Skip to content

Qarepvgg block

QARepVGGBlock

Bases: nn.Module

QARepVGG (S3/S4) block from 'Make RepVGG Greater Again: A Quantization-aware Approach' (https://arxiv.org/pdf/2212.01593.pdf) It consists of three branches:

3x3: a branch of a 3x3 Convolution + BatchNorm 1x1: a branch of a 1x1 Convolution with bias identity: a Residual branch which will only be used if input channel == output channel and use_residual_connection is True (usually in all but the first block of each stage)

BatchNorm is applied after summation of all three branches. In contrast to our implementation of RepVGGBlock, SE is applied AFTER NONLINEARITY in order to fuse Conv+Act in inference frameworks.

This module converts to Conv+Act in a PTQ-friendly way by calling QARepVGGBlock.fuse_block_residual_branches(). Has the same API as RepVGGBlock and is designed to be a plug-and-play replacement but is not compatible parameter-wise. Has less trainable parameters than RepVGGBlock because it has only 2 BatchNorms instead of 3.

                |
                |
|---------------|---------------|
|               |               |

3x3 1x1 | | | | BatchNorm +bias | | | | | *alpha | | | | |---------------+---------------| | BatchNorm | Act | SE

Source code in src/super_gradients/modules/qarepvgg_block.py
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 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
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 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
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
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
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
class QARepVGGBlock(nn.Module):
    """
    QARepVGG (S3/S4) block from 'Make RepVGG Greater Again: A Quantization-aware Approach' (https://arxiv.org/pdf/2212.01593.pdf)
    It consists of three branches:

    3x3: a branch of a 3x3 Convolution + BatchNorm
    1x1: a branch of a 1x1 Convolution with bias
    identity: a Residual branch which will only be used if input channel == output channel and use_residual_connection is True
        (usually in all but the first block of each stage)

    BatchNorm is applied after summation of all three branches.
    In contrast to our implementation of RepVGGBlock, SE is applied AFTER NONLINEARITY in order to fuse Conv+Act in inference frameworks.

    This module converts to Conv+Act in a PTQ-friendly way by calling QARepVGGBlock.fuse_block_residual_branches().
    Has the same API as RepVGGBlock and is designed to be a plug-and-play replacement but is not compatible parameter-wise.
    Has less trainable parameters than RepVGGBlock because it has only 2 BatchNorms instead of 3.


                        |
                        |
        |---------------|---------------|
        |               |               |
       3x3             1x1              |
        |               |               |
    BatchNorm         +bias             |
        |               |               |
        |             *alpha            |
        |               |               |
        |---------------+---------------|
                        |
                    BatchNorm
                        |
                       Act
                        |
                       SE
    """

    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        stride: int = 1,
        dilation: int = 1,
        groups: int = 1,
        activation_type: Type[nn.Module] = nn.ReLU,
        activation_kwargs: Union[Mapping[str, Any], None] = None,
        se_type: Type[nn.Module] = nn.Identity,
        se_kwargs: Union[Mapping[str, Any], None] = None,
        build_residual_branches: bool = True,
        use_residual_connection: bool = True,
        use_alpha: bool = False,
        use_1x1_bias: bool = True,
        use_post_bn: bool = True,
    ):
        """
        :param in_channels: Number of input channels
        :param out_channels: Number of output channels
        :param activation_type: Type of the nonlinearity (nn.ReLU by default)
        :param se_type: Type of the se block (Use nn.Identity to disable SE)
        :param stride: Output stride
        :param dilation: Dilation factor for 3x3 conv
        :param groups: Number of groups used in convolutions
        :param activation_kwargs: Additional arguments for instantiating activation module.
        :param se_kwargs: Additional arguments for instantiating SE module.
        :param build_residual_branches: Whether to initialize block with already fused parameters (for deployment)
        :param use_residual_connection: Whether to add input x to the output (Enabled in RepVGG, disabled in PP-Yolo)
        :param use_alpha: If True, enables additional learnable weighting parameter for 1x1 branch (PP-Yolo-E Plus)
        :param use_1x1_bias: If True, enables bias in the 1x1 convolution, authors don't mention it specifically
        :param use_post_bn: If True, adds BatchNorm after the sum of three branches (S4), if False, BatchNorm is not added (S3)
        """
        super().__init__()

        if activation_kwargs is None:
            activation_kwargs = {}
        if se_kwargs is None:
            se_kwargs = {}

        self.groups = groups
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.stride = stride
        self.dilation = dilation
        self.activation_type = activation_type
        self.activation_kwargs = activation_kwargs
        self.se_type = se_type
        self.se_kwargs = se_kwargs
        self.use_residual_connection = use_residual_connection
        self.use_alpha = use_alpha
        self.use_1x1_bias = use_1x1_bias
        self.use_post_bn = use_post_bn

        self.nonlinearity = activation_type(**activation_kwargs)
        self.se = se_type(**se_kwargs)

        self.branch_3x3 = nn.Sequential()
        self.branch_3x3.add_module(
            "conv",
            nn.Conv2d(
                in_channels=in_channels,
                out_channels=out_channels,
                kernel_size=3,
                stride=stride,
                padding=dilation,
                groups=groups,
                bias=False,
                dilation=dilation,
            ),
        )
        self.branch_3x3.add_module("bn", nn.BatchNorm2d(num_features=out_channels))

        self.branch_1x1 = nn.Conv2d(
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=1,
            stride=stride,
            padding=0,
            groups=groups,
            bias=use_1x1_bias,
        )

        if use_residual_connection:
            assert out_channels == in_channels and stride == 1

            self.identity = Residual()

            input_dim = self.in_channels // self.groups
            id_tensor = torch.zeros((self.in_channels, input_dim, 3, 3))
            for i in range(self.in_channels):
                id_tensor[i, i % input_dim, 1, 1] = 1.0

            self.id_tensor: Optional[torch.Tensor]
            self.register_buffer(
                name="id_tensor",
                tensor=id_tensor.to(dtype=self.branch_1x1.weight.dtype, device=self.branch_1x1.weight.device),
                persistent=False,  # so it's not saved in state_dict
            )
        else:
            self.identity = None

        if use_alpha:
            self.alpha = torch.nn.Parameter(torch.tensor([1.0]), requires_grad=True)
        else:
            self.alpha = 1.0

        if self.use_post_bn:
            self.post_bn = nn.BatchNorm2d(num_features=out_channels)
        else:
            self.post_bn = nn.Identity()

        # placeholder to correctly register parameters
        self.rbr_reparam = nn.Conv2d(
            in_channels=self.branch_3x3.conv.in_channels,
            out_channels=self.branch_3x3.conv.out_channels,
            kernel_size=self.branch_3x3.conv.kernel_size,
            stride=self.branch_3x3.conv.stride,
            padding=self.branch_3x3.conv.padding,
            dilation=self.branch_3x3.conv.dilation,
            groups=self.branch_3x3.conv.groups,
            bias=True,
        )

        self.partially_fused = False
        self.fully_fused = False

        if not build_residual_branches:
            self.fuse_block_residual_branches()

    def forward(self, inputs):
        if self.fully_fused:
            return self.se(self.nonlinearity(self.rbr_reparam(inputs)))

        if self.partially_fused:
            return self.se(self.nonlinearity(self.post_bn(self.rbr_reparam(inputs))))

        if self.identity is None:
            id_out = 0.0
        else:
            id_out = self.identity(inputs)

        x_3x3 = self.branch_3x3(inputs)
        x_1x1 = self.alpha * self.branch_1x1(inputs)

        branches = x_3x3 + x_1x1 + id_out

        out = self.nonlinearity(self.post_bn(branches))
        se = self.se(out)

        return se

    def _get_equivalent_kernel_bias_for_branches(self):
        """
        Fuses the 3x3, 1x1 and identity branches into a single 3x3 conv layer
        """
        kernel3x3, bias3x3 = self._fuse_bn_tensor(
            self.branch_3x3.conv.weight,
            0,
            self.branch_3x3.bn.running_mean,
            self.branch_3x3.bn.running_var,
            self.branch_3x3.bn.weight,
            self.branch_3x3.bn.bias,
            self.branch_3x3.bn.eps,
        )

        kernel1x1 = self._pad_1x1_to_3x3_tensor(self.branch_1x1.weight)
        bias1x1 = self.branch_1x1.bias if self.branch_1x1.bias is not None else 0

        kernelid = self.id_tensor if self.identity is not None else 0
        biasid = 0

        eq_kernel_3x3 = kernel3x3 + self.alpha * kernel1x1 + kernelid
        eq_bias_3x3 = bias3x3 + self.alpha * bias1x1 + biasid

        return eq_kernel_3x3, eq_bias_3x3

    def _pad_1x1_to_3x3_tensor(self, kernel1x1):
        """
        padding the 1x1 convolution weights with zeros to be able to fuse the 3x3 conv layer with the 1x1
        :param kernel1x1: weights of the 1x1 convolution
        :type kernel1x1:
        :return: padded 1x1 weights
        :rtype:
        """
        if kernel1x1 is None:
            return 0
        else:
            return torch.nn.functional.pad(kernel1x1, [1, 1, 1, 1])

    def _fuse_bn_tensor(self, kernel, bias, running_mean, running_var, gamma, beta, eps):
        std = torch.sqrt(running_var + eps)
        b = beta - gamma * running_mean / std
        A = gamma / std
        A_ = A.expand_as(kernel.transpose(0, -1)).transpose(0, -1)

        fused_kernel = kernel * A_
        fused_bias = bias * A + b

        return fused_kernel, fused_bias

    def full_fusion(self):
        """Fuse everything into Conv-Act-SE, non-trainable, parameters detached
        converts a qarepvgg block from training model (with branches) to deployment mode (vgg like model)
        :return:
        :rtype:
        """
        if self.fully_fused:
            return

        if not self.partially_fused:
            self.partial_fusion()

        if self.use_post_bn:
            eq_kernel, eq_bias = self._fuse_bn_tensor(
                self.rbr_reparam.weight,
                self.rbr_reparam.bias,
                self.post_bn.running_mean,
                self.post_bn.running_var,
                self.post_bn.weight,
                self.post_bn.bias,
                self.post_bn.eps,
            )

            self.rbr_reparam.weight.data = eq_kernel
            self.rbr_reparam.bias.data = eq_bias

        for para in self.parameters():
            para.detach_()

        if hasattr(self, "post_bn"):
            self.__delattr__("post_bn")

        self.partially_fused = False
        self.fully_fused = True

    def partial_fusion(self):
        """Fuse branches into a single kernel, leave post_bn unfused, leave parameters differentiable"""
        if self.partially_fused:
            return

        if self.fully_fused:
            # TODO: we actually can, all we need to do is insert the properly initialized post_bn back
            # init is not trivial, so not implemented for now
            raise NotImplementedError("QARepVGGBlock can't be converted to partially fused from fully fused")

        kernel, bias = self._get_equivalent_kernel_bias_for_branches()

        self.rbr_reparam.weight.data = kernel
        self.rbr_reparam.bias.data = bias

        self.__delattr__("branch_3x3")
        self.__delattr__("branch_1x1")
        if hasattr(self, "identity"):
            self.__delattr__("identity")
        if hasattr(self, "alpha"):
            self.__delattr__("alpha")
        if hasattr(self, "id_tensor"):
            self.__delattr__("id_tensor")

        self.partially_fused = True
        self.fully_fused = False

    def fuse_block_residual_branches(self):
        # inference frameworks will take care of resulting conv-bn-act-se
        # no need to fuse post_bn prematurely if it is there
        # call self.full_fusion() if you need it
        self.partial_fusion()

    def from_repvgg(self, src: RepVGGBlock):
        raise NotImplementedError

    def prep_model_for_conversion(self, input_size: Optional[Union[tuple, list]] = None, full_fusion: bool = True, **kwargs):
        """Prepare the QARepVGGBlock for conversion.

        :WARNING: the default `full_fusion=True` will make the block non-trainable.

        :param full_fusion: If True, performs full fusion, converting the block into a non-trainable, fully fused block.
                            If False, performs partial fusion, slower for inference but still trainable.
        """

        if full_fusion:
            self.full_fusion()
        else:
            self.partial_fusion()

__init__(in_channels, out_channels, stride=1, dilation=1, groups=1, activation_type=nn.ReLU, activation_kwargs=None, se_type=nn.Identity, se_kwargs=None, build_residual_branches=True, use_residual_connection=True, use_alpha=False, use_1x1_bias=True, use_post_bn=True)

Parameters:

Name Type Description Default
in_channels int

Number of input channels

required
out_channels int

Number of output channels

required
activation_type Type[nn.Module]

Type of the nonlinearity (nn.ReLU by default)

nn.ReLU
se_type Type[nn.Module]

Type of the se block (Use nn.Identity to disable SE)

nn.Identity
stride int

Output stride

1
dilation int

Dilation factor for 3x3 conv

1
groups int

Number of groups used in convolutions

1
activation_kwargs Union[Mapping[str, Any], None]

Additional arguments for instantiating activation module.

None
se_kwargs Union[Mapping[str, Any], None]

Additional arguments for instantiating SE module.

None
build_residual_branches bool

Whether to initialize block with already fused parameters (for deployment)

True
use_residual_connection bool

Whether to add input x to the output (Enabled in RepVGG, disabled in PP-Yolo)

True
use_alpha bool

If True, enables additional learnable weighting parameter for 1x1 branch (PP-Yolo-E Plus)

False
use_1x1_bias bool

If True, enables bias in the 1x1 convolution, authors don't mention it specifically

True
use_post_bn bool

If True, adds BatchNorm after the sum of three branches (S4), if False, BatchNorm is not added (S3)

True
Source code in src/super_gradients/modules/qarepvgg_block.py
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 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
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
def __init__(
    self,
    in_channels: int,
    out_channels: int,
    stride: int = 1,
    dilation: int = 1,
    groups: int = 1,
    activation_type: Type[nn.Module] = nn.ReLU,
    activation_kwargs: Union[Mapping[str, Any], None] = None,
    se_type: Type[nn.Module] = nn.Identity,
    se_kwargs: Union[Mapping[str, Any], None] = None,
    build_residual_branches: bool = True,
    use_residual_connection: bool = True,
    use_alpha: bool = False,
    use_1x1_bias: bool = True,
    use_post_bn: bool = True,
):
    """
    :param in_channels: Number of input channels
    :param out_channels: Number of output channels
    :param activation_type: Type of the nonlinearity (nn.ReLU by default)
    :param se_type: Type of the se block (Use nn.Identity to disable SE)
    :param stride: Output stride
    :param dilation: Dilation factor for 3x3 conv
    :param groups: Number of groups used in convolutions
    :param activation_kwargs: Additional arguments for instantiating activation module.
    :param se_kwargs: Additional arguments for instantiating SE module.
    :param build_residual_branches: Whether to initialize block with already fused parameters (for deployment)
    :param use_residual_connection: Whether to add input x to the output (Enabled in RepVGG, disabled in PP-Yolo)
    :param use_alpha: If True, enables additional learnable weighting parameter for 1x1 branch (PP-Yolo-E Plus)
    :param use_1x1_bias: If True, enables bias in the 1x1 convolution, authors don't mention it specifically
    :param use_post_bn: If True, adds BatchNorm after the sum of three branches (S4), if False, BatchNorm is not added (S3)
    """
    super().__init__()

    if activation_kwargs is None:
        activation_kwargs = {}
    if se_kwargs is None:
        se_kwargs = {}

    self.groups = groups
    self.in_channels = in_channels
    self.out_channels = out_channels
    self.stride = stride
    self.dilation = dilation
    self.activation_type = activation_type
    self.activation_kwargs = activation_kwargs
    self.se_type = se_type
    self.se_kwargs = se_kwargs
    self.use_residual_connection = use_residual_connection
    self.use_alpha = use_alpha
    self.use_1x1_bias = use_1x1_bias
    self.use_post_bn = use_post_bn

    self.nonlinearity = activation_type(**activation_kwargs)
    self.se = se_type(**se_kwargs)

    self.branch_3x3 = nn.Sequential()
    self.branch_3x3.add_module(
        "conv",
        nn.Conv2d(
            in_channels=in_channels,
            out_channels=out_channels,
            kernel_size=3,
            stride=stride,
            padding=dilation,
            groups=groups,
            bias=False,
            dilation=dilation,
        ),
    )
    self.branch_3x3.add_module("bn", nn.BatchNorm2d(num_features=out_channels))

    self.branch_1x1 = nn.Conv2d(
        in_channels=in_channels,
        out_channels=out_channels,
        kernel_size=1,
        stride=stride,
        padding=0,
        groups=groups,
        bias=use_1x1_bias,
    )

    if use_residual_connection:
        assert out_channels == in_channels and stride == 1

        self.identity = Residual()

        input_dim = self.in_channels // self.groups
        id_tensor = torch.zeros((self.in_channels, input_dim, 3, 3))
        for i in range(self.in_channels):
            id_tensor[i, i % input_dim, 1, 1] = 1.0

        self.id_tensor: Optional[torch.Tensor]
        self.register_buffer(
            name="id_tensor",
            tensor=id_tensor.to(dtype=self.branch_1x1.weight.dtype, device=self.branch_1x1.weight.device),
            persistent=False,  # so it's not saved in state_dict
        )
    else:
        self.identity = None

    if use_alpha:
        self.alpha = torch.nn.Parameter(torch.tensor([1.0]), requires_grad=True)
    else:
        self.alpha = 1.0

    if self.use_post_bn:
        self.post_bn = nn.BatchNorm2d(num_features=out_channels)
    else:
        self.post_bn = nn.Identity()

    # placeholder to correctly register parameters
    self.rbr_reparam = nn.Conv2d(
        in_channels=self.branch_3x3.conv.in_channels,
        out_channels=self.branch_3x3.conv.out_channels,
        kernel_size=self.branch_3x3.conv.kernel_size,
        stride=self.branch_3x3.conv.stride,
        padding=self.branch_3x3.conv.padding,
        dilation=self.branch_3x3.conv.dilation,
        groups=self.branch_3x3.conv.groups,
        bias=True,
    )

    self.partially_fused = False
    self.fully_fused = False

    if not build_residual_branches:
        self.fuse_block_residual_branches()

full_fusion()

Fuse everything into Conv-Act-SE, non-trainable, parameters detached converts a qarepvgg block from training model (with branches) to deployment mode (vgg like model)

Returns:

Type Description
Source code in src/super_gradients/modules/qarepvgg_block.py
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
def full_fusion(self):
    """Fuse everything into Conv-Act-SE, non-trainable, parameters detached
    converts a qarepvgg block from training model (with branches) to deployment mode (vgg like model)
    :return:
    :rtype:
    """
    if self.fully_fused:
        return

    if not self.partially_fused:
        self.partial_fusion()

    if self.use_post_bn:
        eq_kernel, eq_bias = self._fuse_bn_tensor(
            self.rbr_reparam.weight,
            self.rbr_reparam.bias,
            self.post_bn.running_mean,
            self.post_bn.running_var,
            self.post_bn.weight,
            self.post_bn.bias,
            self.post_bn.eps,
        )

        self.rbr_reparam.weight.data = eq_kernel
        self.rbr_reparam.bias.data = eq_bias

    for para in self.parameters():
        para.detach_()

    if hasattr(self, "post_bn"):
        self.__delattr__("post_bn")

    self.partially_fused = False
    self.fully_fused = True

partial_fusion()

Fuse branches into a single kernel, leave post_bn unfused, leave parameters differentiable

Source code in src/super_gradients/modules/qarepvgg_block.py
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
def partial_fusion(self):
    """Fuse branches into a single kernel, leave post_bn unfused, leave parameters differentiable"""
    if self.partially_fused:
        return

    if self.fully_fused:
        # TODO: we actually can, all we need to do is insert the properly initialized post_bn back
        # init is not trivial, so not implemented for now
        raise NotImplementedError("QARepVGGBlock can't be converted to partially fused from fully fused")

    kernel, bias = self._get_equivalent_kernel_bias_for_branches()

    self.rbr_reparam.weight.data = kernel
    self.rbr_reparam.bias.data = bias

    self.__delattr__("branch_3x3")
    self.__delattr__("branch_1x1")
    if hasattr(self, "identity"):
        self.__delattr__("identity")
    if hasattr(self, "alpha"):
        self.__delattr__("alpha")
    if hasattr(self, "id_tensor"):
        self.__delattr__("id_tensor")

    self.partially_fused = True
    self.fully_fused = False

prep_model_for_conversion(input_size=None, full_fusion=True, **kwargs)

Prepare the QARepVGGBlock for conversion.

:WARNING: the default full_fusion=True will make the block non-trainable.

Parameters:

Name Type Description Default
full_fusion bool

If True, performs full fusion, converting the block into a non-trainable, fully fused block. If False, performs partial fusion, slower for inference but still trainable.

True
Source code in src/super_gradients/modules/qarepvgg_block.py
319
320
321
322
323
324
325
326
327
328
329
330
331
def prep_model_for_conversion(self, input_size: Optional[Union[tuple, list]] = None, full_fusion: bool = True, **kwargs):
    """Prepare the QARepVGGBlock for conversion.

    :WARNING: the default `full_fusion=True` will make the block non-trainable.

    :param full_fusion: If True, performs full fusion, converting the block into a non-trainable, fully fused block.
                        If False, performs partial fusion, slower for inference but still trainable.
    """

    if full_fusion:
        self.full_fusion()
    else:
        self.partial_fusion()