blocks_repvgg.py

from basic import *
import numpy as np

import torch
import torch.nn as nn
import torch.nn.functional as F

class RepVGGBlock(nn.Module):

    def __init__(self, in_channels, out_channels, kernel_size,
                 stride=1, padding=0, dilation=1, groups=1, padding_mode='zeros', deploy=False, use_se=False):
        super(RepVGGBlock, self).__init__()
        self.deploy = deploy
        self.groups = groups
        self.in_channels = in_channels

        assert kernel_size == 3
        assert padding == 1

        padding_11 = padding - kernel_size // 2

        self.nonlinearity = nn.ReLU()

        if use_se:
            self.se = SEBlock(out_channels, internal_neurons=out_channels // 16)
        else:
            self.se = nn.Identity()

        if deploy:
            self.rbr_reparam = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride,
                                      padding=padding, dilation=dilation, groups=groups, bias=True, padding_mode=padding_mode)

        else:
            self.rbr_identity = nn.BatchNorm2d(num_features=in_channels) if out_channels == in_channels and stride == 1 else None
            self.rbr_dense = ConvBN(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, groups=groups)
            self.rbr_1x1 = ConvBN(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, padding=padding_11, groups=groups)
            print('RepVGG Block, identity = ', self.rbr_identity)


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

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

        return self.nonlinearity(self.se(self.rbr_dense(inputs) + self.rbr_1x1(inputs) + id_out))


    #   Optional. This improves the accuracy and facilitates quantization.
    #   1.  Cancel the original weight decay on rbr_dense.conv.weight and rbr_1x1.conv.weight.
    #   2.  Use like this.
    #       loss = criterion(....)
    #       for every RepVGGBlock blk:
    #           loss += weight_decay_coefficient * 0.5 * blk.get_cust_L2()
    #       optimizer.zero_grad()
    #       loss.backward()
    def get_custom_L2(self):
        K3 = self.rbr_dense.conv.weight
        K1 = self.rbr_1x1.conv.weight
        t3 = (self.rbr_dense.bn.weight / ((self.rbr_dense.bn.running_var + self.rbr_dense.bn.eps).sqrt())).reshape(-1, 1, 1, 1).detach()
        t1 = (self.rbr_1x1.bn.weight / ((self.rbr_1x1.bn.running_var + self.rbr_1x1.bn.eps).sqrt())).reshape(-1, 1, 1, 1).detach()

        l2_loss_circle = (K3 ** 2).sum() - (K3[:, :, 1:2, 1:2] ** 2).sum()      # The L2 loss of the "circle" of weights in 3x3 kernel. Use regular L2 on them.
        eq_kernel = K3[:, :, 1:2, 1:2] * t3 + K1 * t1                           # The equivalent resultant central point of 3x3 kernel.
        l2_loss_eq_kernel = (eq_kernel ** 2 / (t3 ** 2 + t1 ** 2)).sum()        # Normalize for an L2 coefficient comparable to regular L2.
        return l2_loss_eq_kernel + l2_loss_circle



#   This func derives the equivalent kernel and bias in a DIFFERENTIABLE way.
#   You can get the equivalent kernel and bias at any time and do whatever you want,
    #   for example, apply some penalties or constraints during training, just like you do to the other models.
#   May be useful for quantization or pruning.
    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 torch.nn.functional.pad(kernel1x1, [1,1,1,1])

    def _fuse_bn_tensor(self, branch):
        if branch is None:
            return 0, 0
        if not isinstance(branch, nn.BatchNorm2d):
            kernel = branch.conv.weight
            running_mean = branch.bn.running_mean
            running_var = branch.bn.running_var
            gamma = branch.bn.weight
            beta = branch.bn.bias
            eps = branch.bn.eps
        else:
            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 switch_to_deploy(self):
        if hasattr(self, 'rbr_reparam'):
            return
        kernel, bias = self.get_equivalent_kernel_bias()
        self.rbr_reparam = nn.Conv2d(in_channels=self.rbr_dense.conv.in_channels, out_channels=self.rbr_dense.conv.out_channels,
                                     kernel_size=self.rbr_dense.conv.kernel_size, stride=self.rbr_dense.conv.stride,
                                     padding=self.rbr_dense.conv.padding, dilation=self.rbr_dense.conv.dilation, groups=self.rbr_dense.conv.groups, bias=True)
        self.rbr_reparam.weight.data = kernel
        self.rbr_reparam.bias.data = bias
        for para in self.parameters():
            para.detach_()
        self.__delattr__('rbr_dense')
        self.__delattr__('rbr_1x1')
        if hasattr(self, 'rbr_identity'):
            self.__delattr__('rbr_identity')


class OREPA_3x3_RepVGG(nn.Module):

    def __init__(self, in_channels, out_channels, kernel_size,
                 stride=1, padding=0, dilation=1, groups=1,
                 internal_channels_1x1_3x3=None,
                 deploy=False, nonlinear=None, single_init=False):
        super(OREPA_3x3_RepVGG, self).__init__()
        self.deploy = deploy

        if nonlinear is None:
            self.nonlinear = nn.Identity()
        else:
            self.nonlinear = nonlinear

        self.kernel_size = kernel_size
        self.in_channels = in_channels
        self.out_channels = out_channels
        self.groups = groups
        assert padding == kernel_size // 2

        self.stride = stride
        self.padding = padding
        self.dilation = dilation

        self.branch_counter = 0

        self.weight_rbr_origin = nn.Parameter(torch.Tensor(out_channels, int(in_channels/self.groups), kernel_size, kernel_size))
        init.kaiming_uniform_(self.weight_rbr_origin, a=math.sqrt(1.0))
        self.branch_counter += 1


        if groups < out_channels:
            self.weight_rbr_avg_conv = nn.Parameter(torch.Tensor(out_channels, int(in_channels/self.groups), 1, 1))
            self.weight_rbr_pfir_conv = nn.Parameter(torch.Tensor(out_channels, int(in_channels/self.groups), 1, 1))
            init.kaiming_uniform_(self.weight_rbr_avg_conv, a=1.0)
            init.kaiming_uniform_(self.weight_rbr_pfir_conv, a=1.0)
            self.weight_rbr_avg_conv.data
            self.weight_rbr_pfir_conv.data
            self.register_buffer('weight_rbr_avg_avg', torch.ones(kernel_size, kernel_size).mul(1.0/kernel_size/kernel_size))
            self.branch_counter += 1

        else:
            raise NotImplementedError
        self.branch_counter += 1

        if internal_channels_1x1_3x3 is None:
            internal_channels_1x1_3x3 = in_channels if groups < out_channels else 2 * in_channels   # For mobilenet, it is better to have 2X internal channels

        if internal_channels_1x1_3x3 == in_channels:
            self.weight_rbr_1x1_kxk_idconv1 = nn.Parameter(torch.zeros(in_channels, int(in_channels/self.groups), 1, 1))
            id_value = np.zeros((in_channels, int(in_channels/self.groups), 1, 1))
            for i in range(in_channels):
                id_value[i, i % int(in_channels/self.groups), 0, 0] = 1
            id_tensor = torch.from_numpy(id_value).type_as(self.weight_rbr_1x1_kxk_idconv1)
            self.register_buffer('id_tensor', id_tensor)

        else:
            self.weight_rbr_1x1_kxk_conv1 = nn.Parameter(torch.Tensor(internal_channels_1x1_3x3, int(in_channels/self.groups), 1, 1))
            init.kaiming_uniform_(self.weight_rbr_1x1_kxk_conv1, a=math.sqrt(1.0))
        self.weight_rbr_1x1_kxk_conv2 = nn.Parameter(torch.Tensor(out_channels, int(internal_channels_1x1_3x3/self.groups), kernel_size, kernel_size))
        init.kaiming_uniform_(self.weight_rbr_1x1_kxk_conv2, a=math.sqrt(1.0))
        self.branch_counter += 1

        expand_ratio = 8
        self.weight_rbr_gconv_dw = nn.Parameter(torch.Tensor(in_channels*expand_ratio, 1, kernel_size, kernel_size))
        self.weight_rbr_gconv_pw = nn.Parameter(torch.Tensor(out_channels, in_channels*expand_ratio, 1, 1))
        init.kaiming_uniform_(self.weight_rbr_gconv_dw, a=math.sqrt(1.0))
        init.kaiming_uniform_(self.weight_rbr_gconv_pw, a=math.sqrt(1.0))
        self.branch_counter += 1

        if out_channels == in_channels and stride == 1:
            self.branch_counter += 1

        self.vector = nn.Parameter(torch.Tensor(self.branch_counter, self.out_channels))
        self.bn = nn.BatchNorm2d(out_channels)

        self.fre_init()

        init.constant_(self.vector[0, :], 0.25)    #origin
        init.constant_(self.vector[1, :], 0.25)      #avg
        init.constant_(self.vector[2, :], 0.0)      #prior
        init.constant_(self.vector[3, :], 0.5)    #1x1_kxk
        init.constant_(self.vector[4, :], 0.5)     #dws_conv


    def fre_init(self):
        prior_tensor = torch.Tensor(self.out_channels, self.kernel_size, self.kernel_size)
        half_fg = self.out_channels/2
        for i in range(self.out_channels):
            for h in range(3):
                for w in range(3):
                    if i < half_fg:
                        prior_tensor[i, h, w] = math.cos(math.pi*(h+0.5)*(i+1)/3)
                    else:
                        prior_tensor[i, h, w] = math.cos(math.pi*(w+0.5)*(i+1-half_fg)/3)

        self.register_buffer('weight_rbr_prior', prior_tensor)

    def weight_gen(self):

        weight_rbr_origin = torch.einsum('oihw,o->oihw', self.weight_rbr_origin, self.vector[0, :])

        weight_rbr_avg = torch.einsum('oihw,o->oihw', torch.einsum('oihw,hw->oihw', self.weight_rbr_avg_conv, self.weight_rbr_avg_avg), self.vector[1, :])
        
        weight_rbr_pfir = torch.einsum('oihw,o->oihw', torch.einsum('oihw,ohw->oihw', self.weight_rbr_pfir_conv, self.weight_rbr_prior), self.vector[2, :])

        weight_rbr_1x1_kxk_conv1 = None
        if hasattr(self, 'weight_rbr_1x1_kxk_idconv1'):
            weight_rbr_1x1_kxk_conv1 = (self.weight_rbr_1x1_kxk_idconv1 + self.id_tensor).squeeze()
        elif hasattr(self, 'weight_rbr_1x1_kxk_conv1'):
            weight_rbr_1x1_kxk_conv1 = self.weight_rbr_1x1_kxk_conv1.squeeze()
        else:
            raise NotImplementedError
        weight_rbr_1x1_kxk_conv2 = self.weight_rbr_1x1_kxk_conv2

        if self.groups > 1:
            g = self.groups
            t, ig = weight_rbr_1x1_kxk_conv1.size()
            o, tg, h, w = weight_rbr_1x1_kxk_conv2.size()
            weight_rbr_1x1_kxk_conv1 = weight_rbr_1x1_kxk_conv1.view(g, int(t/g), ig)
            weight_rbr_1x1_kxk_conv2 = weight_rbr_1x1_kxk_conv2.view(g, int(o/g), tg, h, w)
            weight_rbr_1x1_kxk = torch.einsum('gti,gothw->goihw', weight_rbr_1x1_kxk_conv1, weight_rbr_1x1_kxk_conv2).view(o, ig, h, w)
        else:
            weight_rbr_1x1_kxk = torch.einsum('ti,othw->oihw', weight_rbr_1x1_kxk_conv1, weight_rbr_1x1_kxk_conv2)

        weight_rbr_1x1_kxk = torch.einsum('oihw,o->oihw', weight_rbr_1x1_kxk, self.vector[3, :])

        weight_rbr_gconv = self.dwsc2full(self.weight_rbr_gconv_dw, self.weight_rbr_gconv_pw, self.in_channels)
        weight_rbr_gconv = torch.einsum('oihw,o->oihw', weight_rbr_gconv, self.vector[4, :])    

        weight = weight_rbr_origin + weight_rbr_avg + weight_rbr_1x1_kxk + weight_rbr_pfir + weight_rbr_gconv

        return weight

    def dwsc2full(self, weight_dw, weight_pw, groups):
        
        t, ig, h, w = weight_dw.size()
        o, _, _, _ = weight_pw.size()
        tg = int(t/groups)
        i = int(ig*groups)
        weight_dw = weight_dw.view(groups, tg, ig, h, w)
        weight_pw = weight_pw.squeeze().view(o, groups, tg)
        
        weight_dsc = torch.einsum('gtihw,ogt->ogihw', weight_dw, weight_pw)
        return weight_dsc.view(o, i, h, w)

    def forward(self, inputs):
        weight = self.weight_gen()
        out = F.conv2d(inputs, weight, bias=None, stride=self.stride, padding=self.padding, dilation=self.dilation, groups=self.groups)

        return self.nonlinear(self.bn(out))

class RepVGGBlock_OREPA(nn.Module):

    def __init__(self, in_channels, out_channels, kernel_size,
                 stride=1, padding=0, dilation=1, groups=1, padding_mode='zeros', deploy=False, use_se=False, nonlinear=nn.ReLU()):
        super(RepVGGBlock_OREPA, self).__init__()
        self.deploy = deploy
        self.groups = groups
        self.in_channels = in_channels
        self.out_channels = out_channels

        self.padding = padding
        self.dilation = dilation
        self.groups = groups

        assert kernel_size == 3
        assert padding == 1

        padding_11 = padding - kernel_size // 2

        if nonlinear is None:
            self.nonlinearity = nn.Identity()
        else:
            self.nonlinearity = nonlinear

        if use_se:
            self.se = SEBlock(out_channels, internal_neurons=out_channels // 16)
        else:
            self.se = nn.Identity()

        if deploy:
            self.rbr_reparam = nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride,
                                      padding=padding, dilation=dilation, groups=groups, bias=True, padding_mode=padding_mode)

        else:
            self.rbr_identity = nn.BatchNorm2d(num_features=in_channels) if out_channels == in_channels and stride == 1 else None
            self.rbr_dense = OREPA_3x3_RepVGG(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, groups=groups, dilation=1)
            self.rbr_1x1 = ConvBN(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, padding=padding_11, groups=groups, dilation=1)
            print('RepVGG Block, identity = ', self.rbr_identity)


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

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

        out1 = self.rbr_dense(inputs)
        out2 = self.rbr_1x1(inputs)
        out3 = id_out
        out = out1 + out2 + out3

        return self.nonlinearity(self.se(out))


    #   Optional. This improves the accuracy and facilitates quantization.
    #   1.  Cancel the original weight decay on rbr_dense.conv.weight and rbr_1x1.conv.weight.
    #   2.  Use like this.
    #       loss = criterion(....)
    #       for every RepVGGBlock blk:
    #           loss += weight_decay_coefficient * 0.5 * blk.get_cust_L2()
    #       optimizer.zero_grad()
    #       loss.backward()

    # Not used for OREPA
    def get_custom_L2(self):
        K3 = self.rbr_dense.weight_gen()
        K1 = self.rbr_1x1.conv.weight
        t3 = (self.rbr_dense.bn.weight / ((self.rbr_dense.bn.running_var + self.rbr_dense.bn.eps).sqrt())).reshape(-1, 1, 1, 1).detach()
        t1 = (self.rbr_1x1.bn.weight / ((self.rbr_1x1.bn.running_var + self.rbr_1x1.bn.eps).sqrt())).reshape(-1, 1, 1, 1).detach()

        l2_loss_circle = (K3 ** 2).sum() - (K3[:, :, 1:2, 1:2] ** 2).sum()      # The L2 loss of the "circle" of weights in 3x3 kernel. Use regular L2 on them.
        eq_kernel = K3[:, :, 1:2, 1:2] * t3 + K1 * t1                           # The equivalent resultant central point of 3x3 kernel.
        l2_loss_eq_kernel = (eq_kernel ** 2 / (t3 ** 2 + t1 ** 2)).sum()        # Normalize for an L2 coefficient comparable to regular L2.
        return l2_loss_eq_kernel + l2_loss_circle

    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 torch.nn.functional.pad(kernel1x1, [1,1,1,1])

    def _fuse_bn_tensor(self, branch):
        if branch is None:
            return 0, 0
        if not isinstance(branch, nn.BatchNorm2d):
            if isinstance(branch, OREPA_3x3_RepVGG):
                kernel = branch.weight_gen()
            elif isinstance(branch, ConvBN):
                kernel = branch.conv.weight
            else:
                raise NotImplementedError
            running_mean = branch.bn.running_mean
            running_var = branch.bn.running_var
            gamma = branch.bn.weight
            beta = branch.bn.bias
            eps = branch.bn.eps
        else:
            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 switch_to_deploy(self):
        if hasattr(self, 'rbr_reparam'):
            return
        kernel, bias = self.get_equivalent_kernel_bias()
        self.rbr_reparam = nn.Conv2d(in_channels=self.rbr_dense.in_channels, out_channels=self.rbr_dense.out_channels,
                                     kernel_size=self.rbr_dense.kernel_size, stride=self.rbr_dense.stride,
                                     padding=self.rbr_dense.padding, dilation=self.rbr_dense.dilation, groups=self.rbr_dense.groups, bias=True)
        self.rbr_reparam.weight.data = kernel
        self.rbr_reparam.bias.data = bias
        for para in self.parameters():
            para.detach_()
        self.__delattr__('rbr_dense')
        self.__delattr__('rbr_1x1')
        if hasattr(self, 'rbr_identity'):
            self.__delattr__('rbr_identity')

class SEBlock(nn.Module):

    def __init__(self, input_channels, internal_neurons):
        super(SEBlock, self).__init__()
        self.down = nn.Conv2d(in_channels=input_channels, out_channels=internal_neurons, kernel_size=1, stride=1, bias=True)
        self.up = nn.Conv2d(in_channels=internal_neurons, out_channels=input_channels, kernel_size=1, stride=1, bias=True)
        self.input_channels = input_channels

    def forward(self, inputs):
        x = F.avg_pool2d(inputs, kernel_size=inputs.size(3))
        x = self.down(x)
        x = F.relu(x)
        x = self.up(x)
        x = torch.sigmoid(x)
        x = x.view(-1, self.input_channels, 1, 1)
        return inputs * x