close NMS

models/detectors/rtdetr/build.py

@@ -24,7 +24,7 @@ def build_rtdetr(args, cfg, num_classes=80, trainable=False, deploy=False):
                     topk            = 300,
                     deploy          = deploy,
                     no_multi_labels = args.no_multi_labels,
-                    use_nms         = True,   # NMS is beneficial 
+                    use_nms         = False,   # NMS is beneficial 
                     nms_class_agnostic = args.nms_class_agnostic
                     )
             

models/detectors/vitdet/basic_modules/backbone.py

@@ -0,0 +1,129 @@
+import torch
+import torchvision
+from torch import nn
+from torchvision.models._utils import IntermediateLayerGetter
+
+try:
+    from .basic import FrozenBatchNorm2d
+except:
+    from basic  import FrozenBatchNorm2d
+   
+
+# IN1K MIM pretrained weights (from SparK: https://github.com/keyu-tian/SparK)
+pretrained_urls = {
+    # ResNet series
+    'resnet18':  None,
+    'resnet34':  None,
+    'resnet50':  "https://github.com/yjh0410/RT-ODLab/releases/download/backbone_weight/resnet50_in1k_spark_pretrained_timm_style.pth",
+    'resnet101': None,
+    # ShuffleNet series
+}
+
+
+# ----------------- Model functions -----------------
+## Build backbone network
+def build_backbone(cfg, pretrained=False):
+    print('==============================')
+    print('Backbone: {}'.format(cfg['backbone']))
+    # ResNet
+    if 'resnet' in cfg['backbone']:
+        model, feats = build_resnet(cfg, pretrained)
+    else:
+        raise NotImplementedError("Unknown backbone: <>.".format(cfg['backbone']))
+    
+    return model, feats
+
+
+# ----------------- ResNet Backbone -----------------
+class VisionTransformer(nn.Module):
+    """Vision Transformer."""
+    def __init__(self,
+                 name: str,
+                 norm_type: str,
+                 pretrained: bool = False,
+                 freeze_at: int = -1,
+                 freeze_stem_only: bool = False):
+        super().__init__()
+        # Pretrained
+        # Norm layer
+        if norm_type == 'BN':
+            norm_layer = nn.BatchNorm2d
+        elif norm_type == 'FrozeBN':
+            norm_layer = FrozenBatchNorm2d
+        # Backbone
+        backbone = getattr(torchvision.models, name)(norm_layer=norm_layer,)
+        return_layers = {"layer2": "0", "layer3": "1", "layer4": "2"}
+        self.body = IntermediateLayerGetter(backbone, return_layers=return_layers)
+        self.feat_dims = [128, 256, 512] if name in ('resnet18', 'resnet34') else [512, 1024, 2048]
+        
+        # Load pretrained
+        if pretrained:
+            self.load_pretrained(name)
+
+        # Freeze
+        if freeze_at >= 0:
+            for name, parameter in backbone.named_parameters():
+                if freeze_stem_only:
+                    if 'layer1' not in name and 'layer2' not in name and 'layer3' not in name and 'layer4' not in name:
+                        parameter.requires_grad_(False)
+                else:
+                    if 'layer2' not in name and 'layer3' not in name and 'layer4' not in name:
+                        parameter.requires_grad_(False)
+
+    def load_pretrained(self, name):
+        url = pretrained_urls[name]
+        if url is not None:
+            print('Loading pretrained weight from : {}'.format(url))
+            # checkpoint state dict
+            checkpoint_state_dict = torch.hub.load_state_dict_from_url(
+                url=url, map_location="cpu", check_hash=True)
+            # model state dict
+            model_state_dict = self.body.state_dict()
+            # check
+            for k in list(checkpoint_state_dict.keys()):
+                if k in model_state_dict:
+                    shape_model = tuple(model_state_dict[k].shape)
+                    shape_checkpoint = tuple(checkpoint_state_dict[k].shape)
+                    if shape_model != shape_checkpoint:
+                        checkpoint_state_dict.pop(k)
+                else:
+                    checkpoint_state_dict.pop(k)
+                    print('Unused key: ', k)
+            # load the weight
+            self.body.load_state_dict(checkpoint_state_dict)
+        else:
+            print('No backbone pretrained for {}.'.format(name))
+
+    def forward(self, x):
+        xs = self.body(x)
+        fmp_list = []
+        for name, fmp in xs.items():
+            fmp_list.append(fmp)
+
+        return fmp_list
+
+def build_resnet(cfg, pretrained=False):
+    # ResNet series
+    backbone = None
+
+    return backbone
+
+
+if __name__ == '__main__':
+    cfg = {
+        'backbone': 'resnet50',
+        'backbone_norm': 'FrozeBN',
+        'pretrained': True,
+        'freeze_at': 0,
+        'freeze_stem_only': False,
+    }
+    model, feat_dim = build_backbone(cfg, cfg['pretrained'])
+    model.eval()
+    print(feat_dim)
+
+    x = torch.ones(2, 3, 320, 320)
+    output = model(x)
+    for y in output:
+        print(y.size())
+    print(output[-1])
+

models/detectors/vitdet/basic_modules/basic.py

@@ -0,0 +1,238 @@
+import math
+import warnings
+import numpy as np
+import torch
+import torch.nn as nn
+
+
+def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
+    """Copy from timm"""
+    with torch.no_grad():
+        """Copy from timm"""
+        def norm_cdf(x):
+            return (1. + math.erf(x / math.sqrt(2.))) / 2.
+
+        if (mean < a - 2 * std) or (mean > b + 2 * std):
+            warnings.warn("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+                        "The distribution of values may be incorrect.",
+                        stacklevel=2)
+
+        l = norm_cdf((a - mean) / std)
+        u = norm_cdf((b - mean) / std)
+
+        tensor.uniform_(2 * l - 1, 2 * u - 1)
+        tensor.erfinv_()
+
+        tensor.mul_(std * math.sqrt(2.))
+        tensor.add_(mean)
+
+        tensor.clamp_(min=a, max=b)
+
+        return tensor
+
+
+# ---------------------------- NMS ----------------------------
+## basic NMS
+def nms(bboxes, scores, nms_thresh):
+    """"Pure Python NMS."""
+    x1 = bboxes[:, 0]  #xmin
+    y1 = bboxes[:, 1]  #ymin
+    x2 = bboxes[:, 2]  #xmax
+    y2 = bboxes[:, 3]  #ymax
+
+    areas = (x2 - x1) * (y2 - y1)
+    order = scores.argsort()[::-1]
+
+    keep = []
+    while order.size > 0:
+        i = order[0]
+        keep.append(i)
+        # compute iou
+        xx1 = np.maximum(x1[i], x1[order[1:]])
+        yy1 = np.maximum(y1[i], y1[order[1:]])
+        xx2 = np.minimum(x2[i], x2[order[1:]])
+        yy2 = np.minimum(y2[i], y2[order[1:]])
+
+        w = np.maximum(1e-10, xx2 - xx1)
+        h = np.maximum(1e-10, yy2 - yy1)
+        inter = w * h
+
+        iou = inter / (areas[i] + areas[order[1:]] - inter + 1e-14)
+        #reserve all the boundingbox whose ovr less than thresh
+        inds = np.where(iou <= nms_thresh)[0]
+        order = order[inds + 1]
+
+    return keep
+
+## class-agnostic NMS 
+def multiclass_nms_class_agnostic(scores, labels, bboxes, nms_thresh):
+    # nms
+    keep = nms(bboxes, scores, nms_thresh)
+    scores = scores[keep]
+    labels = labels[keep]
+    bboxes = bboxes[keep]
+
+    return scores, labels, bboxes
+
+## class-aware NMS 
+def multiclass_nms_class_aware(scores, labels, bboxes, nms_thresh, num_classes):
+    # nms
+    keep = np.zeros(len(bboxes), dtype=np.int32)
+    for i in range(num_classes):
+        inds = np.where(labels == i)[0]
+        if len(inds) == 0:
+            continue
+        c_bboxes = bboxes[inds]
+        c_scores = scores[inds]
+        c_keep = nms(c_bboxes, c_scores, nms_thresh)
+        keep[inds[c_keep]] = 1
+    keep = np.where(keep > 0)
+    scores = scores[keep]
+    labels = labels[keep]
+    bboxes = bboxes[keep]
+
+    return scores, labels, bboxes
+
+## multi-class NMS 
+def multiclass_nms(scores, labels, bboxes, nms_thresh, num_classes, class_agnostic=False):
+    if class_agnostic:
+        return multiclass_nms_class_agnostic(scores, labels, bboxes, nms_thresh)
+    else:
+        return multiclass_nms_class_aware(scores, labels, bboxes, nms_thresh, num_classes)
+
+
+# ----------------- Customed NormLayer Ops -----------------
+class LayerNorm2D(nn.Module):
+    def __init__(self, normalized_shape, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.ln = norm_layer(normalized_shape) if norm_layer is not None else nn.Identity()
+
+    def forward(self, x):
+        """
+        x: N C H W
+        """
+        x = x.permute(0, 2, 3, 1)
+        x = self.ln(x)
+        x = x.permute(0, 3, 1, 2)
+        return x
+
+
+# ----------------- Basic CNN Ops -----------------
+def get_conv2d(c1, c2, k, p, s, g, bias=False):
+    conv = nn.Conv2d(c1, c2, k, stride=s, padding=p, groups=g, bias=bias)
+
+    return conv
+
+def get_activation(act_type=None):
+    if act_type == 'relu':
+        return nn.ReLU(inplace=True)
+    elif act_type == 'lrelu':
+        return nn.LeakyReLU(0.1, inplace=True)
+    elif act_type == 'mish':
+        return nn.Mish(inplace=True)
+    elif act_type == 'silu':
+        return nn.SiLU(inplace=True)
+    elif act_type == 'gelu':
+        return nn.GELU()
+    elif act_type is None:
+        return nn.Identity()
+    else:
+        raise NotImplementedError
+        
+def get_norm(norm_type, dim):
+    if norm_type == 'BN':
+        return nn.BatchNorm2d(dim)
+    elif norm_type == 'GN':
+        return nn.GroupNorm(num_groups=32, num_channels=dim)
+    elif norm_type is None:
+        return nn.Identity()
+    else:
+        raise NotImplementedError
+
+class BasicConv(nn.Module):
+    def __init__(self, 
+                 in_dim,                   # in channels
+                 out_dim,                  # out channels 
+                 kernel_size=1,            # kernel size 
+                 padding=0,                # padding
+                 stride=1,                 # padding
+                 act_type  :str = 'lrelu', # activation
+                 norm_type :str = 'BN',    # normalization
+                ):
+        super(BasicConv, self).__init__()
+        add_bias = False if norm_type else True
+        self.conv = get_conv2d(in_dim, out_dim, k=kernel_size, p=padding, s=stride, g=1, bias=add_bias)
+        self.norm = get_norm(norm_type, out_dim)
+        self.act  = get_activation(act_type)
+
+    def forward(self, x):
+        return self.act(self.norm(self.conv(x)))
+
+class UpSampleWrapper(nn.Module):
+    """Upsample last feat map to specific stride."""
+    def __init__(self, in_dim, upsample_factor):
+        super(UpSampleWrapper, self).__init__()
+        # ---------- Basic parameters ----------
+        self.upsample_factor = upsample_factor
+
+        # ---------- Network parameters ----------
+        if upsample_factor == 1:
+            self.upsample = nn.Identity()
+        else:
+            scale = int(math.log2(upsample_factor))
+            dim = in_dim
+            layers = []
+            for _ in range(scale-1):
+                layers += [
+                    nn.ConvTranspose2d(dim, dim, kernel_size=2, stride=2),
+                    LayerNorm2D(dim),
+                    nn.GELU()
+                ]
+            layers += [nn.ConvTranspose2d(dim, dim, kernel_size=2, stride=2)]
+            self.upsample = nn.Sequential(*layers)
+            self.out_dim = dim
+
+    def forward(self, x):
+        x = self.upsample(x)
+
+        return x
+
+
+# ----------------- MLP modules -----------------
+class MLP(nn.Module):
+    def __init__(self, in_dim, hidden_dim, out_dim, num_layers):
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([in_dim] + h, h + [out_dim]))
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = nn.functional.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
+        return x
+
+class FFN(nn.Module):
+    def __init__(self, d_model=256, mlp_ratio=4.0, dropout=0., act_type='relu', pre_norm=False):
+        super().__init__()
+        # ----------- Basic parameters -----------
+        self.pre_norm = pre_norm
+        self.fpn_dim = round(d_model * mlp_ratio)
+        # ----------- Network parameters -----------
+        self.linear1 = nn.Linear(d_model, self.fpn_dim)
+        self.activation = get_activation(act_type)
+        self.dropout2 = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(self.fpn_dim, d_model)
+        self.dropout3 = nn.Dropout(dropout)
+        self.norm = nn.LayerNorm(d_model)
+
+    def forward(self, src):
+        if self.pre_norm:
+            src = self.norm(src)
+            src2 = self.linear2(self.dropout2(self.activation(self.linear1(src))))
+            src = src + self.dropout3(src2)
+        else:
+            src2 = self.linear2(self.dropout2(self.activation(self.linear1(src))))
+            src = src + self.dropout3(src2)
+            src = self.norm(src)
+        
+        return src

models/detectors/vitdet/loss.py

@@ -0,0 +1,163 @@
+import torch
+import torch.nn.functional as F
+
+try:
+    from .loss_utils import get_ious, get_world_size, is_dist_avail_and_initialized
+    from .matcher import AlignedSimOtaMatcher
+except:
+    from  loss_utils import get_ious, get_world_size, is_dist_avail_and_initialized
+    from  matcher import AlignedSimOtaMatcher
+
+
+class Criterion(object):
+    def __init__(self, cfg, num_classes=80):
+        # ------------ Basic parameters ------------
+        self.cfg = cfg
+        self.num_classes = num_classes
+        # --------------- Matcher config ---------------
+        self.matcher_hpy = cfg['matcher_hpy']
+        self.matcher = AlignedSimOtaMatcher(soft_center_radius = self.matcher_hpy['soft_center_radius'],
+                                            topk_candidates    = self.matcher_hpy['topk_candidates'],
+                                            num_classes        = num_classes,
+                                            )
+        # ------------- Loss weight -------------
+        self.weight_dict = {'loss_cls':  cfg['loss_coeff']['class'],
+                            'loss_box':  cfg['loss_coeff']['bbox'],
+                            'loss_giou': cfg['loss_coeff']['giou']}
+
+    def loss_classes(self, pred_cls, target, num_gts, beta=2.0):
+        # Quality FocalLoss
+        """
+            pred_cls: (torch.Tensor): [N, C]。
+            target:   (tuple([torch.Tensor], [torch.Tensor])): label -> (N,), score -> (N)
+        """
+        label, score = target
+        pred_sigmoid = pred_cls.sigmoid()
+        scale_factor = pred_sigmoid
+        zerolabel = scale_factor.new_zeros(pred_cls.shape)
+
+        ce_loss = F.binary_cross_entropy_with_logits(
+            pred_cls, zerolabel, reduction='none') * scale_factor.pow(beta)
+        
+        bg_class_ind = pred_cls.shape[-1]
+        pos = ((label >= 0) & (label < bg_class_ind)).nonzero().squeeze(1)
+        pos_label = label[pos].long()
+
+        scale_factor = score[pos] - pred_sigmoid[pos, pos_label]
+
+        ce_loss[pos, pos_label] = F.binary_cross_entropy_with_logits(
+            pred_cls[pos, pos_label], score[pos],
+            reduction='none') * scale_factor.abs().pow(beta)
+        
+        losses = {}
+        losses['loss_cls'] = ce_loss.sum() / num_gts
+
+        return losses
+    
+    def loss_bboxes(self, pred_reg, pred_box, gt_box, anchors, stride_tensors, num_gts):
+        # --------------- Compute L1 loss ---------------
+        ## xyxy -> cxcy&bwbh
+        gt_cxcy = (gt_box[..., :2] + gt_box[..., 2:]) * 0.5
+        gt_bwbh = gt_box[..., 2:] - gt_box[..., :2]
+        ## Encode gt box
+        gt_cxcy_encode = (gt_cxcy - anchors) / stride_tensors
+        gt_bwbh_encode = torch.log(gt_bwbh / stride_tensors)
+        gt_box_encode = torch.cat([gt_cxcy_encode, gt_bwbh_encode], dim=-1)
+        # L1 loss
+        loss_box = F.l1_loss(pred_reg, gt_box_encode, reduction='none')
+
+        # --------------- Compute GIoU loss ---------------
+        gious = get_ious(pred_box, gt_box, box_mode="xyxy", iou_type='giou')
+        loss_giou = 1.0 - gious
+
+        losses = {}
+        losses['loss_box'] = loss_box.sum() / num_gts
+        losses['loss_giou'] = loss_giou.sum() / num_gts
+
+        return losses
+    
+    def __call__(self, outputs, targets):        
+        """
+            outputs['pred_cls']: List(Tensor) [B, M, C]
+            outputs['pred_box']: List(Tensor) [B, M, 4]
+            outputs['pred_box']: List(Tensor) [B, M, 4]
+            outputs['strides']: List(Int) [8, 16, 32] output stride
+            targets: (List) [dict{'boxes': [...], 
+                                 'labels': [...], 
+                                 'orig_size': ...}, ...]
+        """
+        bs = outputs['pred_cls'][0].shape[0]
+        device = outputs['pred_cls'][0].device
+        anchors = outputs['anchors']
+        fpn_strides = outputs['strides']
+        stride_tensors = outputs['stride_tensors']
+        losses = dict()
+        # preds: [B, M, C]
+        cls_preds = torch.cat(outputs['pred_cls'], dim=1)
+        box_preds = torch.cat(outputs['pred_box'], dim=1)
+        reg_preds = torch.cat(outputs['pred_reg'], dim=1)
+        
+        # --------------- label assignment ---------------
+        cls_targets = []
+        box_targets = []
+        assign_metrics = []
+        for batch_idx in range(bs):
+            tgt_labels = targets[batch_idx]["labels"].to(device)  # [N,]
+            tgt_bboxes = targets[batch_idx]["boxes"].to(device)   # [N, 4]
+            assigned_result = self.matcher(fpn_strides=fpn_strides,
+                                           anchors=anchors,
+                                           pred_cls=cls_preds[batch_idx].detach(),
+                                           pred_box=box_preds[batch_idx].detach(),
+                                           gt_labels=tgt_labels,
+                                           gt_bboxes=tgt_bboxes
+                                           )
+            cls_targets.append(assigned_result['assigned_labels'])
+            box_targets.append(assigned_result['assigned_bboxes'])
+            assign_metrics.append(assigned_result['assign_metrics'])
+
+        # List[B, M, C] -> Tensor[BM, C]
+        cls_targets = torch.cat(cls_targets, dim=0)
+        box_targets = torch.cat(box_targets, dim=0)
+        assign_metrics = torch.cat(assign_metrics, dim=0)
+
+        # FG cat_id: [0, num_classes -1], BG cat_id: num_classes
+        bg_class_ind = self.num_classes
+        pos_inds = ((cls_targets >= 0) & (cls_targets < bg_class_ind)).nonzero().squeeze(1)
+        num_fgs = assign_metrics.sum()
+
+        if is_dist_avail_and_initialized():
+            torch.distributed.all_reduce(num_fgs)
+        num_fgs = (num_fgs / get_world_size()).clamp(1.0).item()
+
+        # ------------------ Classification loss ------------------
+        cls_preds = cls_preds.view(-1, self.num_classes)
+        loss_dict = self.loss_classes(cls_preds, (cls_targets, assign_metrics), num_fgs)
+        loss_dict = {k: loss_dict[k] * self.weight_dict[k] for k in loss_dict if k in self.weight_dict}
+        losses.update(loss_dict)
+
+        # ------------------ Regression loss ------------------
+        box_targets_pos = box_targets[pos_inds]
+        ## positive predictions
+        box_preds_pos = box_preds.view(-1, 4)[pos_inds]
+        reg_preds_pos = reg_preds.view(-1, 4)[pos_inds]
+
+        ## anchor tensors
+        anchors_tensors = torch.cat(anchors, dim=0)[None].repeat(bs, 1, 1)
+        anchors_tensors_pos = anchors_tensors.view(-1, 2)[pos_inds]
+
+        ## stride tensors
+        stride_tensors = torch.cat(stride_tensors, dim=0)[None].repeat(bs, 1, 1)
+        stride_tensors_pos = stride_tensors.view(-1, 1)[pos_inds]
+
+        ## aux loss
+        loss_dict = self.loss_bboxes(reg_preds_pos, box_preds_pos, box_targets_pos, anchors_tensors_pos, stride_tensors_pos, num_fgs)
+        loss_dict = {k: loss_dict[k] * self.weight_dict[k] for k in loss_dict if k in self.weight_dict}
+        losses.update(loss_dict)
+
+        return losses
+    
+
+def build_criterion(cfg, num_classes):
+    criterion = Criterion(cfg, num_classes)
+
+    return criterion

models/detectors/vitdet/loss_utils.py

@@ -0,0 +1,87 @@
+import torch
+import torch.distributed as dist
+from torchvision.ops.boxes import box_area
+
+
+# ------------------------- For box -------------------------
+def box_iou(boxes1, boxes2):
+    area1 = box_area(boxes1)
+    area2 = box_area(boxes2)
+
+    lt = torch.max(boxes1[:, None, :2], boxes2[:, :2])  # [N,M,2]
+    rb = torch.min(boxes1[:, None, 2:], boxes2[:, 2:])  # [N,M,2]
+
+    wh = (rb - lt).clamp(min=0)  # [N,M,2]
+    inter = wh[:, :, 0] * wh[:, :, 1]  # [N,M]
+
+    union = area1[:, None] + area2 - inter
+
+    iou = inter / union
+    return iou, union
+
+def get_ious(bboxes1,
+             bboxes2,
+             box_mode="xyxy",
+             iou_type="iou"):
+    """
+    Compute iou loss of type ['iou', 'giou', 'linear_iou']
+
+    Args:
+        inputs (tensor): pred values
+        targets (tensor): target values
+        weight (tensor): loss weight
+        box_mode (str): 'xyxy' or 'ltrb', 'ltrb' is currently supported.
+        loss_type (str): 'giou' or 'iou' or 'linear_iou'
+        reduction (str): reduction manner
+
+    Returns:
+        loss (tensor): computed iou loss.
+    """
+    if box_mode == "ltrb":
+        bboxes1 = torch.cat((-bboxes1[..., :2], bboxes1[..., 2:]), dim=-1)
+        bboxes2 = torch.cat((-bboxes2[..., :2], bboxes2[..., 2:]), dim=-1)
+    elif box_mode != "xyxy":
+        raise NotImplementedError
+
+    eps = torch.finfo(torch.float32).eps
+
+    bboxes1_area = (bboxes1[..., 2] - bboxes1[..., 0]).clamp_(min=0) \
+        * (bboxes1[..., 3] - bboxes1[..., 1]).clamp_(min=0)
+    bboxes2_area = (bboxes2[..., 2] - bboxes2[..., 0]).clamp_(min=0) \
+        * (bboxes2[..., 3] - bboxes2[..., 1]).clamp_(min=0)
+
+    w_intersect = (torch.min(bboxes1[..., 2], bboxes2[..., 2])
+                   - torch.max(bboxes1[..., 0], bboxes2[..., 0])).clamp_(min=0)
+    h_intersect = (torch.min(bboxes1[..., 3], bboxes2[..., 3])
+                   - torch.max(bboxes1[..., 1], bboxes2[..., 1])).clamp_(min=0)
+
+    area_intersect = w_intersect * h_intersect
+    area_union = bboxes2_area + bboxes1_area - area_intersect
+    ious = area_intersect / area_union.clamp(min=eps)
+
+    if iou_type == "iou":
+        return ious
+    elif iou_type == "giou":
+        g_w_intersect = torch.max(bboxes1[..., 2], bboxes2[..., 2]) \
+            - torch.min(bboxes1[..., 0], bboxes2[..., 0])
+        g_h_intersect = torch.max(bboxes1[..., 3], bboxes2[..., 3]) \
+            - torch.min(bboxes1[..., 1], bboxes2[..., 1])
+        ac_uion = g_w_intersect * g_h_intersect
+        gious = ious - (ac_uion - area_union) / ac_uion.clamp(min=eps)
+        return gious
+    else:
+        raise NotImplementedError
+
+
+# ------------------------- For distributed -------------------------
+def is_dist_avail_and_initialized():
+    if not dist.is_available():
+        return False
+    if not dist.is_initialized():
+        return False
+    return True
+
+def get_world_size():
+    if not is_dist_avail_and_initialized():
+        return 1
+    return dist.get_world_size()

models/detectors/vitdet/matcher.py

@@ -0,0 +1,164 @@
+# ------------------------------------------------------------------------------------------
+# This code referenced to https://github.com/open-mmlab/mmyolo/models/task_modules/assigners/batch_dsl_assigner.py
+# ------------------------------------------------------------------------------------------
+import torch
+import torch.nn.functional as F
+
+try:
+    from .loss_utils import box_iou
+except:
+    from  loss_utils import box_iou
+
+
+# -------------------------- Aligned SimOTA assigner --------------------------
+class AlignedSimOtaMatcher(object):
+    def __init__(self, num_classes, soft_center_radius=3.0, topk_candidates=13):
+        self.num_classes = num_classes
+        self.soft_center_radius = soft_center_radius
+        self.topk_candidates = topk_candidates
+
+    @torch.no_grad()
+    def __call__(self, 
+                 fpn_strides, 
+                 anchors, 
+                 pred_cls, 
+                 pred_box, 
+                 gt_labels,
+                 gt_bboxes):
+        # [M,]
+        strides = torch.cat([torch.ones_like(anchor_i[:, 0]) * stride_i
+                                for stride_i, anchor_i in zip(fpn_strides, anchors)], dim=-1)
+        # List[F, M, 2] -> [M, 2]
+        num_gt = len(gt_labels)
+        anchors = torch.cat(anchors, dim=0)
+
+        # check gt
+        if num_gt == 0 or gt_bboxes.max().item() == 0.:
+            return {
+                'assigned_labels': gt_labels.new_full(pred_cls[..., 0].shape,
+                                                      self.num_classes,
+                                                      dtype=torch.long),
+                'assigned_bboxes': gt_bboxes.new_full(pred_box.shape, 0),
+                'assign_metrics': gt_bboxes.new_full(pred_cls[..., 0].shape, 0)
+            }
+        
+        # get inside points: [N, M]
+        is_in_gt = self.find_inside_points(gt_bboxes, anchors)
+        valid_mask = is_in_gt.sum(dim=0) > 0  # [M,]
+
+        # ----------------------------------- soft center prior -----------------------------------
+        gt_center = (gt_bboxes[..., :2] + gt_bboxes[..., 2:]) / 2.0
+        distance = (anchors.unsqueeze(0) - gt_center.unsqueeze(1)
+                    ).pow(2).sum(-1).sqrt() / strides.unsqueeze(0)  # [N, M]
+        distance = distance * valid_mask.unsqueeze(0)
+        soft_center_prior = torch.pow(10, distance - self.soft_center_radius)
+
+        # ----------------------------------- regression cost -----------------------------------
+        pair_wise_ious, _ = box_iou(gt_bboxes, pred_box)  # [N, M]
+        pair_wise_ious_loss = -torch.log(pair_wise_ious + 1e-8) * 3.0
+
+        # ----------------------------------- classification cost -----------------------------------
+        ## select the predicted scores corresponded to the gt_labels
+        pairwise_pred_scores = pred_cls.permute(1, 0)  # [M, C] -> [C, M]
+        pairwise_pred_scores = pairwise_pred_scores[gt_labels.long(), :].float()   # [N, M]
+        ## scale factor
+        scale_factor = (pair_wise_ious - pairwise_pred_scores.sigmoid()).abs().pow(2.0)
+        ## cls cost
+        pair_wise_cls_loss = F.binary_cross_entropy_with_logits(
+            pairwise_pred_scores, pair_wise_ious,
+            reduction="none") * scale_factor # [N, M]
+            
+        del pairwise_pred_scores
+
+        ## foreground cost matrix
+        cost_matrix = pair_wise_cls_loss + pair_wise_ious_loss + soft_center_prior
+        max_pad_value = torch.ones_like(cost_matrix) * 1e9
+        cost_matrix = torch.where(valid_mask[None].repeat(num_gt, 1),   # [N, M]
+                                  cost_matrix, max_pad_value)
+
+        # ----------------------------------- dynamic label assignment -----------------------------------
+        matched_pred_ious, matched_gt_inds, fg_mask_inboxes = self.dynamic_k_matching(
+            cost_matrix, pair_wise_ious, num_gt)
+        del pair_wise_cls_loss, cost_matrix, pair_wise_ious, pair_wise_ious_loss
+
+        # -----------------------------------process assigned labels -----------------------------------
+        assigned_labels = gt_labels.new_full(pred_cls[..., 0].shape,
+                                             self.num_classes)  # [M,]
+        assigned_labels[fg_mask_inboxes] = gt_labels[matched_gt_inds].squeeze(-1)
+        assigned_labels = assigned_labels.long()  # [M,]
+
+        assigned_bboxes = gt_bboxes.new_full(pred_box.shape, 0)        # [M, 4]
+        assigned_bboxes[fg_mask_inboxes] = gt_bboxes[matched_gt_inds]  # [M, 4]
+
+        assign_metrics = gt_bboxes.new_full(pred_cls[..., 0].shape, 0) # [M, 4]
+        assign_metrics[fg_mask_inboxes] = matched_pred_ious            # [M, 4]
+
+        assigned_dict = dict(
+            assigned_labels=assigned_labels,
+            assigned_bboxes=assigned_bboxes,
+            assign_metrics=assign_metrics
+            )
+        
+        return assigned_dict
+
+    def find_inside_points(self, gt_bboxes, anchors):
+        """
+            gt_bboxes: Tensor -> [N, 2]
+            anchors:   Tensor -> [M, 2]
+        """
+        num_anchors = anchors.shape[0]
+        num_gt = gt_bboxes.shape[0]
+
+        anchors_expand = anchors.unsqueeze(0).repeat(num_gt, 1, 1)           # [N, M, 2]
+        gt_bboxes_expand = gt_bboxes.unsqueeze(1).repeat(1, num_anchors, 1)  # [N, M, 4]
+
+        # offset
+        lt = anchors_expand - gt_bboxes_expand[..., :2]
+        rb = gt_bboxes_expand[..., 2:] - anchors_expand
+        bbox_deltas = torch.cat([lt, rb], dim=-1)
+
+        is_in_gts = bbox_deltas.min(dim=-1).values > 0
+
+        return is_in_gts
+    
+    def dynamic_k_matching(self, cost_matrix, pairwise_ious, num_gt):
+        """Use IoU and matching cost to calculate the dynamic top-k positive
+        targets.
+
+        Args:
+            cost_matrix (Tensor): Cost matrix.
+            pairwise_ious (Tensor): Pairwise iou matrix.
+            num_gt (int): Number of gt.
+            valid_mask (Tensor): Mask for valid bboxes.
+        Returns:
+            tuple: matched ious and gt indexes.
+        """
+        matching_matrix = torch.zeros_like(cost_matrix, dtype=torch.uint8)
+        # select candidate topk ious for dynamic-k calculation
+        candidate_topk = min(self.topk_candidates, pairwise_ious.size(1))
+        topk_ious, _ = torch.topk(pairwise_ious, candidate_topk, dim=1)
+        # calculate dynamic k for each gt
+        dynamic_ks = torch.clamp(topk_ious.sum(1).int(), min=1)
+
+        # sorting the batch cost matirx is faster than topk
+        _, sorted_indices = torch.sort(cost_matrix, dim=1)
+        for gt_idx in range(num_gt):
+            topk_ids = sorted_indices[gt_idx, :dynamic_ks[gt_idx]]
+            matching_matrix[gt_idx, :][topk_ids] = 1
+
+        del topk_ious, dynamic_ks, topk_ids
+
+        prior_match_gt_mask = matching_matrix.sum(0) > 1
+        if prior_match_gt_mask.sum() > 0:
+            cost_min, cost_argmin = torch.min(
+                cost_matrix[:, prior_match_gt_mask], dim=0)
+            matching_matrix[:, prior_match_gt_mask] *= 0
+            matching_matrix[cost_argmin, prior_match_gt_mask] = 1
+
+        # get foreground mask inside box and center prior
+        fg_mask_inboxes = matching_matrix.sum(0) > 0
+        matched_pred_ious = (matching_matrix *
+                             pairwise_ious).sum(0)[fg_mask_inboxes]
+        matched_gt_inds = matching_matrix[:, fg_mask_inboxes].argmax(0)
+
+        return matched_pred_ious, matched_gt_inds, fg_mask_inboxes

models/detectors/vitdet/vitdet.py

@@ -0,0 +1,230 @@
+# --------------- Torch components ---------------
+import torch
+import torch.nn as nn
+
+# --------------- Model components ---------------
+try:
+    from .vitdet_encoder import build_image_encoder
+    from .vitdet_decoder import build_decoder
+    from .vitdet_head    import build_predictor
+    from .basic_modules.basic import multiclass_nms
+except:
+    from  vitdet_encoder import build_image_encoder
+    from  vitdet_decoder import build_decoder
+    from  vitdet_head    import build_predictor
+    from  basic_modules.basic import multiclass_nms
+
+
+
+# Real-time ViT-based Object Detector
+class ViTDet(nn.Module):
+    def __init__(self,
+                 cfg,
+                 device,
+                 num_classes = 20,
+                 conf_thresh = 0.01,
+                 nms_thresh  = 0.5,
+                 topk        = 1000,
+                 trainable   = False,
+                 deploy      = False,
+                 no_multi_labels    = False,
+                 nms_class_agnostic = False,
+                 ):
+        super(ViTDet, self).__init__()
+        # ---------------------- Basic Parameters ----------------------
+        self.cfg = cfg
+        self.device = device
+        self.strides = cfg['stride']
+        self.num_classes = num_classes
+        ## Scale hidden channels by width_factor
+        cfg['hidden_dim'] = round(cfg['hidden_dim'] * cfg['width'])
+        cfg['pretrained'] = cfg['pretrained'] & trainable
+        ## Post-process parameters
+        self.conf_thresh = conf_thresh
+        self.nms_thresh = nms_thresh
+        self.topk = topk
+        self.deploy = deploy
+        self.no_multi_labels = no_multi_labels
+        self.nms_class_agnostic = nms_class_agnostic
+        
+        # ---------------------- Network Parameters ----------------------
+        ## ----------- Encoder -----------
+        self.encoder = build_image_encoder(cfg)
+
+        ## ----------- Decoder -----------
+        self.decoder = build_decoder(cfg, self.encoder.fpn_dims, num_levels=3)
+        
+        ## ----------- Preds -----------
+        self.predictor = build_predictor(cfg, self.strides, num_classes, 4, 3)
+
+    def post_process(self, cls_preds, box_preds):
+        """
+        Input:
+            cls_preds: List[np.array] -> [[M, C], ...]
+            box_preds: List[np.array] -> [[M, 4], ...]
+        Output:
+            bboxes: np.array -> [N, 4]
+            scores: np.array -> [N,]
+            labels: np.array -> [N,]
+        """
+        all_scores = []
+        all_labels = []
+        all_bboxes = []
+        
+        for cls_pred_i, box_pred_i in zip(cls_preds, box_preds):
+            cls_pred_i = cls_pred_i[0]
+            box_pred_i = box_pred_i[0]
+            if self.no_multi_labels:
+                # [M,]
+                scores, labels = torch.max(cls_pred_i.sigmoid(), dim=1)
+
+                # Keep top k top scoring indices only.
+                num_topk = min(self.topk_candidates, box_pred_i.size(0))
+
+                # topk candidates
+                predicted_prob, topk_idxs = scores.sort(descending=True)
+                topk_scores = predicted_prob[:num_topk]
+                topk_idxs = topk_idxs[:num_topk]
+
+                # filter out the proposals with low confidence score
+                keep_idxs = topk_scores > self.conf_thresh
+                scores = topk_scores[keep_idxs]
+                topk_idxs = topk_idxs[keep_idxs]
+
+                labels = labels[topk_idxs]
+                bboxes = box_pred_i[topk_idxs]
+            else:
+                # [M, C] -> [MC,]
+                scores_i = cls_pred_i.sigmoid().flatten()
+
+                # Keep top k top scoring indices only.
+                num_topk = min(self.topk_candidates, box_pred_i.size(0))
+
+                # torch.sort is actually faster than .topk (at least on GPUs)
+                predicted_prob, topk_idxs = scores_i.sort(descending=True)
+                topk_scores = predicted_prob[:num_topk]
+                topk_idxs = topk_idxs[:num_topk]
+
+                # filter out the proposals with low confidence score
+                keep_idxs = topk_scores > self.conf_thresh
+                scores = topk_scores[keep_idxs]
+                topk_idxs = topk_idxs[keep_idxs]
+
+                anchor_idxs = torch.div(topk_idxs, self.num_classes, rounding_mode='floor')
+                labels = topk_idxs % self.num_classes
+
+                bboxes = box_pred_i[anchor_idxs]
+
+            all_scores.append(scores)
+            all_labels.append(labels)
+            all_bboxes.append(bboxes)
+
+        scores = torch.cat(all_scores, dim=0)
+        labels = torch.cat(all_labels, dim=0)
+        bboxes = torch.cat(all_bboxes, dim=0)
+
+        if not self.deploy:
+            # to cpu & numpy
+            scores = scores.cpu().numpy()
+            labels = labels.cpu().numpy()
+            bboxes = bboxes.cpu().numpy()
+
+            # nms
+            scores, labels, bboxes = multiclass_nms(
+                scores, labels, bboxes, self.nms_thresh, self.num_classes, self.nms_class_agnostic)
+
+        return bboxes, scores, labels
+    
+    def forward(self, x):
+        # ---------------- Backbone ----------------
+        pyramid_feats = self.encoder(x)
+
+        # ---------------- Heads ----------------
+        outputs = self.decoder(pyramid_feats)
+
+        # ---------------- Preds ----------------
+        outputs = self.predictor(outputs['cls_feats'], outputs['reg_feats'])
+
+        if not self.training:
+            cls_pred = outputs["pred_cls"]
+            box_pred = outputs["pred_box"]
+            # post process
+            bboxes, scores, labels = self.post_process(cls_pred, box_pred)
+
+            outputs = {
+                "scores": scores,
+                "labels": labels,
+                "bboxes": bboxes
+            }
+                    
+        return outputs
+        
+
+if __name__ == '__main__':
+    import time
+    from thop import profile
+    from loss import build_criterion
+
+    # Model config
+    cfg = {
+        'width': 1.0,
+        'depth': 1.0,
+        'out_stride': [8, 16, 32],
+        # Image Encoder - Backbone
+        'backbone': 'resnet18',
+        'backbone_norm': 'BN',
+        'res5_dilation': False,
+        'pretrained': True,
+        'pretrained_weight': 'imagenet1k_v1',
+        'freeze_at': 0,
+        'freeze_stem_only': False,
+        'out_stride': [8, 16, 32],
+        'max_stride': 32,
+        # Convolutional Decoder
+        'hidden_dim': 256,
+        'decoder': 'det_decoder',
+        'de_num_cls_layers': 2,
+        'de_num_reg_layers': 2,
+        'de_act': 'silu',
+        'de_norm': 'BN',
+        # Matcher
+        'matcher_hpy': {'soft_center_radius': 2.5,
+                        'topk_candidates': 13,},
+        # Loss
+        'use_vfl': True,
+        'loss_coeff': {'class': 1,
+                       'bbox': 1,
+                       'giou': 2,},
+        }
+    bs = 1
+    # Create a batch of images & targets
+    image = torch.randn(bs, 3, 640, 640).cuda()
+    targets = [{
+        'labels': torch.tensor([2, 4, 5, 8]).long().cuda(),
+        'boxes':  torch.tensor([[0, 0, 10, 10], [12, 23, 56, 70], [0, 10, 20, 30], [50, 60, 55, 150]]).float().cuda() / 640.
+    }] * bs
+
+    # Create model
+    model = ViTDet(cfg, num_classes=20)
+    model.train().cuda()
+
+    # Create criterion
+    criterion = build_criterion(cfg, num_classes=20)
+
+    # Model inference
+    t0 = time.time()
+    outputs = model(image, targets)
+    t1 = time.time()
+    print('Infer time: ', t1 - t0)
+
+    # Compute loss
+    loss = criterion(outputs, targets)
+    for k in loss.keys():
+        print("{} : {}".format(k, loss[k].item()))
+
+    print('==============================')
+    model.eval()
+    flops, params = profile(model, inputs=(image, ), verbose=False)
+    print('==============================')
+    print('GFLOPs : {:.2f}'.format(flops / 1e9 * 2))
+    print('Params : {:.2f} M'.format(params / 1e6))

models/detectors/vitdet/vitdet_decoder.py

@@ -0,0 +1,187 @@
+import torch
+import torch.nn as nn
+
+try:
+    from .basic_modules.basic import BasicConv
+except:
+    from  basic_modules.basic import BasicConv
+
+
+def build_decoder(cfg, in_dims, num_levels=3):
+    if cfg['decoder'] == "det_decoder":
+        decoder = MultiDetHead(cfg, in_dims, num_levels)
+    elif cfg['decoder'] == "seg_decoder":
+        decoder = MaskHead()
+    elif cfg['decoder'] == "pos_decoder":
+        decoder = PoseHead()
+
+    return decoder
+
+
+# ---------------------------- Detection Head ----------------------------
+## Single-level Detection Head
+class SingleDetHead(nn.Module):
+    def __init__(self,
+                 in_dim       :int  = 256,
+                 cls_head_dim :int  = 256,
+                 reg_head_dim :int  = 256,
+                 num_cls_head :int  = 2,
+                 num_reg_head :int  = 2,
+                 act_type     :str  = "silu",
+                 norm_type    :str  = "BN",
+                 ):
+        super().__init__()
+        # --------- Basic Parameters ----------
+        self.in_dim = in_dim
+        self.num_cls_head = num_cls_head
+        self.num_reg_head = num_reg_head
+        self.act_type = act_type
+        self.norm_type = norm_type
+        
+        # --------- Network Parameters ----------
+        ## cls head
+        cls_feats = []
+        self.cls_head_dim = cls_head_dim
+        for i in range(num_cls_head):
+            if i == 0:
+                cls_feats.append(
+                    BasicConv(in_dim, self.cls_head_dim,
+                              kernel_size=3, padding=1, stride=1, 
+                              act_type=act_type, norm_type=norm_type)
+                              )
+            else:
+                cls_feats.append(
+                    BasicConv(self.cls_head_dim, self.cls_head_dim,
+                              kernel_size=3, padding=1, stride=1, 
+                              act_type=act_type, norm_type=norm_type)
+                              )
+        ## reg head
+        reg_feats = []
+        self.reg_head_dim = reg_head_dim
+        for i in range(num_reg_head):
+            if i == 0:
+                cls_feats.append(
+                    BasicConv(in_dim, self.reg_head_dim,
+                              kernel_size=3, padding=1, stride=1, 
+                              act_type=act_type, norm_type=norm_type)
+                              )
+            else:
+                cls_feats.append(
+                    BasicConv(self.reg_head_dim, self.reg_head_dim,
+                              kernel_size=3, padding=1, stride=1, 
+                              act_type=act_type, norm_type=norm_type)
+                              )
+        self.cls_feats = nn.Sequential(*cls_feats)
+        self.reg_feats = nn.Sequential(*reg_feats)
+
+        self.init_weights()
+        
+    def init_weights(self):
+        """Initialize the parameters."""
+        for m in self.modules():
+            if isinstance(m, torch.nn.Conv2d):
+                # In order to be consistent with the source code,
+                # reset the Conv2d initialization parameters
+                m.reset_parameters()
+
+    def forward(self, x):
+        """
+            in_feats: (Tensor) [B, C, H, W]
+        """
+        cls_feats = self.cls_feats(x)
+        reg_feats = self.reg_feats(x)
+
+        return cls_feats, reg_feats
+    
+## Multi-level Detection Head
+class MultiDetHead(nn.Module):
+    def __init__(self, cfg, in_dims, num_levels=3):
+        super().__init__()
+        ## ----------- Network Parameters -----------
+        self.multi_level_heads = nn.ModuleList(
+            [SingleDetHead(in_dim       = in_dims[level],
+                           cls_head_dim = cfg['hidden_dim'],
+                           reg_head_dim = cfg['hidden_dim'],
+                           num_cls_head = cfg['de_num_cls_layers'],
+                           num_reg_head = cfg['de_num_reg_layers'],
+                           act_type     = cfg['de_act'],
+                           norm_type    = cfg['de_norm'],
+                           )
+                           for level in range(num_levels)
+                           ])
+        # --------- Basic Parameters ----------
+        self.in_dims = in_dims
+        self.cls_head_dim = self.multi_level_heads[0].cls_head_dim
+        self.reg_head_dim = self.multi_level_heads[0].reg_head_dim
+
+    def forward(self, feats):
+        """
+            feats: List[(Tensor)] [[B, C, H, W], ...]
+        """
+        cls_feats = []
+        reg_feats = []
+        for feat, head in zip(feats, self.multi_level_heads):
+            # ---------------- Pred ----------------
+            cls_feat, reg_feat = head(feat)
+
+            cls_feats.append(cls_feat)
+            reg_feats.append(reg_feat)
+
+        outputs = {
+            "cls_feat": cls_feats,
+            "reg_feat": reg_feats
+        }
+
+        return outputs
+
+
+# ---------------------------- Segmentation Head ----------------------------
+class MaskHead(nn.Module):
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+
+    def forward(self, x):
+        return
+
+
+# ---------------------------- Human-Pose Head ----------------------------
+class PoseHead(nn.Module):
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+
+    def forward(self, x):
+        return
+
+
+if __name__ == '__main__':
+    import time
+    from thop import profile
+    cfg = {
+        'width': 1.0,
+        'depth': 1.0,
+        # Decoder parameters
+        'hidden_dim': 256,
+        'decoder': 'det_decoder',
+        'de_num_cls_layers': 2,
+        'de_num_reg_layers': 2,
+        'de_act': 'silu',
+        'de_norm': 'BN',
+    }
+    fpn_dims = [256, 256, 256]
+    out_dim = 256
+    # Head-1
+    model = build_decoder(cfg, fpn_dims, num_levels=3)
+    print(model)
+    fpn_feats = [torch.randn(1, fpn_dims[0], 80, 80), torch.randn(1, fpn_dims[1], 40, 40), torch.randn(1, fpn_dims[2], 20, 20)]
+    t0 = time.time()
+    outputs = model(fpn_feats)
+    t1 = time.time()
+    print('Time: ', t1 - t0)
+    # for out in outputs:
+    #     print(out.shape)
+
+    print('==============================')
+    flops, params = profile(model, inputs=(fpn_feats, ), verbose=False)
+    print('==============================')
+    print('Head-1: GFLOPs : {:.2f}'.format(flops / 1e9 * 2))
+    print('Head-1: Params : {:.2f} M'.format(params / 1e6))

models/detectors/vitdet/vitdet_encoder.py

@@ -0,0 +1,101 @@
+import torch
+import torch.nn as nn
+
+try:
+    from .basic_modules.basic    import BasicConv, UpSampleWrapper
+    from .basic_modules.backbone import build_backbone
+except:
+    from  basic_modules.basic    import BasicConv, UpSampleWrapper
+    from  basic_modules.backbone import build_backbone
+
+
+# ----------------- Image Encoder -----------------
+def build_image_encoder(cfg):
+    return ImageEncoder(cfg)
+
+class ImageEncoder(nn.Module):
+    def __init__(self, cfg):
+        super().__init__()
+        # ---------------- Basic settings ----------------
+        ## Basic parameters
+        self.cfg = cfg
+        ## Network parameters
+        self.stride = 16
+        self.fpn_dims = [cfg['hidden_dim']] * 3
+        self.hidden_dim = cfg['hidden_dim']
+        
+        # ---------------- Network settings ----------------
+        ## Backbone Network
+        self.backbone, backbone_dim = build_backbone(cfg, cfg['pretrained'])
+
+        ## Input projection
+        self.input_proj = BasicConv(backbone_dim, cfg['hidden_dim'],
+                                    kernel_size=1,
+                                    act_type=None, norm_type='BN')
+
+        ## Upsample layer
+        self.upsample = UpSampleWrapper(cfg['hidden_dim'], 2.0)
+        
+        ## Downsample layer
+        self.downsample = BasicConv(cfg['hidden_dim'], cfg['hidden_dim'],
+                                    kernel_size=3, padding=1, stride=2,
+                                    act_type=None, norm_type='BN')
+
+        ## Output projection
+        self.output_projs = nn.ModuleList([BasicConv(cfg['hidden_dim'], cfg['hidden_dim'],
+                                                     kernel_size=3, padding=1,
+                                                     act_type='silu', norm_type='BN')
+                                                     ] * 3)
+
+
+    def forward(self, x):
+        # Backbone
+        feat = self.backbone(x)
+
+        # Input proj
+        feat = self.input_proj(feat)
+
+        # FPN
+        feat_up = self.upsample(feat)
+        feat_ds = self.downsample(feat)
+
+        # Multi level features: [P3, P4, P5]
+        pyramid_feats = [self.output_projs[0](feat_up),
+                         self.output_projs[1](feat),
+                         self.output_projs[2](feat_ds)]
+
+        return pyramid_feats
+
+
+if __name__ == '__main__':
+    import time
+    from thop import profile
+    cfg = {
+        'width': 1.0,
+        'depth': 1.0,
+        'out_stride': 16,
+        'hidden_dim': 256,
+        # Image Encoder - Backbone
+        'backbone': 'resnet50',
+        'backbone_norm': 'FrozeBN',
+        'pretrained': True,
+        'freeze_at': 0,
+        'freeze_stem_only': False,
+    }
+    x = torch.rand(2, 3, 640, 640)
+    model = build_image_encoder(cfg)
+    model.train()
+
+    t0 = time.time()
+    outputs = model(x)
+    t1 = time.time()
+    print('Time: ', t1 - t0)
+    print(outputs.shape)
+
+    print('==============================')
+    model.eval()
+    x = torch.rand(1, 3, 640, 640)
+    flops, params = profile(model, inputs=(x, ), verbose=False)
+    print('==============================')
+    print('GFLOPs : {:.2f}'.format(flops / 1e9 * 2))
+    print('Params : {:.2f} M'.format(params / 1e6))

models/detectors/vitdet/vitdet_head.py

@@ -0,0 +1,181 @@
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def build_predictor(cfg, strides, num_classes, num_coords=4, num_levels=3):
+    if cfg['task'] == 'detection':
+        pred_layer = MultiDetPredLayer(cls_dim     = cfg['hidden_dim'],
+                                       reg_dim     = cfg['hidden_dim'],
+                                       strides     = strides,
+                                       num_classes = num_classes,
+                                       num_coords  = num_coords,
+                                       num_levels  = num_levels
+                                       )
+        
+    elif cfg['task'] == 'segmentation':
+        raise NotImplementedError
+
+    elif cfg['task'] == 'pose_estimation':
+        raise NotImplementedError
+
+    return pred_layer
+
+
+# ---------------------------- Detection predictor ----------------------------
+## Single-level Detection Prediction Layer
+class SingleDetPDLayer(nn.Module):
+    def __init__(self,
+                 cls_dim     :int = 256,
+                 reg_dim     :int = 256,
+                 stride      :int = 32,
+                 num_classes :int = 80,
+                 num_coords  :int = 4):
+        super().__init__()
+        # --------- Basic Parameters ----------
+        self.stride = stride
+        self.cls_dim = cls_dim
+        self.reg_dim = reg_dim
+        self.num_classes = num_classes
+        self.num_coords = num_coords
+
+        # --------- Network Parameters ----------
+        self.cls_pred = nn.Conv2d(cls_dim, num_classes, kernel_size=1)
+        self.reg_pred = nn.Conv2d(reg_dim, num_coords, kernel_size=1)                
+
+        self.init_bias()
+        
+    def init_bias(self):
+        # cls pred bias
+        b = self.cls_pred.bias.view(1, -1)
+        b.data.fill_(math.log(5 / self.num_classes / (640. / self.stride) ** 2))
+        self.cls_pred.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
+        # reg pred bias
+        b = self.reg_pred.bias.view(-1, )
+        b.data.fill_(1.0)
+        self.reg_pred.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
+
+    def generate_anchors(self, fmp_size):
+        """
+            fmp_size: (List) [H, W]
+        """
+        # generate grid cells
+        fmp_h, fmp_w = fmp_size
+        anchor_y, anchor_x = torch.meshgrid([torch.arange(fmp_h), torch.arange(fmp_w)])
+        # [H, W, 2] -> [HW, 2]
+        anchors = torch.stack([anchor_x, anchor_y], dim=-1).float().view(-1, 2)
+        anchors += 0.5  # add center offset
+        anchors *= self.stride
+
+        return anchors
+        
+    def forward(self, cls_feat, reg_feat):
+        # pred
+        cls_pred = self.cls_pred(cls_feat)
+        reg_pred = self.reg_pred(reg_feat)
+
+        # generate anchor boxes: [M, 4]
+        B, _, H, W = cls_pred.size()
+        fmp_size = [H, W]
+        anchors = self.generate_anchors(fmp_size)
+        anchors = anchors.to(cls_pred.device)
+        # stride tensor: [M, 1]
+        stride_tensor = torch.ones_like(anchors[..., :1]) * self.stride
+        
+        # [B, C, H, W] -> [B, H, W, C] -> [B, M, C]
+        cls_pred = cls_pred.permute(0, 2, 3, 1).contiguous().view(B, -1, self.num_classes)
+        reg_pred = reg_pred.permute(0, 2, 3, 1).contiguous().view(B, -1, 4)
+
+        # ---------------- Decode bbox ----------------
+        ctr_pred = reg_pred[..., :2] * self.stride + anchors[..., :2]
+        wh_pred = torch.exp(reg_pred[..., 2:]) * self.stride
+        pred_x1y1 = ctr_pred - wh_pred * 0.5
+        pred_x2y2 = ctr_pred + wh_pred * 0.5
+        box_pred = torch.cat([pred_x1y1, pred_x2y2], dim=-1)
+
+        # output dict
+        outputs = {"pred_cls": cls_pred,             # (Tensor) [B, M, C]
+                   "pred_reg": reg_pred,             # (Tensor) [B, M, 4]
+                   "pred_box": box_pred,             # (Tensor) [B, M, 4] 
+                   "anchors": anchors,               # (Tensor) [M, 2]
+                   "stride": self.stride,            # (Int)
+                   "stride_tensors": stride_tensor   # List(Tensor) [M, 1]
+                   }
+
+        return outputs
+
+# Multi-level pred layer
+class MultiDetPredLayer(nn.Module):
+    def __init__(self,
+                 cls_dim,
+                 reg_dim,
+                 strides,
+                 num_classes :int = 80,
+                 num_coords  :int = 4,
+                 num_levels  :int = 3):
+        super().__init__()
+        # --------- Basic Parameters ----------
+        self.cls_dim = cls_dim
+        self.reg_dim = reg_dim
+        self.strides = strides
+        self.num_classes = num_classes
+        self.num_coords = num_coords
+        self.num_levels = num_levels
+
+        # ----------- Network Parameters -----------
+        ## multi-level pred layers
+        self.multi_level_preds = nn.ModuleList(
+            [SingleDetPDLayer(cls_dim     = cls_dim,
+                              reg_dim     = reg_dim,
+                              stride      = strides[level],
+                              num_classes = num_classes,
+                              num_coords  = num_coords)
+                              for level in range(num_levels)
+                              ])
+        
+    def forward(self, cls_feats, reg_feats):
+        all_anchors = []
+        all_strides = []
+        all_cls_preds = []
+        all_box_preds = []
+        all_reg_preds = []
+        for level in range(self.num_levels):
+            # ---------------- Single level prediction ----------------
+            outputs = self.multi_level_preds[level](cls_feats[level], reg_feats[level])
+
+            # collect results
+            all_cls_preds.append(outputs["pred_cls"])
+            all_box_preds.append(outputs["pred_box"])
+            all_reg_preds.append(outputs["pred_reg"])
+            all_anchors.append(outputs["anchors"])
+            all_strides.append(outputs["stride_tensors"])
+        
+        # output dict
+        outputs = {"pred_cls": all_cls_preds,      # List(Tensor) [B, M, C]
+                   "pred_box": all_box_preds,      # List(Tensor) [B, M, 4]
+                   "pred_reg": all_reg_preds,      # List(Tensor) [B, M, 4]
+                   "anchors": all_anchors,         # List(Tensor) [M, 2]
+                   "strides": self.strides,        # List(Int) [8, 16, 32]
+                   "stride_tensors": all_strides   # List(Tensor) [M, 1]
+                   }
+
+        return outputs
+    
+
+# -------------------- Segmentation predictor --------------------
+class MaskPDLayer(nn.Module):
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+    
+    def forward(self, x):
+        return
+
+
+# -------------------- Human-Pose predictor --------------------
+class PosePDLayer(nn.Module):
+    def __init__(self, *args, **kwargs) -> None:
+        super().__init__(*args, **kwargs)
+    
+    def forward(self, x):
+        return