RT-ODLab/models/yolov7/yolov7_fpn.py

import torch
import torch.nn as nn
import torch.nn.functional as F
from .yolov7_basic import Conv, ELANBlockFPN, DownSample, RepConv


# PaFPN-ELAN (YOLOv7's)
class Yolov7PaFPN(nn.Module):
    def __init__(self, 
                 in_dims=[512, 1024, 512],
                 out_dim=None,
                 width=1.0,
                 depth=1.0,
                 nbranch=4.0,
                 act_type='silu',
                 norm_type='BN',
                 depthwise=False):
        super(Yolov7PaFPN, self).__init__()
        self.in_dims = in_dims
        c3, c4, c5 = in_dims

        # top dwon
        ## P5 -> P4
        self.reduce_layer_1 = Conv(c5, round(256*width), k=1, norm_type=norm_type, act_type=act_type)
        self.reduce_layer_2 = Conv(c4, round(256*width), k=1, norm_type=norm_type, act_type=act_type)
        self.top_down_layer_1 = ELANBlockFPN(in_dim=round(256*width) + round(256*width),
                                             out_dim=round(256*width),
                                             expand_ratio=0.5,
                                             nbranch=nbranch,
                                             depth=depth,
                                             act_type=act_type,
                                             norm_type=norm_type,
                                             depthwise=depthwise
                                             )
        # P4 -> P3
        self.reduce_layer_3 = Conv(round(256*width), round(128*width), k=1, norm_type=norm_type, act_type=act_type)
        self.reduce_layer_4 = Conv(c3, round(128*width), k=1, norm_type=norm_type, act_type=act_type)
        self.top_down_layer_2 = ELANBlockFPN(in_dim=round(128*width) + round(128*width),
                                             out_dim=round(128*width),
                                             expand_ratio=0.5,
                                             nbranch=nbranch,
                                             depth=depth,
                                             act_type=act_type,
                                             norm_type=norm_type,
                                             depthwise=depthwise
                                             )

        # bottom up
        # P3 -> P4
        self.downsample_layer_1 = DownSample(in_dim=round(128*width), out_dim=round(256*width),
                                             act_type=act_type, norm_type=norm_type, depthwise=depthwise)
        self.bottom_up_layer_1 = ELANBlockFPN(in_dim=round(256*width) + round(256*width),
                                              out_dim=round(256*width),
                                              expand_ratio=0.5,
                                              nbranch=nbranch,
                                              depth=depth,
                                              act_type=act_type,
                                              norm_type=norm_type,
                                              depthwise=depthwise
                                              )
        # P4 -> P5
        self.downsample_layer_2 = DownSample(in_dim=round(256*width), out_dim=round(512*width),
                                             act_type=act_type, norm_type=norm_type, depthwise=depthwise)
        self.bottom_up_layer_2 = ELANBlockFPN(in_dim=round(512*width) + c5,
                                              out_dim=round(512*width),
                                              expand_ratio=0.5,
                                              nbranch=nbranch,
                                              depth=depth,
                                              act_type=act_type,
                                              norm_type=norm_type,
                                              depthwise=depthwise
                                              )
        
        # output proj layers
        if out_dim is not None:
            self.out_layers = nn.ModuleList([
                Conv(in_dim, out_dim, k=1,
                     norm_type=norm_type, act_type=act_type)
                     for in_dim in [round(128*width), round(256*width), round(512*width)]
                     ])
            self.out_dim = [out_dim] * 3
        else:
            self.out_layers = None
            self.out_dim = [round(128*width), round(256*width), round(512*width)]


    def forward(self, features):
        c3, c4, c5 = features

        # Top down
        ## P5 -> P4
        c6 = self.reduce_layer_1(c5)
        c7 = F.interpolate(c6, scale_factor=2.0)
        c8 = torch.cat([c7, self.reduce_layer_2(c4)], dim=1)
        c9 = self.top_down_layer_1(c8)
        ## P4 -> P3
        c10 = self.reduce_layer_3(c9)
        c11 = F.interpolate(c10, scale_factor=2.0)
        c12 = torch.cat([c11, self.reduce_layer_4(c3)], dim=1)
        c13 = self.top_down_layer_2(c12)

        # Bottom up
        ## p3 -> P4
        c14 = self.downsample_layer_1(c13)
        c15 = torch.cat([c14, c9], dim=1)
        c16 = self.bottom_up_layer_1(c15)
        ## P4 -> P5
        c17 = self.downsample_layer_2(c16)
        c18 = torch.cat([c17, c5], dim=1)
        c19 = self.bottom_up_layer_2(c18)

        out_feats = [c13, c16, c19] # [P3, P4, P5]
        
        # output proj layers
        if self.out_layers is not None:
            out_feats_proj = []
            for feat, layer in zip(out_feats, self.out_layers):
                out_feats_proj.append(layer(feat))
            return out_feats_proj

        return out_feats


def build_fpn(cfg, in_dims, out_dim=None):
    model = cfg['fpn']
    # build pafpn
    if model == 'yolov7_pafpn':
        fpn_net = Yolov7PaFPN(in_dims=in_dims,
                             out_dim=out_dim,
                             width=cfg['width'],
                             depth=cfg['depth'],
                             nbranch=cfg['nbranch'],
                             act_type=cfg['fpn_act'],
                             norm_type=cfg['fpn_norm'],
                             depthwise=cfg['fpn_depthwise']
                             )


    return fpn_net