, nG, nG]).float()
        self.scaled_anchors = torch.FloatTensor([(a_w / stride, a_h / stride) for a_w, a_h in anchors])
        self.anchor_w = self.scaled_anchors[:, 0:1].view((1, nA, 1, 1))
        self.anchor_h = self.scaled_anchors[:, 1:2].view((1, nA, 1, 1))

    def forward(self, p, targets=None, requestPrecision=False):
        FT = torch.cuda.FloatTensor if p.is_cuda else torch.FloatTensor

        bs = p.shape[0]  # batch size
        nG = p.shape[2]  # number of grid points
        stride = self.img_dim / nG

        if p.is_cuda and not self.grid_x.is_cuda:
            self.grid_x, self.grid_y = self.grid_x.cuda(), self.grid_y.cuda()
            self.anchor_w, self.anchor_h = self.anchor_w.cuda(), self.anchor_h.cuda()

        # p.view(12, 255, 13, 13) -- > (12, 3, 13, 13, 80)  # (bs, anchors, grid, grid, classes + xywh)
        p = p.view(bs, self.nA, self.bbox_attrs, nG, nG).permute(0, 1, 3, 4, 2).contiguous()  # prediction

        # Get outputs
        x = torch.sigmoid(p[..., 0])  # Center x
        y = torch.sigmoid(p[..., 1])  # Center y

        # Width and height (yolo method)
        w = p[..., 2]  # Width
        h = p[..., 3]  # Height
        width = torch.exp(w.data) * self.anchor_w
        height = torch.exp(h.data) * self.anchor_h

        # Width and height (power method)
        # w = torch.sigmoid(p[..., 2])  # Width
        # h = torch.sigmoid(p[..., 3])  # Height
        # width = ((w.data * 2) ** 2) * self.anchor_w
        # height = ((h.data * 2) ** 2) * self.anchor_h

        # Add offset and scale with anchors (in grid space, i.e. 0-13)
        pred_boxes = FT(bs, self.nA, nG, nG, 4)
        pred_conf = p[..., 4]  # Conf
        pred_cls = p[..., 5:]  # Class

        # Training
        if targets is not None:
            MSELoss = nn.MSELoss(size_average=True)
            BCEWithLogitsLoss = nn.BCEWithLogitsLoss(size_average=True)
            CrossEntropyLoss = nn.CrossEntropyLoss()

            if requestPrecision:
                gx = self.grid_x[:, :, :nG, :nG]
                gy = self.grid_y[:, :, :nG, :nG]
                pred_boxes[..., 0] = x.data + gx - width / 2
                pred_boxes[..., 1] = y.data + gy - height / 2
                pred_boxes[..., 2] = x.data + gx + width / 2
                pred_boxes[..., 3] = y.data + gy + height / 2

            tx, ty, tw, th, mask, tcls, TP, FP, FN, TC = \
                build_targets(pred_boxes, pred_conf, pred_cls, targets, self.scaled_anchors, self.nA, self.nC, nG,
                              requestPrecision)
            tcls = tcls[mask]
            if x.is_cuda:
                tx, ty, tw, th, mask, tcls = tx.cuda(), ty.cuda(), tw.cuda(), th.cuda(), mask.cuda(), tcls.cuda()

            # Mask outputs to ignore non-existing objects (but keep confidence predictions)
            nT = sum([len(x) for x in targets])  # number of targets
            nM = mask.sum().float()  # number of anchors (assigned to targets)
            nB = len(targets)  # batch size
            k = nM / nB
            if nM > 0:
                lx = k * MSELoss(x[mask], tx[mask])
                ly = k * MSELoss(y[mask], ty[mask])
                lw = k * MSELoss(w[mask], tw[mask])
                lh = k * MSELoss(h[mask], th[mask])

                # lconf = k * BCEWithLogitsLoss(pred_conf[mask], mask[mask].float())
                lconf = k * BCEWithLogitsLoss(pred_conf, mask.float())

                lcls = k * CrossEntropyLoss(pred_cls[mask], torch.argmax(tcls, 1))
                # lcls = k * BCEWithLogitsLoss(pred_cls[mask], tcls.float())
            else:
                lx, ly, lw, lh, lcls, lconf = FT([0]), FT([0]), FT([0]), FT([0]), FT([0]), FT([0])

            # Add confidence loss for background anchors (noobj)
            #lconf += k * BCEWithLogitsLoss(pred_conf[~mask], mask[~mask].float())

            # Sum loss components
            loss = lx + ly + lw + lh + lconf + lcls

            # Sum False Positives from unassigned anchors
            i = torch.sigmoid(pred_conf[~mask]) > 0.9
            if i.sum() > 0:
                FP_classes = torch.argmax(pred_cls[~mask][i], 1)
                FPe = torch.bincount(FP_classes, minlength=self.nC).float().cpu()  # extra FPs
            else:
                FPe = torch.zeros(self.nC)

            return loss, loss.item(), lx.item(), ly.item(), lw.item(), lh.item(), lconf.item(), lcls.item(), \
                   nT, TP, FP, FPe, FN, TC

        else:
            pred_boxes[..., 0] = x.data + self.grid_x
            pred_boxes[..., 1] = y.data + self.grid_y
            pred_boxes[..., 2] = width
            pred_boxes[..., 3] = height

            # If not in training phase return predictions
            output = torch.cat((pred_boxes.view(bs, -1, 4) * stride,
                                torch.sigmoid(pred_conf.view(bs, -1, 1)), pred_cls.view(bs, -1, self.nC)), -1)
            return output.data

暂且放到这里，之后在做解析

1.6 初始化模型

model = Darknet(opt.cfg, opt.img_size)

转到定义：

class Darknet(nn.Module):
    """YOLOv3 object detection model"""

    def __init__(self, config_path, img_size=416):
        super(Darknet, self).__init__()
        self.module_defs = parse_model_config(config_path)
        self.module_defs[0]['height'] = img_size
        self.hyperparams, self.module_list = create_modules(self.module_defs)
        self.img_size = img_size
        self.loss_names = ['loss', 'x', 'y', 'w', 'h', 'conf', 'cls', 'nT', 'TP', 'FP', 'FPe', 'FN', 'TC']

    def forward(self, x, targets=None, requestPrecision=False):
        is_training = targets is not None
        output = []
        self.losses = defaultdict(float)
        layer_outputs = []

        for i, (module_def, module) in enumerate(zip(self.module_defs, self.module_list)):
            if module_def['type'] in ['convolutional', 'upsample']:
                x = module(x)
            elif module_def['type'] == 'route':
                layer_i = [int(x) for x in module_def['layers'].split(',')]
                x = torch.cat([layer_outputs[i] for i in layer_i], 1)
            elif module_def['type'] == 'shortcut':
                layer_i = int(module_def['from'])
                x = layer_outputs[-1] + layer_outputs[layer_i]
            elif module_def['type'] == 'yolo':
                # Train phase: get loss
                if is_training:
                    x, *losses = module[0](x, targets, requestPrecision)
                    for name, loss in zip(self.loss_names, losses):
                        self.losses[name] += loss
                # Test phase: Get detections
                else:
                    x = module(x)
                output.append(x)
            layer_outputs.append(x)

        if is_training:
            self.losses['nT'] /= 3
            self.losses['TC'] /= 3
            metrics = torch.zeros(4, len(self.losses['FPe']))  # TP, FP, FN, target_count

            ui = np.unique(self.losses['TC'])[1:]
            for i in ui:
                j = self.losses['TC'] == float(i)
                metrics[0, i] = (self.losses['TP'][j] > 0).sum().float()  # TP
                metrics[1, i] = (self.losses['FP'][j] > 0).sum().float()  # FP
                metrics[2, i] = (self.losses['FN'][j] == 3).sum().float()  # FN
            metrics[3] = metrics.sum(0)
            metrics[1] += self.losses['FPe']

            self.losses['TP'] = metrics[0].sum()
            self.losses['FP'] = metrics[1].sum()
            self.losses['FN'] = metrics[2].sum()
            self.losses['TC'] = 0
            self.losses['metrics'] = metrics

        return sum(output) if is_training else torch.cat(output, 1)

梳理一下属性值，以便更好理解：

• module_def: dict类型，存储cfg文件中
• hyperparams: 超参数，整个网络需要的参数被存储到改属性中
• module_list：整个网络所有的模型加载到pytorch中的nn.ModuleList()
• loss_names: 有必要理解一下这里的loss中参数的含义
  • loss
  • x,y,w,h
  • conf
  • cls
  • nT
  • TP,FP,FPe,FN,TC

loss参数含义还不是很明白，留坑，待填坑

1.7 加载权重

都知道，pytorch版的yolov3权重文件是.pt结尾的，darknet版本的yolov3权重文件是.weights结尾的。

所以得知了这个版本可以使用加载weights文件。

# Load weights
if opt.weights_path.endswith('.weights'):  # darknet format
    load_weights(model, opt.weights_path)
elif opt.weights_path.endswith('.pt'):  # pytorch format
    checkpoint = torch.load(opt.weights_path, map_location='cpu')
    model.load_state_dict(checkpoint['model'])
    del checkpoint

1.8 计算mAP

print('Compute mAP...')

correct = 0
targets = None
outputs, mAPs, TP, confidence, pred_class, target_class = [], [], [], [], [], []
for batch_i, (imgs, targets) in enumerate(dataloader):
    imgs = imgs.to(device)

    with torch.no_grad():
        output = model(imgs)
        output = non_max_suppression(output, conf_thres=opt.conf_thres, nms_thres=opt.nms_thres)

    # Compute average precision for each sample
    for sample_i in range(len(targets)):
        correct = []

        # Get labels for sample where width is not zero (dummies)
        annotations = targets[sample_i]
        # Extract detections
        detections = output[sample_i]

        if detections is None:
            # If there are no detections but there are annotations mask as zero AP
            if annotations.size(0) != 0:
                mAPs.append(0)
            continue

        # Get detections sorted by decreasing confidence scores
        detections = detections[np.argsort(-detections[:, 4])]

        # If no annotations add number of detections as incorrect
        if annotations.size(0) == 0:
            target_cls = []
            #correct.extend([0 for _ in range(len(detections))])
            mAPs.append(0)
            continue
        else:
            target_cls = annotations[:, 0]

            # Extract target boxes as (x1, y1, x2, y2)
            target_boxes = xywh2xyxy(annotations[:, 1:5])
            target_boxes *= opt.img_size

            detected = []
            for *pred_bbox, conf, obj_conf, obj_pred in detections:

                pred_bbox = torch.FloatTensor(pred_bbox).view(1, -1)
                # Compute iou with target boxes
                iou = bbox_iou(pred_bbox, target_boxes)
                # Extract index of largest overlap
                best_i = np.argmax(iou)
                # If overlap exceeds threshold and classification is correct mark as correct
                if iou[best_i] > opt.iou_thres and obj_pred == annotations[best_i, 0] and best_i not in detected:
                    correct.append(1)
                    detected.append(best_i)
                else:
                    correct.append(0)

        # Compute Average Precision (AP) per class
        AP = ap_per_class(tp=correct, conf=detections[:, 4], pred_cls=detections[:, 6], target_cls=target_cls)

        # Compute mean AP for this image
        mAP = AP.mean()

        # Append image mAP to list
        mAPs.append(mAP)

        # Print image mAP and running mean mAP
        print('+ Sample [%d/%d] AP: %.4f (%.4f)' % (len(mAPs), len(dataloader) * opt.batch_size, mAP, np.mean(mAPs)))

print('Mean Average Precision: %.4f' % np.mean(mAPs))

留坑，待填

2. 阅读train.py

2.1 参数解读

parser = argparse.ArgumentParser()
parser.add_argument('-epochs', type=int, default=68, help='number of epochs')
parser.add_argument('-batch_size', type=int, default=12, help='size of each image batch')
parser.add_argument('-data_config_path', type=str, default='cfg/coco.data', help='data config file path')
parser.add_argument('-cfg', type=str, default='cfg/yolov3.cfg', help='cfg file path')
parser.add_argument('-img_size', type=int, default=32 * 13, help='size of each image dimension')
parser.add_argument('-resume', default=False, help='resume training flag')
opt = parser.parse_args()
print(opt)

• epochs 设置循环的参数
• batch_size: 设置batch
• data_config_path: data文件位置
• cfg: 记录cfg文件的位置
• img_size: 设置图片大小
• resume: 是否恢复训练（True or False）

2.2 随机初始化

random.seed(0)
np.random.seed(0)
torch.manual_seed(0)
if cuda:
    torch.cuda.manual_seed(0)
    torch.cuda.manual_seed_all(0)
    torch.backends.cudnn.benchmark = True

2.3 设置优化器

optimizer = torch.optim.SGD(filter(lambda p: p.requires_grad, model.parameters()), lr=1e-3,momentum=.9, weight_decay=5e-4, nesterov=True)

使用SGD优化器，learning_rate=0.001,momentum=0.9,weight_decay=5e-4,使用nesterov动量

2.4 更新优化器

根据当前epoch来确定使用哪一个lr:

        # Update scheduler (automatic)
        # scheduler.step()

        # Update scheduler (manual)
        if epoch < 54:
            lr = 1e-3
        elif epoch < 61:
            lr = 1e-4
        else:
            lr = 1e-5
        for g in optimizer.param_groups:
            g['lr'] = lr

可以自动更新参数，也可以手工更新参数。

2.5 loss指标

• mean_precision:

            # Precision
            precision = metrics[0] / (metrics[0] + metrics[1] + 1e-16)
            k = (metrics[0] + metrics[1]) > 0
            if k.sum() > 0:
                mean_precision = precision[k].mean()
            else:
                mean_precision = 0

• mean_recall:

            # Recall
            recall = metrics[0] / (metrics[0] + metrics[2] + 1e-16)
            k = (metrics[0] + metrics[2]) > 0
            if k.sum() > 0:
                mean_recall = recall[k].mean()
            else:
                mean_recall = 0

然后将所有指标写到results.txt文件中

2.6 checkpoint相关

checkpoint参数：epoch, best_loss,model,optimizer

latest.pt: 最新的权重文件

best.pt: 当前最好的权重文件

        # Save latest checkpoint
        checkpoint = {'epoch': epoch,
                      'best_loss': best_loss,
                      'model': model.state_dict(),
                      'optimizer': optimizer.state_dict()}
        torch.save(checkpoint, 'checkpoints/latest.pt')

        # Save best checkpoint
        if best_loss == loss_per_target:
            os.system('cp checkpoints/latest.pt checkpoints/best.pt')

        # Save backup checkpoint
        if (epoch > 0) & (epoch % 5 == 0):
            os.system('cp checkpoints/latest.pt checkpoints/backup' + str(epoch) + '.pt')

3. 阅读detect.py

3.1 参数解读

parser.add_argument('-image_folder', type=str, default='data/samples', help='path to images')
parser.add_argument('-output_folder', type=str, default='output', help='path to outputs')
parser.add_argument('-plot_flag', type=bool, default=True)
parser.add_argument('-txt_out', type=bool, default=False)
parser.add_argument('-cfg', type=str, default='cfg/yolov3.cfg', help='cfg file path')
parser.add_argument('-class_path', type=str, default='data/coco.names', help='path to class label file')
parser.add_argument('-conf_thres', type=float, default=0.50, help='object confidence threshold')
parser.add_argument('-nms_thres', type=float, default=0.45, help='iou threshold for non-maximum suppression')
parser.add_argument('-batch_size', type=int, default=1, help='size of the batches')
parser.add_argument('-img_size', type=int, default=32 * 13, help='size of each image dimension')
opt = parser.parse_args()
print(opt)

• image_folder: data/samples, 待检测的图片的文件夹
• output_folder: output,结果输出文件
• plot_flag: True or False, 添加bbox, 保存图片
• txt_out: True or False, 是否保存图片检测结果
• cfg: cfg文件路径
• class_path: 类别名称文件位置
• conf_thres, nms_thres: 目标检测置信度，非极大抑制阈值
• batch_size: 一般设置为1，选用默认的即可
• img_size: 设置加载图片时候的图片大小

3.2 预测框的获取

        # Get detections
        with torch.no_grad():
            chip = torch.from_numpy(img).unsqueeze(0).to(device)
            pred = model(chip)
            pred = pred[pred[:, :, 4] > opt.conf_thres]

            if len(pred) > 0:
                detections = non_max_suppression(pred.unsqueeze(0), opt.conf_thres, opt.nms_thres)
                img_detections.extend(detections)
                imgs.extend(img_paths)

获取预测框，非极大值抑制。

3.2 核心-迭代图片画出预测框

# Iterate through images and save plot of detections
    for img_i, (path, detections) in enumerate(zip(imgs, img_detections)):
        print("image %g: '%s'" % (img_i, path))

        if opt.plot_flag:
            img = cv2.imread(path)

        # The amount of padding that was added
        pad_x = max(img.shape[0] - img.shape[1], 0) * (opt.img_size / max(img.shape))
        pad_y = max(img.shape[1] - img.shape[0], 0) * (opt.img_size / max(img.shape))
        # Image height and width after padding is removed
        unpad_h = opt.img_size - pad_y
        unpad_w = opt.img_size - pad_x

        # Draw bounding boxes and labels of detections
        if detections is not None:
            unique_classes = detections[:, -1].cpu().unique()
            bbox_colors = random.sample(color_list, len(unique_classes))

            # write results to .txt file
            results_img_path = os.path.join(opt.output_folder, path.split('/')[-1])
            results_txt_path = results_img_path + '.txt'
            if os.path.isfile(results_txt_path):
                os.remove(results_txt_path)

            for i in unique_classes:
                n = (detections[:, -1].cpu() == i).sum()
                print('%g %ss' % (n, classes[int(i)]))

            for x1, y1, x2, y2, conf, cls_conf, cls_pred in detections:
                # Rescale coordinates to original dimensions
                box_h = ((y2 - y1) / unpad_h) * img.shape[0]
                box_w = ((x2 - x1) / unpad_w) * img.shape[1]
                y1 = (((y1 - pad_y // 2) / unpad_h) * img.shape[0]).round().item()
                x1 = (((x1 - pad_x // 2) / unpad_w) * img.shape[1]).round().item()
                x2 = (x1 + box_w).round().item()
                y2 = (y1 + box_h).round().item()
                x1, y1, x2, y2 = max(x1, 0), max(y1, 0), max(x2, 0), max(y2, 0)

                # write to file
                if opt.txt_out:
                    with open(results_txt_path, 'a') as file:
                        file.write(('%g %g %g %g %g %g \n') % (x1, y1, x2, y2, cls_pred, cls_conf * conf))

                if opt.plot_flag:
                    # Add the bbox to the plot
                    label = '%s %.2f' % (classes[int(cls_pred)], conf)
                    color = bbox_colors[int(np.where(unique_classes == int(cls_pred))[0])]
                    plot_one_box([x1, y1, x2, y2], img, label=label, color=color)

        if opt.plot_flag:
            # Save generated image with detections
            cv2.imwrite(results_img_path.replace('.bmp', '.jpg').replace('.tif', '.jpg'), img)