Add Conv2dNormActivation and Conv3dNormActivation Blocks (#5445)

* Add ops.conv3d

* Refactor for conv2d and 3d

* Refactor

* Fix bug

* Addres review

* Fix bug

* nit fix

* Fix flake

* Final fix

* remove documentation

* fix linter

* Update all the implementations to use new Conv

* Small doc fix

Co-authored-by: Vasilis Vryniotis <[email protected]>
Co-authored-by: Joao Gomes <[email protected]>

Author: 3 people

docs/source/ops.rst

@@ -45,4 +45,6 @@ Operators
     FeaturePyramidNetwork
     StochasticDepth
     FrozenBatchNorm2d
+    Conv2dNormActivation
+    Conv3dNormActivation
     SqueezeExcitation

torchvision/models/convnext.py

@@ -6,7 +6,7 @@
 from torch.nn import functional as F
 
 from .._internally_replaced_utils import load_state_dict_from_url
-from ..ops.misc import ConvNormActivation
+from ..ops.misc import Conv2dNormActivation
 from ..ops.stochastic_depth import StochasticDepth
 from ..utils import _log_api_usage_once
 
@@ -127,7 +127,7 @@ def __init__(
         # Stem
         firstconv_output_channels = block_setting[0].input_channels
         layers.append(
-            ConvNormActivation(
+            Conv2dNormActivation(
                 3,
                 firstconv_output_channels,
                 kernel_size=4,

torchvision/models/detection/ssdlite.py

@@ -7,7 +7,7 @@
 from torch import nn, Tensor
 
 from ..._internally_replaced_utils import load_state_dict_from_url
-from ...ops.misc import ConvNormActivation
+from ...ops.misc import Conv2dNormActivation
 from ...utils import _log_api_usage_once
 from .. import mobilenet
 from . import _utils as det_utils
@@ -29,7 +29,7 @@ def _prediction_block(
 ) -> nn.Sequential:
     return nn.Sequential(
         # 3x3 depthwise with stride 1 and padding 1
-        ConvNormActivation(
+        Conv2dNormActivation(
             in_channels,
             in_channels,
             kernel_size=kernel_size,
@@ -47,11 +47,11 @@ def _extra_block(in_channels: int, out_channels: int, norm_layer: Callable[...,
     intermediate_channels = out_channels // 2
     return nn.Sequential(
         # 1x1 projection to half output channels
-        ConvNormActivation(
+        Conv2dNormActivation(
             in_channels, intermediate_channels, kernel_size=1, norm_layer=norm_layer, activation_layer=activation
         ),
         # 3x3 depthwise with stride 2 and padding 1
-        ConvNormActivation(
+        Conv2dNormActivation(
             intermediate_channels,
             intermediate_channels,
             kernel_size=3,
@@ -61,7 +61,7 @@ def _extra_block(in_channels: int, out_channels: int, norm_layer: Callable[...,
             activation_layer=activation,
         ),
         # 1x1 projetion to output channels
-        ConvNormActivation(
+        Conv2dNormActivation(
             intermediate_channels, out_channels, kernel_size=1, norm_layer=norm_layer, activation_layer=activation
         ),
     )

torchvision/models/efficientnet.py

@@ -8,7 +8,7 @@
 from torchvision.ops import StochasticDepth
 
 from .._internally_replaced_utils import load_state_dict_from_url
-from ..ops.misc import ConvNormActivation, SqueezeExcitation
+from ..ops.misc import Conv2dNormActivation, SqueezeExcitation
 from ..utils import _log_api_usage_once
 from ._utils import _make_divisible
 
@@ -104,7 +104,7 @@ def __init__(
         expanded_channels = cnf.adjust_channels(cnf.input_channels, cnf.expand_ratio)
         if expanded_channels != cnf.input_channels:
             layers.append(
-                ConvNormActivation(
+                Conv2dNormActivation(
                     cnf.input_channels,
                     expanded_channels,
                     kernel_size=1,
@@ -115,7 +115,7 @@ def __init__(
 
         # depthwise
         layers.append(
-            ConvNormActivation(
+            Conv2dNormActivation(
                 expanded_channels,
                 expanded_channels,
                 kernel_size=cnf.kernel,
@@ -132,7 +132,7 @@ def __init__(
 
         # project
         layers.append(
-            ConvNormActivation(
+            Conv2dNormActivation(
                 expanded_channels, cnf.out_channels, kernel_size=1, norm_layer=norm_layer, activation_layer=None
             )
         )
@@ -193,7 +193,7 @@ def __init__(
         # building first layer
         firstconv_output_channels = inverted_residual_setting[0].input_channels
         layers.append(
-            ConvNormActivation(
+            Conv2dNormActivation(
                 3, firstconv_output_channels, kernel_size=3, stride=2, norm_layer=norm_layer, activation_layer=nn.SiLU
             )
         )
@@ -224,7 +224,7 @@ def __init__(
         lastconv_input_channels = inverted_residual_setting[-1].out_channels
         lastconv_output_channels = 4 * lastconv_input_channels
         layers.append(
-            ConvNormActivation(
+            Conv2dNormActivation(
                 lastconv_input_channels,
                 lastconv_output_channels,
                 kernel_size=1,

torchvision/models/mobilenetv2.py

@@ -6,7 +6,7 @@
 from torch import nn
 
 from .._internally_replaced_utils import load_state_dict_from_url
-from ..ops.misc import ConvNormActivation
+from ..ops.misc import Conv2dNormActivation
 from ..utils import _log_api_usage_once
 from ._utils import _make_divisible
 
@@ -20,11 +20,11 @@
 
 
 # necessary for backwards compatibility
-class _DeprecatedConvBNAct(ConvNormActivation):
+class _DeprecatedConvBNAct(Conv2dNormActivation):
     def __init__(self, *args, **kwargs):
         warnings.warn(
             "The ConvBNReLU/ConvBNActivation classes are deprecated since 0.12 and will be removed in 0.14. "
-            "Use torchvision.ops.misc.ConvNormActivation instead.",
+            "Use torchvision.ops.misc.Conv2dNormActivation instead.",
             FutureWarning,
         )
         if kwargs.get("norm_layer", None) is None:
@@ -56,12 +56,12 @@ def __init__(
         if expand_ratio != 1:
             # pw
             layers.append(
-                ConvNormActivation(inp, hidden_dim, kernel_size=1, norm_layer=norm_layer, activation_layer=nn.ReLU6)
+                Conv2dNormActivation(inp, hidden_dim, kernel_size=1, norm_layer=norm_layer, activation_layer=nn.ReLU6)
             )
         layers.extend(
             [
                 # dw
-                ConvNormActivation(
+                Conv2dNormActivation(
                     hidden_dim,
                     hidden_dim,
                     stride=stride,
@@ -144,7 +144,7 @@ def __init__(
         input_channel = _make_divisible(input_channel * width_mult, round_nearest)
         self.last_channel = _make_divisible(last_channel * max(1.0, width_mult), round_nearest)
         features: List[nn.Module] = [
-            ConvNormActivation(3, input_channel, stride=2, norm_layer=norm_layer, activation_layer=nn.ReLU6)
+            Conv2dNormActivation(3, input_channel, stride=2, norm_layer=norm_layer, activation_layer=nn.ReLU6)
         ]
         # building inverted residual blocks
         for t, c, n, s in inverted_residual_setting:
@@ -155,7 +155,7 @@ def __init__(
                 input_channel = output_channel
         # building last several layers
         features.append(
-            ConvNormActivation(
+            Conv2dNormActivation(
                 input_channel, self.last_channel, kernel_size=1, norm_layer=norm_layer, activation_layer=nn.ReLU6
             )
         )

torchvision/models/mobilenetv3.py

@@ -6,7 +6,7 @@
 from torch import nn, Tensor
 
 from .._internally_replaced_utils import load_state_dict_from_url
-from ..ops.misc import ConvNormActivation, SqueezeExcitation as SElayer
+from ..ops.misc import Conv2dNormActivation, SqueezeExcitation as SElayer
 from ..utils import _log_api_usage_once
 from ._utils import _make_divisible
 
@@ -83,7 +83,7 @@ def __init__(
         # expand
         if cnf.expanded_channels != cnf.input_channels:
             layers.append(
-                ConvNormActivation(
+                Conv2dNormActivation(
                     cnf.input_channels,
                     cnf.expanded_channels,
                     kernel_size=1,
@@ -95,7 +95,7 @@ def __init__(
         # depthwise
         stride = 1 if cnf.dilation > 1 else cnf.stride
         layers.append(
-            ConvNormActivation(
+            Conv2dNormActivation(
                 cnf.expanded_channels,
                 cnf.expanded_channels,
                 kernel_size=cnf.kernel,
@@ -112,7 +112,7 @@ def __init__(
 
         # project
         layers.append(
-            ConvNormActivation(
+            Conv2dNormActivation(
                 cnf.expanded_channels, cnf.out_channels, kernel_size=1, norm_layer=norm_layer, activation_layer=None
             )
         )
@@ -172,7 +172,7 @@ def __init__(
         # building first layer
         firstconv_output_channels = inverted_residual_setting[0].input_channels
         layers.append(
-            ConvNormActivation(
+            Conv2dNormActivation(
                 3,
                 firstconv_output_channels,
                 kernel_size=3,
@@ -190,7 +190,7 @@ def __init__(
         lastconv_input_channels = inverted_residual_setting[-1].out_channels
         lastconv_output_channels = 6 * lastconv_input_channels
         layers.append(
-            ConvNormActivation(
+            Conv2dNormActivation(
                 lastconv_input_channels,
                 lastconv_output_channels,
                 kernel_size=1,

torchvision/models/optical_flow/raft.py

@@ -6,7 +6,7 @@
 from torch import Tensor
 from torch.nn.modules.batchnorm import BatchNorm2d
 from torch.nn.modules.instancenorm import InstanceNorm2d
-from torchvision.ops.misc import ConvNormActivation
+from torchvision.ops import Conv2dNormActivation
 
 from ..._internally_replaced_utils import load_state_dict_from_url
 from ...utils import _log_api_usage_once
@@ -38,17 +38,17 @@ def __init__(self, in_channels, out_channels, *, norm_layer, stride=1):
         # and frozen for the rest of the training process (i.e. set as eval()). The bias term is thus still useful
         # for the rest of the datasets. Technically, we could remove the bias for other norm layers like Instance norm
         # because these aren't frozen, but we don't bother (also, we woudn't be able to load the original weights).
-        self.convnormrelu1 = ConvNormActivation(
+        self.convnormrelu1 = Conv2dNormActivation(
             in_channels, out_channels, norm_layer=norm_layer, kernel_size=3, stride=stride, bias=True
         )
-        self.convnormrelu2 = ConvNormActivation(
+        self.convnormrelu2 = Conv2dNormActivation(
             out_channels, out_channels, norm_layer=norm_layer, kernel_size=3, bias=True
         )
 
         if stride == 1:
             self.downsample = nn.Identity()
         else:
-            self.downsample = ConvNormActivation(
+            self.downsample = Conv2dNormActivation(
                 in_channels,
                 out_channels,
                 norm_layer=norm_layer,
@@ -77,21 +77,21 @@ def __init__(self, in_channels, out_channels, *, norm_layer, stride=1):
         super().__init__()
 
         # See note in ResidualBlock for the reason behind bias=True
-        self.convnormrelu1 = ConvNormActivation(
+        self.convnormrelu1 = Conv2dNormActivation(
             in_channels, out_channels // 4, norm_layer=norm_layer, kernel_size=1, bias=True
         )
-        self.convnormrelu2 = ConvNormActivation(
+        self.convnormrelu2 = Conv2dNormActivation(
             out_channels // 4, out_channels // 4, norm_layer=norm_layer, kernel_size=3, stride=stride, bias=True
         )
-        self.convnormrelu3 = ConvNormActivation(
+        self.convnormrelu3 = Conv2dNormActivation(
             out_channels // 4, out_channels, norm_layer=norm_layer, kernel_size=1, bias=True
         )
         self.relu = nn.ReLU(inplace=True)
 
         if stride == 1:
             self.downsample = nn.Identity()
         else:
-            self.downsample = ConvNormActivation(
+            self.downsample = Conv2dNormActivation(
                 in_channels,
                 out_channels,
                 norm_layer=norm_layer,
@@ -124,7 +124,9 @@ def __init__(self, *, block=ResidualBlock, layers=(64, 64, 96, 128, 256), norm_l
         assert len(layers) == 5
 
         # See note in ResidualBlock for the reason behind bias=True
-        self.convnormrelu = ConvNormActivation(3, layers[0], norm_layer=norm_layer, kernel_size=7, stride=2, bias=True)
+        self.convnormrelu = Conv2dNormActivation(
+            3, layers[0], norm_layer=norm_layer, kernel_size=7, stride=2, bias=True
+        )
 
         self.layer1 = self._make_2_blocks(block, layers[0], layers[1], norm_layer=norm_layer, first_stride=1)
         self.layer2 = self._make_2_blocks(block, layers[1], layers[2], norm_layer=norm_layer, first_stride=2)
@@ -170,17 +172,17 @@ def __init__(self, *, in_channels_corr, corr_layers=(256, 192), flow_layers=(128
         assert len(flow_layers) == 2
         assert len(corr_layers) in (1, 2)
 
-        self.convcorr1 = ConvNormActivation(in_channels_corr, corr_layers[0], norm_layer=None, kernel_size=1)
+        self.convcorr1 = Conv2dNormActivation(in_channels_corr, corr_layers[0], norm_layer=None, kernel_size=1)
         if len(corr_layers) == 2:
-            self.convcorr2 = ConvNormActivation(corr_layers[0], corr_layers[1], norm_layer=None, kernel_size=3)
+            self.convcorr2 = Conv2dNormActivation(corr_layers[0], corr_layers[1], norm_layer=None, kernel_size=3)
         else:
             self.convcorr2 = nn.Identity()
 
-        self.convflow1 = ConvNormActivation(2, flow_layers[0], norm_layer=None, kernel_size=7)
-        self.convflow2 = ConvNormActivation(flow_layers[0], flow_layers[1], norm_layer=None, kernel_size=3)
+        self.convflow1 = Conv2dNormActivation(2, flow_layers[0], norm_layer=None, kernel_size=7)
+        self.convflow2 = Conv2dNormActivation(flow_layers[0], flow_layers[1], norm_layer=None, kernel_size=3)
 
         # out_channels - 2 because we cat the flow (2 channels) at the end
-        self.conv = ConvNormActivation(
+        self.conv = Conv2dNormActivation(
             corr_layers[-1] + flow_layers[-1], out_channels - 2, norm_layer=None, kernel_size=3
         )
 
@@ -301,7 +303,7 @@ class MaskPredictor(nn.Module):
 
     def __init__(self, *, in_channels, hidden_size, multiplier=0.25):
         super().__init__()
-        self.convrelu = ConvNormActivation(in_channels, hidden_size, norm_layer=None, kernel_size=3)
+        self.convrelu = Conv2dNormActivation(in_channels, hidden_size, norm_layer=None, kernel_size=3)
         # 8 * 8 * 9 because the predicted flow is downsampled by 8, from the downsampling of the initial FeatureEncoder
         # and we interpolate with all 9 surrounding neighbors. See paper and appendix B.
         self.conv = nn.Conv2d(hidden_size, 8 * 8 * 9, 1, padding=0)

torchvision/models/quantization/mobilenetv2.py

@@ -6,7 +6,7 @@
 from torchvision.models.mobilenetv2 import InvertedResidual, MobileNetV2, model_urls
 
 from ..._internally_replaced_utils import load_state_dict_from_url
-from ...ops.misc import ConvNormActivation
+from ...ops.misc import Conv2dNormActivation
 from .utils import _fuse_modules, _replace_relu, quantize_model
 
 
@@ -54,7 +54,7 @@ def forward(self, x: Tensor) -> Tensor:
 
     def fuse_model(self, is_qat: Optional[bool] = None) -> None:
         for m in self.modules():
-            if type(m) is ConvNormActivation:
+            if type(m) is Conv2dNormActivation:
                 _fuse_modules(m, ["0", "1", "2"], is_qat, inplace=True)
             if type(m) is QuantizableInvertedResidual:
                 m.fuse_model(is_qat)

torchvision/models/quantization/mobilenetv3.py

@@ -5,7 +5,7 @@
 from torch.ao.quantization import QuantStub, DeQuantStub
 
 from ..._internally_replaced_utils import load_state_dict_from_url
-from ...ops.misc import ConvNormActivation, SqueezeExcitation
+from ...ops.misc import Conv2dNormActivation, SqueezeExcitation
 from ..mobilenetv3 import InvertedResidual, InvertedResidualConfig, MobileNetV3, model_urls, _mobilenet_v3_conf
 from .utils import _fuse_modules, _replace_relu
 
@@ -103,7 +103,7 @@ def forward(self, x: Tensor) -> Tensor:
 
     def fuse_model(self, is_qat: Optional[bool] = None) -> None:
         for m in self.modules():
-            if type(m) is ConvNormActivation:
+            if type(m) is Conv2dNormActivation:
                 modules_to_fuse = ["0", "1"]
                 if len(m) == 3 and type(m[2]) is nn.ReLU:
                     modules_to_fuse.append("2")