model/recurrent_flow_completion.py

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

from model.modules.deformconv import ModulatedDeformConv2d
from .misc import constant_init

class SecondOrderDeformableAlignment(ModulatedDeformConv2d):
    """Second-order deformable alignment module."""
    def __init__(self, *args, **kwargs):
        self.max_residue_magnitude = kwargs.pop('max_residue_magnitude', 5)

        super(SecondOrderDeformableAlignment, self).__init__(*args, **kwargs)

        self.conv_offset = nn.Sequential(
            nn.Conv2d(3 * self.out_channels, self.out_channels, 3, 1, 1),
            nn.LeakyReLU(negative_slope=0.1, inplace=True),
            nn.Conv2d(self.out_channels, self.out_channels, 3, 1, 1),
            nn.LeakyReLU(negative_slope=0.1, inplace=True),
            nn.Conv2d(self.out_channels, self.out_channels, 3, 1, 1),
            nn.LeakyReLU(negative_slope=0.1, inplace=True),
            nn.Conv2d(self.out_channels, 27 * self.deform_groups, 3, 1, 1),
        )
        self.init_offset()

    def init_offset(self):
        constant_init(self.conv_offset[-1], val=0, bias=0)

    def forward(self, x, extra_feat):
        out = self.conv_offset(extra_feat)
        o1, o2, mask = torch.chunk(out, 3, dim=1)

        # offset
        offset = self.max_residue_magnitude * torch.tanh(torch.cat((o1, o2), dim=1))
        offset_1, offset_2 = torch.chunk(offset, 2, dim=1)
        offset = torch.cat([offset_1, offset_2], dim=1)

        # mask
        mask = torch.sigmoid(mask)

        return torchvision.ops.deform_conv2d(x, offset, self.weight, self.bias, 
                                             self.stride, self.padding,
                                             self.dilation, mask)

class BidirectionalPropagation(nn.Module):
    def __init__(self, channel):
        super(BidirectionalPropagation, self).__init__()
        modules = ['backward_', 'forward_']
        self.deform_align = nn.ModuleDict()
        self.backbone = nn.ModuleDict()
        self.channel = channel

        for i, module in enumerate(modules):
            self.deform_align[module] = SecondOrderDeformableAlignment(
                2 * channel, channel, 3, padding=1, deform_groups=16)

            self.backbone[module] = nn.Sequential(
                nn.Conv2d((2 + i) * channel, channel, 3, 1, 1),
                nn.LeakyReLU(negative_slope=0.1, inplace=True),
                nn.Conv2d(channel, channel, 3, 1, 1),
            )

        self.fusion = nn.Conv2d(2 * channel, channel, 1, 1, 0)

    def forward(self, x):
        """
        x shape : [b, t, c, h, w]
        return [b, t, c, h, w]
        """
        b, t, c, h, w = x.shape
        feats = {}
        feats['spatial'] = [x[:, i, :, :, :] for i in range(0, t)]

        for module_name in ['backward_', 'forward_']:

            feats[module_name] = []

            frame_idx = range(0, t)
            mapping_idx = list(range(0, len(feats['spatial'])))
            mapping_idx += mapping_idx[::-1]

            if 'backward' in module_name:
                frame_idx = frame_idx[::-1]

            feat_prop = x.new_zeros(b, self.channel, h, w)
            for i, idx in enumerate(frame_idx):
                feat_current = feats['spatial'][mapping_idx[idx]]
                if i > 0:
                    cond_n1 = feat_prop

                    # initialize second-order features
                    feat_n2 = torch.zeros_like(feat_prop)
                    cond_n2 = torch.zeros_like(cond_n1)
                    if i > 1:  # second-order features
                        feat_n2 = feats[module_name][-2]
                        cond_n2 = feat_n2

                    cond = torch.cat([cond_n1, feat_current, cond_n2], dim=1) # condition information, cond(flow warped 1st/2nd feature)
                    feat_prop = torch.cat([feat_prop, feat_n2], dim=1) # two order feat_prop -1 & -2
                    feat_prop = self.deform_align[module_name](feat_prop, cond)

                # fuse current features
                feat = [feat_current] + \
                    [feats[k][idx] for k in feats if k not in ['spatial', module_name]] \
                    + [feat_prop]

                feat = torch.cat(feat, dim=1)
                # embed current features
                feat_prop = feat_prop + self.backbone[module_name](feat)

                feats[module_name].append(feat_prop)

            # end for
            if 'backward' in module_name:
                feats[module_name] = feats[module_name][::-1]

        outputs = []
        for i in range(0, t):
            align_feats = [feats[k].pop(0) for k in feats if k != 'spatial']
            align_feats = torch.cat(align_feats, dim=1)
            outputs.append(self.fusion(align_feats))

        return torch.stack(outputs, dim=1) + x


class deconv(nn.Module):
    def __init__(self,
                 input_channel,
                 output_channel,
                 kernel_size=3,
                 padding=0):
        super().__init__()
        self.conv = nn.Conv2d(input_channel,
                              output_channel,
                              kernel_size=kernel_size,
                              stride=1,
                              padding=padding)

    def forward(self, x):
        x = F.interpolate(x,
                          scale_factor=2,
                          mode='bilinear',
                          align_corners=True)
        return self.conv(x)


class P3DBlock(nn.Module):
    def __init__(self, in_channels, out_channels, kernel_size, stride, padding, use_residual=0, bias=True):
        super().__init__()
        self.conv1 = nn.Sequential(
                        nn.Conv3d(in_channels, out_channels, kernel_size=(1, kernel_size, kernel_size),
                                    stride=(1, stride, stride), padding=(0, padding, padding), bias=bias),
                        nn.LeakyReLU(0.2, inplace=True)
        )
        self.conv2 = nn.Sequential(
                        nn.Conv3d(out_channels, out_channels, kernel_size=(3, 1, 1), stride=(1, 1, 1),
                                    padding=(2, 0, 0), dilation=(2, 1, 1), bias=bias)
        )
        self.use_residual = use_residual

    def forward(self, feats):
        feat1 = self.conv1(feats)
        feat2 = self.conv2(feat1)
        if self.use_residual:
            output = feats + feat2
        else:
            output = feat2
        return output


class EdgeDetection(nn.Module):
    def __init__(self, in_ch=2, out_ch=1, mid_ch=16):
        super().__init__()
        self.projection = nn.Sequential(
            nn.Conv2d(in_ch, mid_ch, 3, 1, 1),
            nn.LeakyReLU(0.2, inplace=True)
        )

        self.mid_layer_1 = nn.Sequential(
            nn.Conv2d(mid_ch, mid_ch, 3, 1, 1),
            nn.LeakyReLU(0.2, inplace=True)
        )

        self.mid_layer_2 = nn.Sequential(
            nn.Conv2d(mid_ch, mid_ch, 3, 1, 1)
        )        

        self.l_relu = nn.LeakyReLU(0.01, inplace=True)

        self.out_layer = nn.Conv2d(mid_ch, out_ch, 1, 1, 0)

    def forward(self, flow):
        flow = self.projection(flow)
        edge = self.mid_layer_1(flow)
        edge = self.mid_layer_2(edge)
        edge = self.l_relu(flow + edge)
        edge = self.out_layer(edge)
        edge = torch.sigmoid(edge)
        return edge


class RecurrentFlowCompleteNet(nn.Module):
    def __init__(self, model_path=None):
        super().__init__()
        self.downsample = nn.Sequential(
                        nn.Conv3d(3, 32, kernel_size=(1, 5, 5), stride=(1, 2, 2), 
                                        padding=(0, 2, 2), padding_mode='replicate'),
                        nn.LeakyReLU(0.2, inplace=True)
        )
        
        self.encoder1 = nn.Sequential(
            P3DBlock(32, 32, 3, 1, 1),
            nn.LeakyReLU(0.2, inplace=True),
            P3DBlock(32, 64, 3, 2, 1),
            nn.LeakyReLU(0.2, inplace=True)
        ) # 4x

        self.encoder2 = nn.Sequential(
            P3DBlock(64, 64, 3, 1, 1),
            nn.LeakyReLU(0.2, inplace=True),
            P3DBlock(64, 128, 3, 2, 1),
            nn.LeakyReLU(0.2, inplace=True)
        ) # 8x

        self.mid_dilation = nn.Sequential(
            nn.Conv3d(128, 128, (1, 3, 3), (1, 1, 1), padding=(0, 3, 3), dilation=(1, 3, 3)), # p = d*(k-1)/2
            nn.LeakyReLU(0.2, inplace=True),
            nn.Conv3d(128, 128, (1, 3, 3), (1, 1, 1), padding=(0, 2, 2), dilation=(1, 2, 2)),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Conv3d(128, 128, (1, 3, 3), (1, 1, 1), padding=(0, 1, 1), dilation=(1, 1, 1)),
            nn.LeakyReLU(0.2, inplace=True)
        )

        # feature propagation module
        self.feat_prop_module = BidirectionalPropagation(128)

        self.decoder2 = nn.Sequential(
            nn.Conv2d(128, 128, 3, 1, 1),
            nn.LeakyReLU(0.2, inplace=True),
            deconv(128, 64, 3, 1),
            nn.LeakyReLU(0.2, inplace=True)
        ) # 4x

        self.decoder1 = nn.Sequential(
            nn.Conv2d(64, 64, 3, 1, 1),
            nn.LeakyReLU(0.2, inplace=True),
            deconv(64, 32, 3, 1),
            nn.LeakyReLU(0.2, inplace=True)
        ) # 2x

        self.upsample = nn.Sequential(
            nn.Conv2d(32, 32, 3, padding=1),
            nn.LeakyReLU(0.2, inplace=True),
            deconv(32, 2, 3, 1)
        )

        # edge loss
        self.edgeDetector = EdgeDetection(in_ch=2, out_ch=1, mid_ch=16)

        # Need to initial the weights of MSDeformAttn specifically
        for m in self.modules():
            if isinstance(m, SecondOrderDeformableAlignment):
                m.init_offset()

        if model_path is not None:
            print('Pretrained flow completion model has loaded...')
            ckpt = torch.load(model_path, map_location='cpu')
            self.load_state_dict(ckpt, strict=True)


    def forward(self, masked_flows, masks):
        # masked_flows: b t-1 2 h w
        # masks: b t-1 2 h w
        b, t, _, h, w = masked_flows.size()
        masked_flows = masked_flows.permute(0,2,1,3,4)
        masks = masks.permute(0,2,1,3,4)

        inputs = torch.cat((masked_flows, masks), dim=1)
        
        x = self.downsample(inputs)

        feat_e1 = self.encoder1(x)
        feat_e2 = self.encoder2(feat_e1) # b c t h w
        feat_mid = self.mid_dilation(feat_e2) # b c t h w
        feat_mid = feat_mid.permute(0,2,1,3,4) # b t c h w

        feat_prop = self.feat_prop_module(feat_mid)
        feat_prop = feat_prop.view(-1, 128, h//8, w//8) # b*t c h w

        _, c, _, h_f, w_f = feat_e1.shape
        feat_e1 = feat_e1.permute(0,2,1,3,4).contiguous().view(-1, c, h_f, w_f) # b*t c h w
        feat_d2 = self.decoder2(feat_prop) + feat_e1

        _, c, _, h_f, w_f = x.shape
        x = x.permute(0,2,1,3,4).contiguous().view(-1, c, h_f, w_f) # b*t c h w

        feat_d1 = self.decoder1(feat_d2)

        flow = self.upsample(feat_d1)
        if self.training:
            edge = self.edgeDetector(flow)
            edge = edge.view(b, t, 1, h, w)
        else:
            edge = None

        flow = flow.view(b, t, 2, h, w)

        return flow, edge
        

    def forward_bidirect_flow(self, masked_flows_bi, masks):
        """
        Args:
            masked_flows_bi: [masked_flows_f, masked_flows_b] | (b t-1 2 h w), (b t-1 2 h w)
            masks: b t 1 h w
        """
        masks_forward = masks[:, :-1, ...].contiguous()
        masks_backward = masks[:, 1:, ...].contiguous()

        # mask flow
        masked_flows_forward = masked_flows_bi[0] * (1-masks_forward)
        masked_flows_backward = masked_flows_bi[1] * (1-masks_backward)
        
        # -- completion --
        # forward
        pred_flows_forward, pred_edges_forward = self.forward(masked_flows_forward, masks_forward)

        # backward
        masked_flows_backward = torch.flip(masked_flows_backward, dims=[1])
        masks_backward = torch.flip(masks_backward, dims=[1])
        pred_flows_backward, pred_edges_backward = self.forward(masked_flows_backward, masks_backward)
        pred_flows_backward = torch.flip(pred_flows_backward, dims=[1])
        if self.training:
            pred_edges_backward = torch.flip(pred_edges_backward, dims=[1])

        return [pred_flows_forward, pred_flows_backward], [pred_edges_forward, pred_edges_backward]


    def combine_flow(self, masked_flows_bi, pred_flows_bi, masks):
        masks_forward = masks[:, :-1, ...].contiguous()
        masks_backward = masks[:, 1:, ...].contiguous()

        pred_flows_forward = pred_flows_bi[0] * masks_forward + masked_flows_bi[0] * (1-masks_forward)
        pred_flows_backward = pred_flows_bi[1] * masks_backward + masked_flows_bi[1] * (1-masks_backward)

        return pred_flows_forward, pred_flows_backward