RT-ODLab/models/detectors/yolov7/yolov7_basic.py

import numpy as np
import torch
import torch.nn as nn


# ---------------------------- 2D CNN ----------------------------
class SiLU(nn.Module):
    """export-friendly version of nn.SiLU()"""

    @staticmethod
    def forward(x):
        return x * torch.sigmoid(x)

def get_conv2d(c1, c2, k, p, s, d, g, bias=False):
    conv = nn.Conv2d(c1, c2, k, stride=s, padding=p, dilation=d, 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)

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)

## Basic conv layer
class Conv(nn.Module):
    def __init__(self, 
                 c1,                   # in channels
                 c2,                   # out channels 
                 k=1,                  # kernel size 
                 p=0,                  # padding
                 s=1,                  # padding
                 d=1,                  # dilation
                 act_type='lrelu',     # activation
                 norm_type='BN',       # normalization
                 depthwise=False):
        super(Conv, self).__init__()
        convs = []
        add_bias = False if norm_type else True
        if depthwise:
            convs.append(get_conv2d(c1, c1, k=k, p=p, s=s, d=d, g=c1, bias=add_bias))
            # depthwise conv
            if norm_type:
                convs.append(get_norm(norm_type, c1))
            if act_type:
                convs.append(get_activation(act_type))
            # pointwise conv
            convs.append(get_conv2d(c1, c2, k=1, p=0, s=1, d=d, g=1, bias=add_bias))
            if norm_type:
                convs.append(get_norm(norm_type, c2))
            if act_type:
                convs.append(get_activation(act_type))

        else:
            convs.append(get_conv2d(c1, c2, k=k, p=p, s=s, d=d, g=1, bias=add_bias))
            if norm_type:
                convs.append(get_norm(norm_type, c2))
            if act_type:
                convs.append(get_activation(act_type))
            
        self.convs = nn.Sequential(*convs)


    def forward(self, x):
        return self.convs(x)


# ---------------------------- YOLOv7 Modules ----------------------------
## ELAN-Block proposed by YOLOv7
class ELANBlock(nn.Module):
    def __init__(self, in_dim, out_dim, squeeze_ratio=0.5, branch_depth :int=2, act_type='silu', norm_type='BN', depthwise=False):
        super(ELANBlock, self).__init__()
        inter_dim = int(in_dim * squeeze_ratio)
        self.cv1 = Conv(in_dim, inter_dim, k=1, act_type=act_type, norm_type=norm_type)
        self.cv2 = Conv(in_dim, inter_dim, k=1, act_type=act_type, norm_type=norm_type)
        self.cv3 = nn.Sequential(*[
            Conv(inter_dim, inter_dim, k=3, p=1, act_type=act_type, norm_type=norm_type, depthwise=depthwise)
            for _ in range(round(branch_depth))
        ])
        self.cv4 = nn.Sequential(*[
            Conv(inter_dim, inter_dim, k=3, p=1, act_type=act_type, norm_type=norm_type, depthwise=depthwise)
            for _ in range(round(branch_depth))
        ])

        self.out = Conv(inter_dim*4, out_dim, k=1, act_type=act_type, norm_type=norm_type)



    def forward(self, x):
        x1 = self.cv1(x)
        x2 = self.cv2(x)
        x3 = self.cv3(x2)
        x4 = self.cv4(x3)
        out = self.out(torch.cat([x1, x2, x3, x4], dim=1))

        return out

## PaFPN's ELAN-Block proposed by YOLOv7
class ELANBlockFPN(nn.Module):
    def __init__(self, in_dim, out_dim, squeeze_ratio=0.5, branch_width :int=4, branch_depth :int=1, act_type='silu', norm_type='BN', depthwise=False):
        super(ELANBlockFPN, self).__init__()
        # Basic parameters
        inter_dim = int(in_dim * squeeze_ratio)
        inter_dim2 = int(inter_dim * squeeze_ratio) 
        # Network structure
        self.cv1 = Conv(in_dim, inter_dim, k=1, act_type=act_type, norm_type=norm_type)
        self.cv2 = Conv(in_dim, inter_dim, k=1, act_type=act_type, norm_type=norm_type)
        self.cv3 = nn.ModuleList()
        for idx in range(round(branch_width)):
            if idx == 0:
                cvs = [Conv(inter_dim, inter_dim2, k=3, p=1, act_type=act_type, norm_type=norm_type, depthwise=depthwise)]
            else:
                cvs = [Conv(inter_dim2, inter_dim2, k=3, p=1, act_type=act_type, norm_type=norm_type, depthwise=depthwise)]
            # deeper
            if round(branch_depth) > 1:
                for _ in range(1, round(branch_depth)):
                    cvs.append(Conv(inter_dim2, inter_dim2, k=3, p=1, act_type=act_type, norm_type=norm_type, depthwise=depthwise))
                self.cv3.append(nn.Sequential(*cvs))
            else:
                self.cv3.append(cvs[0])

        self.out = Conv(inter_dim*2+inter_dim2*len(self.cv3), out_dim, k=1, act_type=act_type, norm_type=norm_type)


    def forward(self, x):
        x1 = self.cv1(x)
        x2 = self.cv2(x)
        inter_outs = [x1, x2]
        for m in self.cv3:
            y1 = inter_outs[-1]
            y2 = m(y1)
            inter_outs.append(y2)
        out = self.out(torch.cat(inter_outs, dim=1))

        return out

## DownSample Block proposed by YOLOv7
class DownSample(nn.Module):
    def __init__(self, in_dim, out_dim, act_type='silu', norm_type='BN', depthwise=False):
        super().__init__()
        inter_dim = out_dim // 2
        self.mp = nn.MaxPool2d((2, 2), 2)
        self.cv1 = Conv(in_dim, inter_dim, k=1, act_type=act_type, norm_type=norm_type)
        self.cv2 = nn.Sequential(
            Conv(in_dim, inter_dim, k=1, act_type=act_type, norm_type=norm_type),
            Conv(inter_dim, inter_dim, k=3, p=1, s=2, act_type=act_type, norm_type=norm_type, depthwise=depthwise)
        )

    def forward(self, x):
        x1 = self.cv1(self.mp(x))
        x2 = self.cv2(x)
        out = torch.cat([x1, x2], dim=1)

        return out


# ---------------------------- RepConv Modules ----------------------------
class RepConv(nn.Module):
    """
        The code referenced to https://github.com/WongKinYiu/yolov7/models/common.py
    """
    # Represented convolution
    # https://arxiv.org/abs/2101.03697

    def __init__(self, c1, c2, k=3, s=1, p=1, g=1, act_type='silu', deploy=False):
        super(RepConv, self).__init__()
        # -------------- Basic parameters --------------
        self.deploy = deploy
        self.groups = g
        self.in_channels = c1
        self.out_channels = c2

        # -------------- Network parameters --------------
        if deploy:
            self.rbr_reparam = nn.Conv2d(c1, c2, k, s, p, groups=g, bias=True)

        else:
            self.rbr_identity = (nn.BatchNorm2d(num_features=c1) if c2 == c1 and s == 1 else None)

            self.rbr_dense = nn.Sequential(
                nn.Conv2d(c1, c2, k, s, p, groups=g, bias=False),
                nn.BatchNorm2d(num_features=c2),
            )

            self.rbr_1x1 = nn.Sequential(
                nn.Conv2d(c1, c2, kernel_size=1, stride=s, bias=False),
                nn.BatchNorm2d(num_features=c2),
            )
        self.act = get_activation(act_type)


    def forward(self, inputs):
        if hasattr(self, "rbr_reparam"):
            return self.act(self.rbr_reparam(inputs))

        if self.rbr_identity is None:
            id_out = 0
        else:
            id_out = self.rbr_identity(inputs)

        return self.act(self.rbr_dense(inputs) + self.rbr_1x1(inputs) + id_out)
    
    def get_equivalent_kernel_bias(self):
        kernel3x3, bias3x3 = self._fuse_bn_tensor(self.rbr_dense)
        kernel1x1, bias1x1 = self._fuse_bn_tensor(self.rbr_1x1)
        kernelid, biasid = self._fuse_bn_tensor(self.rbr_identity)
        return (
            kernel3x3 + self._pad_1x1_to_3x3_tensor(kernel1x1) + kernelid,
            bias3x3 + bias1x1 + biasid,
        )

    def _pad_1x1_to_3x3_tensor(self, kernel1x1):
        if kernel1x1 is None:
            return 0
        else:
            return nn.functional.pad(kernel1x1, [1, 1, 1, 1])

    def _fuse_bn_tensor(self, branch):
        if branch is None:
            return 0, 0
        if isinstance(branch, nn.Sequential):
            kernel = branch[0].weight
            running_mean = branch[1].running_mean
            running_var = branch[1].running_var
            gamma = branch[1].weight
            beta = branch[1].bias
            eps = branch[1].eps
        else:
            assert isinstance(branch, nn.BatchNorm2d)
            if not hasattr(self, "id_tensor"):
                input_dim = self.in_channels // self.groups
                kernel_value = np.zeros(
                    (self.in_channels, input_dim, 3, 3), dtype=np.float32
                )
                for i in range(self.in_channels):
                    kernel_value[i, i % input_dim, 1, 1] = 1
                self.id_tensor = torch.from_numpy(kernel_value).to(branch.weight.device)
            kernel = self.id_tensor
            running_mean = branch.running_mean
            running_var = branch.running_var
            gamma = branch.weight
            beta = branch.bias
            eps = branch.eps
        std = (running_var + eps).sqrt()
        t = (gamma / std).reshape(-1, 1, 1, 1)
        return kernel * t, beta - running_mean * gamma / std

    def repvgg_convert(self):
        kernel, bias = self.get_equivalent_kernel_bias()
        return (
            kernel.detach().cpu().numpy(),
            bias.detach().cpu().numpy(),
        )

    def fuse_conv_bn(self, conv, bn):

        std = (bn.running_var + bn.eps).sqrt()
        bias = bn.bias - bn.running_mean * bn.weight / std

        t = (bn.weight / std).reshape(-1, 1, 1, 1)
        weights = conv.weight * t

        bn = nn.Identity()
        conv = nn.Conv2d(in_channels = conv.in_channels,
                              out_channels = conv.out_channels,
                              kernel_size = conv.kernel_size,
                              stride=conv.stride,
                              padding = conv.padding,
                              dilation = conv.dilation,
                              groups = conv.groups,
                              bias = True,
                              padding_mode = conv.padding_mode)

        conv.weight = torch.nn.Parameter(weights)
        conv.bias = torch.nn.Parameter(bias)
        return conv

    def fuse_repvgg_block(self):    
        if self.deploy:
            return
                
        self.rbr_dense = self.fuse_conv_bn(self.rbr_dense[0], self.rbr_dense[1])
        
        self.rbr_1x1 = self.fuse_conv_bn(self.rbr_1x1[0], self.rbr_1x1[1])
        rbr_1x1_bias = self.rbr_1x1.bias
        weight_1x1_expanded = torch.nn.functional.pad(self.rbr_1x1.weight, [1, 1, 1, 1])
        
        # Fuse self.rbr_identity
        if (isinstance(self.rbr_identity, nn.BatchNorm2d) or isinstance(self.rbr_identity, nn.modules.batchnorm.SyncBatchNorm)):
            identity_conv_1x1 = nn.Conv2d(
                    in_channels=self.in_channels,
                    out_channels=self.out_channels,
                    kernel_size=1,
                    stride=1,
                    padding=0,
                    groups=self.groups, 
                    bias=False)
            identity_conv_1x1.weight.data = identity_conv_1x1.weight.data.to(self.rbr_1x1.weight.data.device)
            identity_conv_1x1.weight.data = identity_conv_1x1.weight.data.squeeze().squeeze()

            identity_conv_1x1.weight.data.fill_(0.0)
            identity_conv_1x1.weight.data.fill_diagonal_(1.0)
            identity_conv_1x1.weight.data = identity_conv_1x1.weight.data.unsqueeze(2).unsqueeze(3)

            identity_conv_1x1 = self.fuse_conv_bn(identity_conv_1x1, self.rbr_identity)
            bias_identity_expanded = identity_conv_1x1.bias
            weight_identity_expanded = torch.nn.functional.pad(identity_conv_1x1.weight, [1, 1, 1, 1])            
        else:
            bias_identity_expanded = torch.nn.Parameter( torch.zeros_like(rbr_1x1_bias) )
            weight_identity_expanded = torch.nn.Parameter( torch.zeros_like(weight_1x1_expanded) )            
        
        self.rbr_dense.weight = torch.nn.Parameter(self.rbr_dense.weight + weight_1x1_expanded + weight_identity_expanded)
        self.rbr_dense.bias = torch.nn.Parameter(self.rbr_dense.bias + rbr_1x1_bias + bias_identity_expanded)
                
        self.rbr_reparam = self.rbr_dense
        self.deploy = True

        if self.rbr_identity is not None:
            del self.rbr_identity
            self.rbr_identity = None

        if self.rbr_1x1 is not None:
            del self.rbr_1x1
            self.rbr_1x1 = None

        if self.rbr_dense is not None:
            del self.rbr_dense
            self.rbr_dense = None