def __init__(self, phase, size, Basenet, Neck, Head, cfg): super(SSD, self).__init__() self.phase = phase self.cfg = cfg self.priorbox = PriorBox(self.cfg) self.priors = self.priorbox.forward() self.size = size # SSD network self.basenet = Basenet self.neck = Neck self.head = Head self.num_classes = cfg['num_classes'] self.softmax = nn.Softmax(dim=-1) self.detect = Detect(self.num_classes, 0, 200, 0.01, 0.45, variance=cfg['variance'], nms_kind=cfg['nms_kind'], beta1=cfg['beta1'])
def __init__(self): super().__init__() self.backbone = darknet53 # Compute mask_dim here and add it back to the config. Make sure Yolact's constructor is called early! self.num_grids = 0 self.proto_src = 0 in_channels = 256 in_channels += self.num_grids mask_proto_net = [(256, 3, { 'padding': 1 }), (256, 3, { 'padding': 1 }), (256, 3, { 'padding': 1 }), (None, -2, {}), (256, 3, { 'padding': 1 }), (32, 1, {})] # The include_last_relu=false here is because we might want to change it to another function self.proto_net, mask_dim = make_net(in_channels, mask_proto_net, include_last_relu=False) self.selected_layers = [2, 3, 4] src_channels = self.backbone.channels # Some hacky rewiring to accomodate the FPN self.fpn = FPN([src_channels[i] for i in self.selected_layers]) self.selected_layers = list(range(len(self.selected_layers) + 2)) src_channels = [256] * len(self.selected_layers) self.prediction_layers = nn.ModuleList() pred_aspect_ratios = [[[1, 1 / 2, 2]]] * 5 pred_scales = [[24], [48], [96], [192], [384]] for idx, layer_idx in enumerate(self.selected_layers): # If we're sharing prediction module weights, have every module's parent be the first one parent = None if idx > 0: parent = self.prediction_layers[0] pred = PredictionModule(src_channels[layer_idx], src_channels[layer_idx], aspect_ratios=pred_aspect_ratios[idx], scales=pred_scales[idx], parent=parent) self.prediction_layers.append(pred) self.semantic_seg_conv = nn.Conv2d(src_channels[0], 80, kernel_size=1) self.detect = Detect(81, bkg_label=0, top_k=200, conf_thresh=0.05, nms_thresh=0.5)
def __init__(self, phase, size, base, extras, head, num_classes): super(SSD, self).__init__() self.phase = phase # 'train'或'test' self.size = size # 输入尺寸=300 self.num_classes = num_classes # 类别数 self.cfg = VOC # 配置信息 self.priorbox = PriorBox(self.cfg) # 产生先验框 with torch.no_grad(): # 不使用梯度 self.priors = self.priorbox.forward() self.vgg = nn.ModuleList(base) # 根据'base'建立nn.ModuleList对象 self.L2Norm = L2Norm(512, 20) # conv4_3后使用L2正则化 self.extras = nn.ModuleList(extras) # VGG-16后额外添加的四层 self.loc = nn.ModuleList(head[0]) # 位置预测 self.conf = nn.ModuleList(head[1]) # 置信度预测 if phase == 'test': # 测试阶段需要不同处理 self.softmax = nn.Softmax(dim=-1) self.detect = Detect(num_classes, 200, 0.01, 0.45)
class SSD(nn.Module): """Single Shot Multibox Architecture The network is composed of a base VGG network followed by the added multibox conv layers. Each multibox layer branches into 1) conv2d for class conf scores 2) conv2d for localization predictions 3) associated priorbox layer to produce default bounding boxes specific to the layer's feature map size. See: https://arxiv.org/pdf/1512.02325.pdf for more details. Args: phase: (string) Can be "test" or "train" size: input image size base: VGG16 layers for input, size of either 300 or 500 extras: extra layers that feed to multibox loc and conf layers head: "multibox head" consists of loc and conf conv layers """ def __init__(self, phase, size, Basenet, Neck, Head, cfg): super(SSD, self).__init__() self.phase = phase self.cfg = cfg self.priorbox = PriorBox(self.cfg) self.priors = self.priorbox.forward() self.size = size # SSD network self.basenet = Basenet self.neck = Neck self.head = Head self.num_classes = cfg['num_classes'] self.softmax = nn.Softmax(dim=-1) self.detect = Detect(self.num_classes, 0, 200, 0.01, 0.45, variance=cfg['variance'], nms_kind=cfg['nms_kind'], beta1=cfg['beta1']) def forward(self, x, phase): """Applies network layers and ops on input image(s) x. Args: x: input image or batch of images. Shape: [batch,3,300,300]. Return: Depending on phase: test: Variable(tensor) of output class label predictions, confidence score, and corresponding location predictions for each object detected. Shape: [batch,topk,7] train: list of concat outputs from: 1: confidence layers, Shape: [batch*num_priors,num_classes] 2: localization layers, Shape: [batch,num_priors*4] 3: priorbox layers, Shape: [2,num_priors*4] """ x = self.basenet(x) if self.neck is not None: x = self.neck(x) conf, loc = self.head(x) loc = torch.cat([o.view(o.size(0), -1) for o in loc], 1) conf = torch.cat([o.view(o.size(0), -1) for o in conf], 1) if phase == "test": output = self.detect.trans( loc.view(loc.size(0), -1, 4), # loc preds self.softmax(conf.view(conf.size(0), -1, self.num_classes)), # conf preds #self.priors.type(type(x.data)) # default boxes self.priors ) else: output = ( loc.view(loc.size(0), -1, 4), conf.view(conf.size(0), -1, self.num_classes), self.priors ) return output def load_weights(self, base_file): other, ext = os.path.splitext(base_file) if ext == '.pkl' or '.pth': print('Loading weights into state dict...') self.load_state_dict(torch.load(base_file, map_location=lambda storage, loc: storage)) print('Finished!') else: print('Sorry only .pth and .pkl files supported.')
print('Epoch {}, iter {}, lr {:.6f}, loss {:.2f}, time {:.2f}s, eta {:.2f}h'.format( epoch, iteration, optimizer.param_groups[0]['lr'], loss.item(), load_time, load_time * (max_iter - iteration) / 3600, )) timer.clear() save_weights(model) print('Start Evaluation...') thresh=0.005 max_per_image=300 model.eval() detector = Detect(num_classes) transform = BaseTransform(args.size) num_images = len(testset) all_boxes = [[[] for _ in range(num_images)] for _ in range(num_classes)] rgbs = dict() os.makedirs("draw/", exist_ok=True) os.makedirs("draw/{}/".format(args.dataset), exist_ok=True) _t = {'im_detect': Timer(), 'im_nms': Timer()} for i in range(num_images): img = testset.pull_image(i) scale = torch.Tensor([img.shape[1], img.shape[0], img.shape[1], img.shape[0]]) with torch.no_grad(): x = transform(img).unsqueeze(0) (x, scale) = (x.cuda(), scale.cuda()) _t['im_detect'].tic()
def eval_model( model, num_classes, testset, priors, thresh=0.005, max_per_image=300, ): # Testing after training print('Start Evaluation...') model.eval() detector = Detect(num_classes) transform = BaseTransform(args.size, (104, 117, 123), (2, 0, 1)) num_images = len(testset) all_boxes = [[[] for _ in range(num_images)] for _ in range(num_classes)] rgbs = dict() _t = {'im_detect': Timer(), 'im_nms': Timer()} for i in range(num_images): img = testset.pull_image(i) scale = torch.Tensor( [img.shape[1], img.shape[0], img.shape[1], img.shape[0]]) with torch.no_grad(): x = transform(img).unsqueeze(0) (x, scale) = (x.cuda(), scale.cuda()) _t['im_detect'].tic() out = model(x) # forward pass (boxes, scores) = detector.forward(out, priors) detect_time = _t['im_detect'].toc() boxes *= scale # scale each detection back up to the image boxes = boxes.cpu().numpy() scores = scores.cpu().numpy() _t['im_nms'].tic() for j in range(1, num_classes): inds = np.where(scores[:, j - 1] > thresh)[0] if len(inds) == 0: all_boxes[j][i] = np.empty([0, 5], dtype=np.float32) continue c_bboxes = boxes[inds] c_scores = scores[inds, j - 1] c_dets = np.hstack( (c_bboxes, c_scores[:, np.newaxis])).astype(np.float32, copy=False) keep = nms(c_dets, thresh=args.nms_thresh) # non maximum suppression c_dets = c_dets[keep, :] all_boxes[j][i] = c_dets if max_per_image > 0: image_scores = np.hstack( [all_boxes[j][i][:, -1] for j in range(1, num_classes)]) if len(image_scores) > max_per_image: image_thresh = np.sort(image_scores)[-max_per_image] for j in range(1, num_classes): keep = np.where(all_boxes[j][i][:, -1] >= image_thresh)[0] all_boxes[j][i] = all_boxes[j][i][keep, :] nms_time = _t['im_nms'].toc() if i == 10: _t['im_detect'].clear() _t['im_nms'].clear() if i % math.floor(num_images / 10) == 0 and i > 0: print('[{}/{}]Time results: detect={:.2f}ms,nms={:.2f}ms,'.format( i, num_images, detect_time * 1000, nms_time * 1000)) testset.evaluate_detections(all_boxes, 'eval/{}/'.format(args.dataset)) model.train()