Skip to content

Reference for ultralytics/nn/modules/head.py

Note

This file is available at https://github.com/ultralytics/ultralytics/blob/main/ultralytics/nn/modules/head.py. If you spot a problem please help fix it by contributing a Pull Request 🛠️. Thank you 🙏!


ultralytics.nn.modules.head.Detect

Detect(nc=80, ch=())

Bases: Module

YOLO Detect head for detection models.

Source code in ultralytics/nn/modules/head.py
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
def __init__(self, nc=80, ch=()):
    """Initialize the YOLO detection layer with specified number of classes and channels."""
    super().__init__()
    self.nc = nc  # number of classes
    self.nl = len(ch)  # number of detection layers
    self.reg_max = 16  # DFL channels (ch[0] // 16 to scale 4/8/12/16/20 for n/s/m/l/x)
    self.no = nc + self.reg_max * 4  # number of outputs per anchor
    self.stride = torch.zeros(self.nl)  # strides computed during build
    c2, c3 = max((16, ch[0] // 4, self.reg_max * 4)), max(ch[0], min(self.nc, 100))  # channels
    self.cv2 = nn.ModuleList(
        nn.Sequential(Conv(x, c2, 3), Conv(c2, c2, 3), nn.Conv2d(c2, 4 * self.reg_max, 1)) for x in ch
    )
    self.cv3 = (
        nn.ModuleList(nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, self.nc, 1)) for x in ch)
        if self.legacy
        else nn.ModuleList(
            nn.Sequential(
                nn.Sequential(DWConv(x, x, 3), Conv(x, c3, 1)),
                nn.Sequential(DWConv(c3, c3, 3), Conv(c3, c3, 1)),
                nn.Conv2d(c3, self.nc, 1),
            )
            for x in ch
        )
    )
    self.dfl = DFL(self.reg_max) if self.reg_max > 1 else nn.Identity()

    if self.end2end:
        self.one2one_cv2 = copy.deepcopy(self.cv2)
        self.one2one_cv3 = copy.deepcopy(self.cv3)

bias_init

bias_init()

Initialize Detect() biases, WARNING: requires stride availability.

Source code in ultralytics/nn/modules/head.py
146
147
148
149
150
151
152
153
154
155
156
157
def bias_init(self):
    """Initialize Detect() biases, WARNING: requires stride availability."""
    m = self  # self.model[-1]  # Detect() module
    # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
    # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
    for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
        a[-1].bias.data[:] = 1.0  # box
        b[-1].bias.data[: m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (.01 objects, 80 classes, 640 img)
    if self.end2end:
        for a, b, s in zip(m.one2one_cv2, m.one2one_cv3, m.stride):  # from
            a[-1].bias.data[:] = 1.0  # box
            b[-1].bias.data[: m.nc] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (.01 objects, 80 classes, 640 img)

decode_bboxes

decode_bboxes(bboxes, anchors, xywh=True)

Decode bounding boxes.

Source code in ultralytics/nn/modules/head.py
159
160
161
def decode_bboxes(self, bboxes, anchors, xywh=True):
    """Decode bounding boxes."""
    return dist2bbox(bboxes, anchors, xywh=xywh and not (self.end2end or self.xyxy), dim=1)

forward

forward(x)

Concatenates and returns predicted bounding boxes and class probabilities.

Source code in ultralytics/nn/modules/head.py
67
68
69
70
71
72
73
74
75
76
77
def forward(self, x):
    """Concatenates and returns predicted bounding boxes and class probabilities."""
    if self.end2end:
        return self.forward_end2end(x)

    for i in range(self.nl):
        x[i] = torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1)
    if self.training:  # Training path
        return x
    y = self._inference(x)
    return y if self.export else (y, x)

forward_end2end

forward_end2end(x)

Performs forward pass of the v10Detect module.

Parameters:

Name Type Description Default
x List[Tensor]

Input feature maps from different levels.

required

Returns:

Type Description
dict | tuple
  • If in training mode, returns a dictionary containing outputs of both one2many and one2one detections.
  • If not in training mode, returns processed detections or a tuple with processed detections and raw outputs.
Source code in ultralytics/nn/modules/head.py
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
def forward_end2end(self, x):
    """
    Performs forward pass of the v10Detect module.

    Args:
        x (List[torch.Tensor]): Input feature maps from different levels.

    Returns:
        (dict | tuple):

            - If in training mode, returns a dictionary containing outputs of both one2many and one2one detections.
            - If not in training mode, returns processed detections or a tuple with processed detections and raw outputs.
    """
    x_detach = [xi.detach() for xi in x]
    one2one = [
        torch.cat((self.one2one_cv2[i](x_detach[i]), self.one2one_cv3[i](x_detach[i])), 1) for i in range(self.nl)
    ]
    for i in range(self.nl):
        x[i] = torch.cat((self.cv2[i](x[i]), self.cv3[i](x[i])), 1)
    if self.training:  # Training path
        return {"one2many": x, "one2one": one2one}

    y = self._inference(one2one)
    y = self.postprocess(y.permute(0, 2, 1), self.max_det, self.nc)
    return y if self.export else (y, {"one2many": x, "one2one": one2one})

postprocess staticmethod

postprocess(preds: Tensor, max_det: int, nc: int = 80)

Post-processes YOLO model predictions.

Parameters:

Name Type Description Default
preds Tensor

Raw predictions with shape (batch_size, num_anchors, 4 + nc) with last dimension format [x, y, w, h, class_probs].

required
max_det int

Maximum detections per image.

required
nc int

Number of classes. Default: 80.

80

Returns:

Type Description
Tensor

Processed predictions with shape (batch_size, min(max_det, num_anchors), 6) and last dimension format [x, y, w, h, max_class_prob, class_index].

Source code in ultralytics/nn/modules/head.py
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
@staticmethod
def postprocess(preds: torch.Tensor, max_det: int, nc: int = 80):
    """
    Post-processes YOLO model predictions.

    Args:
        preds (torch.Tensor): Raw predictions with shape (batch_size, num_anchors, 4 + nc) with last dimension
            format [x, y, w, h, class_probs].
        max_det (int): Maximum detections per image.
        nc (int, optional): Number of classes. Default: 80.

    Returns:
        (torch.Tensor): Processed predictions with shape (batch_size, min(max_det, num_anchors), 6) and last
            dimension format [x, y, w, h, max_class_prob, class_index].
    """
    batch_size, anchors, _ = preds.shape  # i.e. shape(16,8400,84)
    boxes, scores = preds.split([4, nc], dim=-1)
    index = scores.amax(dim=-1).topk(min(max_det, anchors))[1].unsqueeze(-1)
    boxes = boxes.gather(dim=1, index=index.repeat(1, 1, 4))
    scores = scores.gather(dim=1, index=index.repeat(1, 1, nc))
    scores, index = scores.flatten(1).topk(min(max_det, anchors))
    i = torch.arange(batch_size)[..., None]  # batch indices
    return torch.cat([boxes[i, index // nc], scores[..., None], (index % nc)[..., None].float()], dim=-1)





ultralytics.nn.modules.head.Segment

Segment(nc=80, nm=32, npr=256, ch=())

Bases: Detect

YOLO Segment head for segmentation models.

Source code in ultralytics/nn/modules/head.py
191
192
193
194
195
196
197
198
199
def __init__(self, nc=80, nm=32, npr=256, ch=()):
    """Initialize the YOLO model attributes such as the number of masks, prototypes, and the convolution layers."""
    super().__init__(nc, ch)
    self.nm = nm  # number of masks
    self.npr = npr  # number of protos
    self.proto = Proto(ch[0], self.npr, self.nm)  # protos

    c4 = max(ch[0] // 4, self.nm)
    self.cv4 = nn.ModuleList(nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, self.nm, 1)) for x in ch)

forward

forward(x)

Return model outputs and mask coefficients if training, otherwise return outputs and mask coefficients.

Source code in ultralytics/nn/modules/head.py
201
202
203
204
205
206
207
208
209
210
def forward(self, x):
    """Return model outputs and mask coefficients if training, otherwise return outputs and mask coefficients."""
    p = self.proto(x[0])  # mask protos
    bs = p.shape[0]  # batch size

    mc = torch.cat([self.cv4[i](x[i]).view(bs, self.nm, -1) for i in range(self.nl)], 2)  # mask coefficients
    x = Detect.forward(self, x)
    if self.training:
        return x, mc, p
    return (torch.cat([x, mc], 1), p) if self.export else (torch.cat([x[0], mc], 1), (x[1], mc, p))





ultralytics.nn.modules.head.OBB

OBB(nc=80, ne=1, ch=())

Bases: Detect

YOLO OBB detection head for detection with rotation models.

Source code in ultralytics/nn/modules/head.py
216
217
218
219
220
221
222
def __init__(self, nc=80, ne=1, ch=()):
    """Initialize OBB with number of classes `nc` and layer channels `ch`."""
    super().__init__(nc, ch)
    self.ne = ne  # number of extra parameters

    c4 = max(ch[0] // 4, self.ne)
    self.cv4 = nn.ModuleList(nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, self.ne, 1)) for x in ch)

decode_bboxes

decode_bboxes(bboxes, anchors)

Decode rotated bounding boxes.

Source code in ultralytics/nn/modules/head.py
238
239
240
def decode_bboxes(self, bboxes, anchors):
    """Decode rotated bounding boxes."""
    return dist2rbox(bboxes, self.angle, anchors, dim=1)

forward

forward(x)

Concatenates and returns predicted bounding boxes and class probabilities.

Source code in ultralytics/nn/modules/head.py
224
225
226
227
228
229
230
231
232
233
234
235
236
def forward(self, x):
    """Concatenates and returns predicted bounding boxes and class probabilities."""
    bs = x[0].shape[0]  # batch size
    angle = torch.cat([self.cv4[i](x[i]).view(bs, self.ne, -1) for i in range(self.nl)], 2)  # OBB theta logits
    # NOTE: set `angle` as an attribute so that `decode_bboxes` could use it.
    angle = (angle.sigmoid() - 0.25) * math.pi  # [-pi/4, 3pi/4]
    # angle = angle.sigmoid() * math.pi / 2  # [0, pi/2]
    if not self.training:
        self.angle = angle
    x = Detect.forward(self, x)
    if self.training:
        return x, angle
    return torch.cat([x, angle], 1) if self.export else (torch.cat([x[0], angle], 1), (x[1], angle))





ultralytics.nn.modules.head.Pose

Pose(nc=80, kpt_shape=(17, 3), ch=())

Bases: Detect

YOLO Pose head for keypoints models.

Source code in ultralytics/nn/modules/head.py
246
247
248
249
250
251
252
253
def __init__(self, nc=80, kpt_shape=(17, 3), ch=()):
    """Initialize YOLO network with default parameters and Convolutional Layers."""
    super().__init__(nc, ch)
    self.kpt_shape = kpt_shape  # number of keypoints, number of dims (2 for x,y or 3 for x,y,visible)
    self.nk = kpt_shape[0] * kpt_shape[1]  # number of keypoints total

    c4 = max(ch[0] // 4, self.nk)
    self.cv4 = nn.ModuleList(nn.Sequential(Conv(x, c4, 3), Conv(c4, c4, 3), nn.Conv2d(c4, self.nk, 1)) for x in ch)

forward

forward(x)

Perform forward pass through YOLO model and return predictions.

Source code in ultralytics/nn/modules/head.py
255
256
257
258
259
260
261
262
263
def forward(self, x):
    """Perform forward pass through YOLO model and return predictions."""
    bs = x[0].shape[0]  # batch size
    kpt = torch.cat([self.cv4[i](x[i]).view(bs, self.nk, -1) for i in range(self.nl)], -1)  # (bs, 17*3, h*w)
    x = Detect.forward(self, x)
    if self.training:
        return x, kpt
    pred_kpt = self.kpts_decode(bs, kpt)
    return torch.cat([x, pred_kpt], 1) if self.export else (torch.cat([x[0], pred_kpt], 1), (x[1], kpt))

kpts_decode

kpts_decode(bs, kpts)

Decodes keypoints.

Source code in ultralytics/nn/modules/head.py
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
def kpts_decode(self, bs, kpts):
    """Decodes keypoints."""
    ndim = self.kpt_shape[1]
    if self.export:
        if self.format in {
            "tflite",
            "edgetpu",
        }:  # required for TFLite export to avoid 'PLACEHOLDER_FOR_GREATER_OP_CODES' bug
            # Precompute normalization factor to increase numerical stability
            y = kpts.view(bs, *self.kpt_shape, -1)
            grid_h, grid_w = self.shape[2], self.shape[3]
            grid_size = torch.tensor([grid_w, grid_h], device=y.device).reshape(1, 2, 1)
            norm = self.strides / (self.stride[0] * grid_size)
            a = (y[:, :, :2] * 2.0 + (self.anchors - 0.5)) * norm
        else:
            # NCNN fix
            y = kpts.view(bs, *self.kpt_shape, -1)
            a = (y[:, :, :2] * 2.0 + (self.anchors - 0.5)) * self.strides
        if ndim == 3:
            a = torch.cat((a, y[:, :, 2:3].sigmoid()), 2)
        return a.view(bs, self.nk, -1)
    else:
        y = kpts.clone()
        if ndim == 3:
            y[:, 2::ndim] = y[:, 2::ndim].sigmoid()  # sigmoid (WARNING: inplace .sigmoid_() Apple MPS bug)
        y[:, 0::ndim] = (y[:, 0::ndim] * 2.0 + (self.anchors[0] - 0.5)) * self.strides
        y[:, 1::ndim] = (y[:, 1::ndim] * 2.0 + (self.anchors[1] - 0.5)) * self.strides
        return y





ultralytics.nn.modules.head.Classify

Classify(c1, c2, k=1, s=1, p=None, g=1)

Bases: Module

YOLO classification head, i.e. x(b,c1,20,20) to x(b,c2).

Source code in ultralytics/nn/modules/head.py
300
301
302
303
304
305
306
307
def __init__(self, c1, c2, k=1, s=1, p=None, g=1):
    """Initializes YOLO classification head to transform input tensor from (b,c1,20,20) to (b,c2) shape."""
    super().__init__()
    c_ = 1280  # efficientnet_b0 size
    self.conv = Conv(c1, c_, k, s, p, g)
    self.pool = nn.AdaptiveAvgPool2d(1)  # to x(b,c_,1,1)
    self.drop = nn.Dropout(p=0.0, inplace=True)
    self.linear = nn.Linear(c_, c2)  # to x(b,c2)

forward

forward(x)

Performs a forward pass of the YOLO model on input image data.

Source code in ultralytics/nn/modules/head.py
309
310
311
312
313
314
315
316
317
def forward(self, x):
    """Performs a forward pass of the YOLO model on input image data."""
    if isinstance(x, list):
        x = torch.cat(x, 1)
    x = self.linear(self.drop(self.pool(self.conv(x)).flatten(1)))
    if self.training:
        return x
    y = x.softmax(1)  # get final output
    return y if self.export else (y, x)





ultralytics.nn.modules.head.WorldDetect

WorldDetect(nc=80, embed=512, with_bn=False, ch=())

Bases: Detect

Head for integrating YOLO detection models with semantic understanding from text embeddings.

Source code in ultralytics/nn/modules/head.py
323
324
325
326
327
328
def __init__(self, nc=80, embed=512, with_bn=False, ch=()):
    """Initialize YOLO detection layer with nc classes and layer channels ch."""
    super().__init__(nc, ch)
    c3 = max(ch[0], min(self.nc, 100))
    self.cv3 = nn.ModuleList(nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, embed, 1)) for x in ch)
    self.cv4 = nn.ModuleList(BNContrastiveHead(embed) if with_bn else ContrastiveHead() for _ in ch)

bias_init

bias_init()

Initialize Detect() biases, WARNING: requires stride availability.

Source code in ultralytics/nn/modules/head.py
340
341
342
343
344
345
346
def bias_init(self):
    """Initialize Detect() biases, WARNING: requires stride availability."""
    m = self  # self.model[-1]  # Detect() module
    # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
    # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
    for a, b, s in zip(m.cv2, m.cv3, m.stride):  # from
        a[-1].bias.data[:] = 1.0  # box

forward

forward(x, text)

Concatenates and returns predicted bounding boxes and class probabilities.

Source code in ultralytics/nn/modules/head.py
330
331
332
333
334
335
336
337
338
def forward(self, x, text):
    """Concatenates and returns predicted bounding boxes and class probabilities."""
    for i in range(self.nl):
        x[i] = torch.cat((self.cv2[i](x[i]), self.cv4[i](self.cv3[i](x[i]), text)), 1)
    if self.training:
        return x
    self.no = self.nc + self.reg_max * 4  # self.nc could be changed when inference with different texts
    y = self._inference(x)
    return y if self.export else (y, x)





ultralytics.nn.modules.head.LRPCHead

LRPCHead(vocab, pf, loc, enabled=True)

Bases: Module

Lightweight Region Proposal and Classification Head for efficient object detection.

Source code in ultralytics/nn/modules/head.py
353
354
355
356
357
358
359
def __init__(self, vocab, pf, loc, enabled=True):
    """Initialize LRPCHead with vocabulary, proposal filter, and localization components."""
    super().__init__()
    self.vocab = self.conv2linear(vocab) if enabled else vocab
    self.pf = pf
    self.loc = loc
    self.enabled = enabled

conv2linear

conv2linear(conv)

Convert a 1x1 convolutional layer to a linear layer.

Source code in ultralytics/nn/modules/head.py
361
362
363
364
365
366
367
def conv2linear(self, conv):
    """Convert a 1x1 convolutional layer to a linear layer."""
    assert isinstance(conv, nn.Conv2d) and conv.kernel_size == (1, 1)
    linear = nn.Linear(conv.in_channels, conv.out_channels)
    linear.weight.data = conv.weight.view(conv.out_channels, -1).data
    linear.bias.data = conv.bias.data
    return linear

forward

forward(cls_feat, loc_feat, conf)

Process classification and localization features to generate detection proposals.

Source code in ultralytics/nn/modules/head.py
369
370
371
372
373
374
375
376
377
378
379
380
381
382
def forward(self, cls_feat, loc_feat, conf):
    """Process classification and localization features to generate detection proposals."""
    if self.enabled:
        pf_score = self.pf(cls_feat)[0, 0].flatten(0)
        mask = pf_score.sigmoid() > conf
        cls_feat = cls_feat.flatten(2).transpose(-1, -2)
        cls_feat = self.vocab(cls_feat[:, mask] if conf else cls_feat * mask.unsqueeze(-1).int())
        return (self.loc(loc_feat), cls_feat.transpose(-1, -2)), mask
    else:
        cls_feat = self.vocab(cls_feat)
        loc_feat = self.loc(loc_feat)
        return (loc_feat, cls_feat.flatten(2)), torch.ones(
            cls_feat.shape[2] * cls_feat.shape[3], device=cls_feat.device, dtype=torch.bool
        )





ultralytics.nn.modules.head.YOLOEDetect

YOLOEDetect(nc=80, embed=512, with_bn=False, ch=())

Bases: Detect

Head for integrating YOLO detection models with semantic understanding from text embeddings.

Source code in ultralytics/nn/modules/head.py
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
def __init__(self, nc=80, embed=512, with_bn=False, ch=()):
    """Initialize YOLO detection layer with nc classes and layer channels ch."""
    super().__init__(nc, ch)
    c3 = max(ch[0], min(self.nc, 100))
    assert c3 <= embed
    assert with_bn is True
    self.cv3 = (
        nn.ModuleList(nn.Sequential(Conv(x, c3, 3), Conv(c3, c3, 3), nn.Conv2d(c3, embed, 1)) for x in ch)
        if self.legacy
        else nn.ModuleList(
            nn.Sequential(
                nn.Sequential(DWConv(x, x, 3), Conv(x, c3, 1)),
                nn.Sequential(DWConv(c3, c3, 3), Conv(c3, c3, 1)),
                nn.Conv2d(c3, embed, 1),
            )
            for x in ch
        )
    )

    self.cv4 = nn.ModuleList(BNContrastiveHead(embed) if with_bn else ContrastiveHead() for _ in ch)

    self.reprta = Residual(SwiGLUFFN(embed, embed))
    self.savpe = SAVPE(ch, c3, embed)
    self.embed = embed

bias_init

bias_init()

Initialize biases for detection heads.

Source code in ultralytics/nn/modules/head.py
525
526
527
528
529
530
531
532
533
534
def bias_init(self):
    """Initialize biases for detection heads."""
    m = self  # self.model[-1]  # Detect() module
    # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1
    # ncf = math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # nominal class frequency
    for a, b, c, s in zip(m.cv2, m.cv3, m.cv4, m.stride):  # from
        a[-1].bias.data[:] = 1.0  # box
        # b[-1].bias.data[:] = math.log(5 / m.nc / (640 / s) ** 2)  # cls (.01 objects, 80 classes, 640 img)
        b[-1].bias.data[:] = 0.0
        c.bias.data[:] = math.log(5 / m.nc / (640 / s) ** 2)

forward

forward(x, cls_pe, return_mask=False)

Process features with class prompt embeddings to generate detections.

Source code in ultralytics/nn/modules/head.py
513
514
515
516
517
518
519
520
521
522
523
def forward(self, x, cls_pe, return_mask=False):
    """Process features with class prompt embeddings to generate detections."""
    if hasattr(self, "lrpc"):  # for prompt-free inference
        return self.forward_lrpc(x, return_mask)
    for i in range(self.nl):
        x[i] = torch.cat((self.cv2[i](x[i]), self.cv4[i](self.cv3[i](x[i]), cls_pe)), 1)
    if self.training:
        return x
    self.no = self.nc + self.reg_max * 4  # self.nc could be changed when inference with different texts
    y = self._inference(x)
    return y if self.export else (y, x)

forward_lrpc

forward_lrpc(x, return_mask=False)

Process features with fused text embeddings to generate detections for prompt-free model.

Source code in ultralytics/nn/modules/head.py
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
def forward_lrpc(self, x, return_mask=False):
    """Process features with fused text embeddings to generate detections for prompt-free model."""
    masks = []
    assert self.is_fused, "Prompt-free inference requires model to be fused!"
    for i in range(self.nl):
        cls_feat = self.cv3[i](x[i])
        loc_feat = self.cv2[i](x[i])
        assert isinstance(self.lrpc[i], LRPCHead)
        x[i], mask = self.lrpc[i](
            cls_feat, loc_feat, 0 if self.export and not self.dynamic else getattr(self, "conf", 0.001)
        )
        masks.append(mask)
    shape = x[0][0].shape
    if self.dynamic or self.shape != shape:
        self.anchors, self.strides = (x.transpose(0, 1) for x in make_anchors([b[0] for b in x], self.stride, 0.5))
        self.shape = shape
    box = torch.cat([xi[0].view(shape[0], self.reg_max * 4, -1) for xi in x], 2)
    cls = torch.cat([xi[1] for xi in x], 2)

    if self.export and self.format in {"tflite", "edgetpu"}:
        # Precompute normalization factor to increase numerical stability
        # See https://github.com/ultralytics/ultralytics/issues/7371
        grid_h = shape[2]
        grid_w = shape[3]
        grid_size = torch.tensor([grid_w, grid_h, grid_w, grid_h], device=box.device).reshape(1, 4, 1)
        norm = self.strides / (self.stride[0] * grid_size)
        dbox = self.decode_bboxes(self.dfl(box) * norm, self.anchors.unsqueeze(0) * norm[:, :2])
    else:
        dbox = self.decode_bboxes(self.dfl(box), self.anchors.unsqueeze(0)) * self.strides

    mask = torch.cat(masks)
    y = torch.cat((dbox if self.export and not self.dynamic else dbox[..., mask], cls.sigmoid()), 1)

    if return_mask:
        return (y, mask) if self.export else ((y, x), mask)
    else:
        return y if self.export else (y, x)

fuse

fuse(txt_feats)

Fuse text features with model weights for efficient inference.

Source code in ultralytics/nn/modules/head.py
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
@smart_inference_mode()
def fuse(self, txt_feats):
    """Fuse text features with model weights for efficient inference."""
    if self.is_fused:
        return

    assert not self.training
    txt_feats = txt_feats.to(torch.float32).squeeze(0)
    for cls_head, bn_head in zip(self.cv3, self.cv4):
        assert isinstance(cls_head, nn.Sequential)
        assert isinstance(bn_head, BNContrastiveHead)
        conv = cls_head[-1]
        assert isinstance(conv, nn.Conv2d)
        logit_scale = bn_head.logit_scale
        bias = bn_head.bias
        norm = bn_head.norm

        t = txt_feats * logit_scale.exp()
        conv: nn.Conv2d = fuse_conv_and_bn(conv, norm)

        w = conv.weight.data.squeeze(-1).squeeze(-1)
        b = conv.bias.data

        w = t @ w
        b1 = (t @ b.reshape(-1).unsqueeze(-1)).squeeze(-1)
        b2 = torch.ones_like(b1) * bias

        conv = (
            nn.Conv2d(
                conv.in_channels,
                w.shape[0],
                kernel_size=1,
            )
            .requires_grad_(False)
            .to(conv.weight.device)
        )

        conv.weight.data.copy_(w.unsqueeze(-1).unsqueeze(-1))
        conv.bias.data.copy_(b1 + b2)
        cls_head[-1] = conv

        bn_head.fuse()

    del self.reprta
    self.reprta = nn.Identity()
    self.is_fused = True

get_tpe

get_tpe(tpe)

Get text prompt embeddings with normalization.

Source code in ultralytics/nn/modules/head.py
462
463
464
def get_tpe(self, tpe):
    """Get text prompt embeddings with normalization."""
    return None if tpe is None else F.normalize(self.reprta(tpe), dim=-1, p=2)

get_vpe

get_vpe(x, vpe)

Get visual prompt embeddings with spatial awareness.

Source code in ultralytics/nn/modules/head.py
466
467
468
469
470
471
472
473
def get_vpe(self, x, vpe):
    """Get visual prompt embeddings with spatial awareness."""
    if vpe.shape[1] == 0:  # no visual prompt embeddings
        return torch.zeros(x[0].shape[0], 0, self.embed, device=x[0].device)
    if vpe.ndim == 4:  # (B, N, H, W)
        vpe = self.savpe(x, vpe)
    assert vpe.ndim == 3  # (B, N, D)
    return vpe





ultralytics.nn.modules.head.YOLOESegment

YOLOESegment(nc=80, nm=32, npr=256, embed=512, with_bn=False, ch=())

Bases: YOLOEDetect

YOLO segmentation head with text embedding capabilities.

Source code in ultralytics/nn/modules/head.py
540
541
542
543
544
545
546
547
548
def __init__(self, nc=80, nm=32, npr=256, embed=512, with_bn=False, ch=()):
    """Initialize YOLOESegment with class count, mask parameters, and embedding dimensions."""
    super().__init__(nc, embed, with_bn, ch)
    self.nm = nm
    self.npr = npr
    self.proto = Proto(ch[0], self.npr, self.nm)

    c5 = max(ch[0] // 4, self.nm)
    self.cv5 = nn.ModuleList(nn.Sequential(Conv(x, c5, 3), Conv(c5, c5, 3), nn.Conv2d(c5, self.nm, 1)) for x in ch)

forward

forward(x, text)

Return model outputs and mask coefficients if training, otherwise return outputs and mask coefficients.

Source code in ultralytics/nn/modules/head.py
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
def forward(self, x, text):
    """Return model outputs and mask coefficients if training, otherwise return outputs and mask coefficients."""
    p = self.proto(x[0])  # mask protos
    bs = p.shape[0]  # batch size

    mc = torch.cat([self.cv5[i](x[i]).view(bs, self.nm, -1) for i in range(self.nl)], 2)  # mask coefficients
    has_lrpc = hasattr(self, "lrpc")

    if not has_lrpc:
        x = YOLOEDetect.forward(self, x, text)
    else:
        x, mask = YOLOEDetect.forward(self, x, text, return_mask=True)

    if self.training:
        return x, mc, p

    if has_lrpc:
        mc = (mc * mask.int()) if self.export and not self.dynamic else mc[..., mask]

    return (torch.cat([x, mc], 1), p) if self.export else (torch.cat([x[0], mc], 1), (x[1], mc, p))





ultralytics.nn.modules.head.RTDETRDecoder

RTDETRDecoder(
    nc=80,
    ch=(512, 1024, 2048),
    hd=256,
    nq=300,
    ndp=4,
    nh=8,
    ndl=6,
    d_ffn=1024,
    dropout=0.0,
    act=nn.ReLU(),
    eval_idx=-1,
    nd=100,
    label_noise_ratio=0.5,
    box_noise_scale=1.0,
    learnt_init_query=False,
)

Bases: Module

Real-Time Deformable Transformer Decoder (RTDETRDecoder) module for object detection.

This decoder module utilizes Transformer architecture along with deformable convolutions to predict bounding boxes and class labels for objects in an image. It integrates features from multiple layers and runs through a series of Transformer decoder layers to output the final predictions.

Parameters:

Name Type Description Default
nc int

Number of classes. Default is 80.

80
ch tuple

Channels in the backbone feature maps. Default is (512, 1024, 2048).

(512, 1024, 2048)
hd int

Dimension of hidden layers. Default is 256.

256
nq int

Number of query points. Default is 300.

300
ndp int

Number of decoder points. Default is 4.

4
nh int

Number of heads in multi-head attention. Default is 8.

8
ndl int

Number of decoder layers. Default is 6.

6
d_ffn int

Dimension of the feed-forward networks. Default is 1024.

1024
dropout float

Dropout rate. Default is 0.0.

0.0
act Module

Activation function. Default is nn.ReLU.

ReLU()
eval_idx int

Evaluation index. Default is -1.

-1
nd int

Number of denoising. Default is 100.

100
label_noise_ratio float

Label noise ratio. Default is 0.5.

0.5
box_noise_scale float

Box noise scale. Default is 1.0.

1.0
learnt_init_query bool

Whether to learn initial query embeddings. Default is False.

False
Source code in ultralytics/nn/modules/head.py
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
def __init__(
    self,
    nc=80,
    ch=(512, 1024, 2048),
    hd=256,  # hidden dim
    nq=300,  # num queries
    ndp=4,  # num decoder points
    nh=8,  # num head
    ndl=6,  # num decoder layers
    d_ffn=1024,  # dim of feedforward
    dropout=0.0,
    act=nn.ReLU(),
    eval_idx=-1,
    # Training args
    nd=100,  # num denoising
    label_noise_ratio=0.5,
    box_noise_scale=1.0,
    learnt_init_query=False,
):
    """
    Initializes the RTDETRDecoder module with the given parameters.

    Args:
        nc (int): Number of classes. Default is 80.
        ch (tuple): Channels in the backbone feature maps. Default is (512, 1024, 2048).
        hd (int): Dimension of hidden layers. Default is 256.
        nq (int): Number of query points. Default is 300.
        ndp (int): Number of decoder points. Default is 4.
        nh (int): Number of heads in multi-head attention. Default is 8.
        ndl (int): Number of decoder layers. Default is 6.
        d_ffn (int): Dimension of the feed-forward networks. Default is 1024.
        dropout (float): Dropout rate. Default is 0.0.
        act (nn.Module): Activation function. Default is nn.ReLU.
        eval_idx (int): Evaluation index. Default is -1.
        nd (int): Number of denoising. Default is 100.
        label_noise_ratio (float): Label noise ratio. Default is 0.5.
        box_noise_scale (float): Box noise scale. Default is 1.0.
        learnt_init_query (bool): Whether to learn initial query embeddings. Default is False.
    """
    super().__init__()
    self.hidden_dim = hd
    self.nhead = nh
    self.nl = len(ch)  # num level
    self.nc = nc
    self.num_queries = nq
    self.num_decoder_layers = ndl

    # Backbone feature projection
    self.input_proj = nn.ModuleList(nn.Sequential(nn.Conv2d(x, hd, 1, bias=False), nn.BatchNorm2d(hd)) for x in ch)
    # NOTE: simplified version but it's not consistent with .pt weights.
    # self.input_proj = nn.ModuleList(Conv(x, hd, act=False) for x in ch)

    # Transformer module
    decoder_layer = DeformableTransformerDecoderLayer(hd, nh, d_ffn, dropout, act, self.nl, ndp)
    self.decoder = DeformableTransformerDecoder(hd, decoder_layer, ndl, eval_idx)

    # Denoising part
    self.denoising_class_embed = nn.Embedding(nc, hd)
    self.num_denoising = nd
    self.label_noise_ratio = label_noise_ratio
    self.box_noise_scale = box_noise_scale

    # Decoder embedding
    self.learnt_init_query = learnt_init_query
    if learnt_init_query:
        self.tgt_embed = nn.Embedding(nq, hd)
    self.query_pos_head = MLP(4, 2 * hd, hd, num_layers=2)

    # Encoder head
    self.enc_output = nn.Sequential(nn.Linear(hd, hd), nn.LayerNorm(hd))
    self.enc_score_head = nn.Linear(hd, nc)
    self.enc_bbox_head = MLP(hd, hd, 4, num_layers=3)

    # Decoder head
    self.dec_score_head = nn.ModuleList([nn.Linear(hd, nc) for _ in range(ndl)])
    self.dec_bbox_head = nn.ModuleList([MLP(hd, hd, 4, num_layers=3) for _ in range(ndl)])

    self._reset_parameters()

forward

forward(x, batch=None)

Runs the forward pass of the module, returning bounding box and classification scores for the input.

Parameters:

Name Type Description Default
x List[Tensor]

List of feature maps from the backbone.

required
batch dict

Batch information for training.

None

Returns:

Type Description
tuple | Tensor

During training, returns a tuple of bounding boxes, scores, and other metadata. During inference, returns a tensor of shape (bs, 300, 4+nc) containing bounding boxes and class scores.

Source code in ultralytics/nn/modules/head.py
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
def forward(self, x, batch=None):
    """
    Runs the forward pass of the module, returning bounding box and classification scores for the input.

    Args:
        x (List[torch.Tensor]): List of feature maps from the backbone.
        batch (dict, optional): Batch information for training.

    Returns:
        (tuple | torch.Tensor): During training, returns a tuple of bounding boxes, scores, and other metadata.
            During inference, returns a tensor of shape (bs, 300, 4+nc) containing bounding boxes and class scores.
    """
    from ultralytics.models.utils.ops import get_cdn_group

    # Input projection and embedding
    feats, shapes = self._get_encoder_input(x)

    # Prepare denoising training
    dn_embed, dn_bbox, attn_mask, dn_meta = get_cdn_group(
        batch,
        self.nc,
        self.num_queries,
        self.denoising_class_embed.weight,
        self.num_denoising,
        self.label_noise_ratio,
        self.box_noise_scale,
        self.training,
    )

    embed, refer_bbox, enc_bboxes, enc_scores = self._get_decoder_input(feats, shapes, dn_embed, dn_bbox)

    # Decoder
    dec_bboxes, dec_scores = self.decoder(
        embed,
        refer_bbox,
        feats,
        shapes,
        self.dec_bbox_head,
        self.dec_score_head,
        self.query_pos_head,
        attn_mask=attn_mask,
    )
    x = dec_bboxes, dec_scores, enc_bboxes, enc_scores, dn_meta
    if self.training:
        return x
    # (bs, 300, 4+nc)
    y = torch.cat((dec_bboxes.squeeze(0), dec_scores.squeeze(0).sigmoid()), -1)
    return y if self.export else (y, x)





ultralytics.nn.modules.head.v10Detect

v10Detect(nc=80, ch=())

Bases: Detect

v10 Detection head from https://arxiv.org/pdf/2405.14458.

Parameters:

Name Type Description Default
nc int

Number of classes.

80
ch tuple

Tuple of channel sizes.

()

Attributes:

Name Type Description
max_det int

Maximum number of detections.

Methods:

Name Description
forward

Performs forward pass of the v10Detect module.

bias_init

Initializes biases of the Detect module.

Source code in ultralytics/nn/modules/head.py
863
864
865
866
867
868
869
870
871
872
873
874
875
876
def __init__(self, nc=80, ch=()):
    """Initializes the v10Detect object with the specified number of classes and input channels."""
    super().__init__(nc, ch)
    c3 = max(ch[0], min(self.nc, 100))  # channels
    # Light cls head
    self.cv3 = nn.ModuleList(
        nn.Sequential(
            nn.Sequential(Conv(x, x, 3, g=x), Conv(x, c3, 1)),
            nn.Sequential(Conv(c3, c3, 3, g=c3), Conv(c3, c3, 1)),
            nn.Conv2d(c3, self.nc, 1),
        )
        for x in ch
    )
    self.one2one_cv3 = copy.deepcopy(self.cv3)

fuse

fuse()

Removes the one2many head.

Source code in ultralytics/nn/modules/head.py
878
879
880
def fuse(self):
    """Removes the one2many head."""
    self.cv2 = self.cv3 = nn.ModuleList([nn.Identity()] * self.nl)





📅 Created 1 year ago ✏️ Updated 1 month ago