RT-ODLab/models/detectors/yolo_free_v2/yolo_free_v2_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)
    elif act_type is None:
        return nn.Identity()


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
        p = p if d == 1 else d
        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)


# ---------------------------- YOLOv8's Modules ----------------------------
# BottleNeck
class Bottleneck(nn.Module):
    def __init__(self,
                 in_dim,
                 out_dim,
                 expand_ratio=0.5,
                 shortcut=False,
                 depthwise=False,
                 act_type='silu',
                 norm_type='BN'):
        super(Bottleneck, self).__init__()
        inter_dim = int(out_dim * expand_ratio)  # hidden channels            
        self.cv1 = Conv(in_dim, inter_dim, k=3, p=1, norm_type=norm_type, act_type=act_type, depthwise=depthwise)
        self.cv2 = Conv(inter_dim, out_dim, k=3, p=1, norm_type=norm_type, act_type=act_type, depthwise=depthwise)
        self.shortcut = shortcut and in_dim == out_dim

    def forward(self, x):
        h = self.cv2(self.cv1(x))

        return x + h if self.shortcut else h


# ELAN-CSP-Block
class ELAN_CSP_Block(nn.Module):
    def __init__(self,
                 in_dim,
                 out_dim,
                 expand_ratio=0.5,
                 nblocks=1,
                 shortcut=False,
                 act_type='silu',
                 norm_type='BN',
                 depthwise=False):
        super(ELAN_CSP_Block, self).__init__()
        inter_dim = int(out_dim * expand_ratio)
        self.cv1 = Conv(in_dim, inter_dim, k=1, norm_type=norm_type, act_type=act_type)
        self.cv2 = Conv(in_dim, inter_dim, k=1, norm_type=norm_type, act_type=act_type)
        self.m = nn.Sequential(*(
            Bottleneck(inter_dim, inter_dim, 1.0, shortcut, depthwise, act_type, norm_type)
            for _ in range(nblocks)))
        self.cv3 = Conv((2 + nblocks) * inter_dim, out_dim, k=1, act_type=act_type, norm_type=norm_type)


    def forward(self, x):
        x1 = self.cv1(x)
        x2 = self.cv2(x)
        out = list([x1, x2])

        out.extend(m(out[-1]) for m in self.m)

        out = self.cv3(torch.cat(out, dim=1))

        return out


# ---------------------------- FPN Modules ----------------------------
## build fpn's core block
def build_fpn_block(cfg, in_dim, out_dim):
    if cfg['fpn_core_block'] == 'elan_cspblock':
        layer = ELAN_CSP_Block(in_dim=in_dim,
                               out_dim=out_dim,
                               expand_ratio=0.5,
                               nblocks=round(3*cfg['depth']),
                               shortcut=False,
                               act_type=cfg['fpn_act'],
                               norm_type=cfg['fpn_norm'],
                               depthwise=cfg['fpn_depthwise']
                               )
        
    return layer

## build fpn's reduce layer
def build_reduce_layer(cfg, in_dim, out_dim):
    if cfg['fpn_reduce_layer'] == 'Conv':
        layer = Conv(in_dim, out_dim, k=1, act_type=cfg['fpn_act'], norm_type=cfg['fpn_norm'])
        
    return layer

## build fpn's downsample layer
def build_downsample_layer(cfg, in_dim, out_dim):
    if cfg['fpn_downsample_layer'] == 'Conv':
        layer = Conv(in_dim, out_dim, k=3, s=2, p=1, act_type=cfg['fpn_act'], norm_type=cfg['fpn_norm'])
        
    return layer