深度学习实战之布匹缺陷检测

 Cython numpy==1.17 opencv-python torch>=1.4 matplotlib pillow tensorboard PyYAML>=5.3torchvisionscipytqdmgit+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI

 for fold in [0]:         val_index = index[len(index) * fold // 5:len(index) * (fold + 1) // 5]         print(len(val_index))         for num, name in enumerate(name_list):             print(c_list[num], x_center_list[num], y_center_list[num], w_list[num], h_list[num])             row = [c_list[num], x_center_list[num], y_center_list[num], w_list[num], h_list[num]]             if name in val_index:                 path2save = 'val/'             else:                path2save = 'train/'                      if not os.path.exists('convertor/fold{}/labels/'.format(fold) + path2save):                os.makedirs('convertor/fold{}/labels/'.format(fold) + path2save)            with open('convertor/fold{}/labels/'.format(fold) + path2save + name.split('.')[0] + ".txt", 'a+') as f:                for data in row:                    f.write('{} '.format(data))                f.write('\n')                if not os.path.exists('convertor/fold{}/images/{}'.format(fold, path2save)):                    os.makedirs('convertor/fold{}/images/{}'.format(fold, path2save))                sh.copy(os.path.join(image_path, name.split('.')[0], name),                        'convertor/fold{}/images/{}/{}'.format(fold, path2save, name))

 # Hyperparameters hyp = {'lr0': 0.01,  # initial learning rate (SGD=1E-2, Adam=1E-3)        'momentum': 0.937,  # SGD momentum        'weight_decay': 5e-4,  # optimizer weight decay        'giou': 0.05,  # giou loss gain        'cls': 0.58,  # cls loss gain        'cls_pw': 1.0,  # cls BCELoss positive_weight        'obj': 1.0,  # obj loss gain (*=img_size/320 if img_size != 320)        'obj_pw': 1.0,  # obj BCELoss positive_weight       'iou_t': 0.20,  # iou training threshold       'anchor_t': 4.0,  # anchor-multiple threshold       'fl_gamma': 0.0,  # focal loss gamma (efficientDet default is gamma=1.5)       'hsv_h': 0.014,  # image HSV-Hue augmentation (fraction)       'hsv_s': 0.68,  # image HSV-Saturation augmentation (fraction)       'hsv_v': 0.36,  # image HSV-Value augmentation (fraction)       'degrees': 0.0,  # image rotation (+/- deg)       'translate': 0.0,  # image translation (+/- fraction)       'scale': 0.5,  # image scale (+/- gain)       'shear': 0.0}  # image shear (+/- deg)}

import argparse    from models.experimental import *      class Detect(nn.Module):      def __init__(self, nc=80, anchors=()):  # detection layer          super(Detect, self).__init__()          self.stride = None  # strides computed during build         self.nc = nc  # number of classes         self.no = nc + 5  # number of outputs per anchor         self.nl = len(anchors)  # number of detection layers         self.na = len(anchors[0]) // 2  # number of anchors         self.grid = [torch.zeros(1)] * self.nl  # init grid         a = torch.tensor(anchors).float().view(self.nl, -1, 2)         self.register_buffer('anchors', a)  # shape(nl,na,2)         self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2))  # shape(nl,1,na,1,1,2)         self.export = False  # onnx export      def forward(self, x):         # x = x.copy()  # for profiling         z = []  # inference output         self.training |= self.export         for i in range(self.nl):             bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)             x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()              if not self.training:  # inference                 if self.grid[i].shape[2:4] != x[i].shape[2:4]:                     self.grid[i] = self._make_grid(nx, ny).to(x[i].device)                  y = x[i].sigmoid()                 y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + self.grid[i].to(x[i].device)) * self.stride[i]  # xy                 y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh                 z.append(y.view(bs, -1, self.no))          return x if self.training else (torch.cat(z, 1), x)      @staticmethod     def _make_grid(nx=20, ny=20):         yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])         return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()   class Model(nn.Module):     def __init__(self, model_cfg='yolov5s.yaml', ch=3, nc=None):  # model, input channels, number of classes         super(Model, self).__init__()         if type(model_cfg) is dict:             self.md = model_cfg  # model dict         else:  # is *.yaml             import yaml  # for torch hub             with open(model_cfg) as f:                 self.md = yaml.load(f, Loader=yaml.FullLoader)  # model dict          # Define model         if nc and nc != self.md['nc']:             print('Overriding %s nc=%g with nc=%g' % (model_cfg, self.md['nc'], nc))             self.md['nc'] = nc  # override yaml value         self.model, self.save = parse_model(self.md, ch=[ch])  # model, savelist, ch_out         # print([x.shape for x in self.forward(torch.zeros(1, ch, 64, 64))])          # Build strides, anchors         m = self.model[-1]  # Detect()         if isinstance(m, Detect):             s = 128  # 2x min stride             m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))])  # forward             m.anchors /= m.stride.view(-1, 1, 1)             check_anchor_order(m)             self.stride = m.stride             self._initialize_biases()  # only run once             # print('Strides: %s' % m.stride.tolist())          # Init weights, biases         torch_utils.initialize_weights(self)         self._initialize_biases()  # only run once         torch_utils.model_info(self)         print('')      def forward(self, x, augment=False, profile=False):         if augment:             img_size = x.shape[-2:]  # height, width             s = [0.83, 0.67]  # scales #1.2 0.83             y = []             for i, xi in enumerate((x,                                     torch_utils.scale_img(x.flip(3), s[0]),  # flip-lr and scale                                     torch_utils.scale_img(x, s[1]),  # scale                                     )):                 # cv2.imwrite('img%g.jpg' % i, 255 * xi[0].numpy().transpose((1, 2, 0))[:, :, ::-1])                 y.append(self.forward_once(xi)[0])              y[1][..., :4] /= s[0]  # scale             y[1][..., 0] = img_size[1] - y[1][..., 0]  # flip lr             y[2][..., :4] /= s[1]  # scale             return torch.cat(y, 1), None  # augmented inference, train         else:             return self.forward_once(x, profile)  # single-scale inference, train      def forward_once(self, x, profile=False):         y, dt = [], []  # outputs        for m in self.model:            if m.f != -1:  # if not from previous layer                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
            if profile:                try:                    import thop                    o = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2  # FLOPS                except:                    o = 0                t = torch_utils.time_synchronized()                for _ in range(10):                    _ = m(x)                dt.append((torch_utils.time_synchronized() - t) * 100)                print('%10.1f%10.0f%10.1fms %-40s' % (o, m.np, dt[-1], m.type))            x = m(x)  # run            y.append(x if m.i in self.save else None)  # save output
        if profile:            print('%.1fms total' % sum(dt))        return x
    def _initialize_biases(self, cf=None):  # initialize biases into Detect(), cf is class frequency        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.        m = self.model[-1]  # Detect() module        for f, s in zip(m.f, m.stride):  #  from            mi = self.model[f % m.i]            b = mi.bias.view(m.na, -1)  # conv.bias(255) to (3,85)            b[:, 4] += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)            b[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum())  # cls            mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
    def _print_biases(self):        m = self.model[-1]  # Detect() module        for f in sorted([x % m.i for x in m.f]):  #  from            b = self.model[f].bias.detach().view(m.na, -1).T  # conv.bias(255) to (3,85)            print(('%g Conv2d.bias:' + '%10.3g' * 6) % (f, *b[:5].mean(1).tolist(), b[5:].mean()))
    # def _print_weights(self):    #     for m in self.model.modules():    #         if type(m) is Bottleneck:    #             print('%10.3g' % (m.w.detach().sigmoid() * 2))  # shortcut weights
    def fuse(self):  # fuse model Conv2d() + BatchNorm2d() layers        print('Fusing layers... ', end='')        for m in self.model.modules():            if type(m) is Conv:                m.conv = torch_utils.fuse_conv_and_bn(m.conv, m.bn)  # update conv                m.bn = None  # remove batchnorm                m.forward = m.fuseforward  # update forward        torch_utils.model_info(self)        return self
def parse_model(md, ch):  # model_dict, input_channels(3)    print('\n%3s%18s%3s%10s  %-40s%-30s' % ('', 'from', 'n', 'params', 'module', 'arguments'))    anchors, nc, gd, gw = md['anchors'], md['nc'], md['depth_multiple'], md['width_multiple']    na = (len(anchors[0]) // 2)  # number of anchors    no = na * (nc + 5)  # number of outputs = anchors * (classes + 5)
    layers, save, c2 = [], [], ch[-1]  # layers, savelist, ch out    for i, (f, n, m, args) in enumerate(md['backbone'] + md['head']):  # from, number, module, args        m = eval(m) if isinstance(m, str) else m  # eval strings        for j, a in enumerate(args):            try:                args[j] = eval(a) if isinstance(a, str) else a  # eval strings            except:                pass
        n = max(round(n * gd), 1) if n > 1 else n  # depth gain        if m in [nn.Conv2d, Conv, PW_Conv,Bottleneck, SPP, DWConv, MixConv2d, Focus, CrossConv, BottleneckCSP, C3, BottleneckMOB]:            c1, c2 = ch[f], args[0]
            # Normal            # if i > 0 and args[0] != no:  # channel expansion factor            #     ex = 1.75  # exponential (default 2.0)            #     e = math.log(c2 / ch[1]) / math.log(2)            #     c2 = int(ch[1] * ex ** e)            # if m != Focus:            c2 = make_divisible(c2 * gw, 8) if c2 != no else c2
            # Experimental            # if i > 0 and args[0] != no:  # channel expansion factor            #     ex = 1 + gw  # exponential (default 2.0)            #     ch1 = 32  # ch[1]            #     e = math.log(c2 / ch1) / math.log(2)  # level 1-n            #     c2 = int(ch1 * ex ** e)            # if m != Focus:            #     c2 = make_divisible(c2, 8) if c2 != no else c2
            args = [c1, c2, *args[1:]]            if m in [BottleneckCSP, C3]:                args.insert(2, n)                n = 1        elif m is nn.BatchNorm2d:            args = [ch[f]]        elif m is Concat:            c2 = sum([ch[-1 if x == -1 else x + 1] for x in f])        elif m is Detect:            f = f or list(reversed([(-1 if j == i else j - 1) for j, x in enumerate(ch) if x == no]))        else:            c2 = ch[f]
        m_ = nn.Sequential(*[m(*args) for _ in range(n)]) if n > 1 else m(*args)  # module        t = str(m)[8:-2].replace('__main__.', '')  # module type        np = sum([x.numel() for x in m_.parameters()])  # number params        m_.i, m_.f, m_.type, m_.np = i, f, t, np  # attach index, 'from' index, type, number params        print('%3s%18s%3s%10.0f  %-40s%-30s' % (i, f, n, np, t, args))  # print        save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1)  # append to savelist        layers.append(m_)        ch.append(c2)    return nn.Sequential(*layers), sorted(save)

if __name__ == '__main__':    parser = argparse.ArgumentParser()    parser.add_argument('--cfg', type=str, default='yolov5s.yaml', help='model.yaml')    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')    opt = parser.parse_args()    opt.cfg = check_file(opt.cfg)  # check file    device = torch_utils.select_device(opt.device)
    # Create model    model = Model(opt.cfg).to(device)    model.train()

for *xyxy, conf, cls in det:                     if save_txt:  # Write to file                         xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) /                                 gn).view(-1).tolist()  # normalized xywh                         with open(txt_path + '.txt', 'a') as f:                             f.write(('%g ' * 5 + '\n') %                                     (cls, *xywh))  # label format                      # write to json                    if save_json:                        name = os.path.split(txt_path)[-1]                        print(name)
                        x1, y1, x2, y2 = float(xyxy[0]), float(xyxy[1]), float(                            xyxy[2]), float(xyxy[3])                        bbox = [x1, y1, x2, y2]                        img_name = name                        conf = float(conf)
                        #add solution remove other                        result.append({                            'name': img_name + '.jpg',                            'category': int(cls + 1),                            'bbox': bbox,                            'score': conf                        })