models/yolo.py

"""YOLOv5-specific modules

Usage:
    $ python path/to/models/yolo.py --cfg yolov5s.yaml
"""
# Copyright (c) Alibaba, Inc. and its affiliates.

import argparse
import logging
import sys
from copy import deepcopy
from pathlib import Path
import warnings

sys.path.append(Path(__file__).parent.parent.absolute().__str__())  # to run '$ python *.py' files in subdirectories
logger = logging.getLogger(__name__)

import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from models.common import *
from models.replknet import *
try:
    from models.gpvit import *
except:
    warnings.warn('Package mmdet is not installed. You can follow https://github.com/ChenhongyiYang/GPViT to install dependencies.')
    SpatialPriorModule = GPViTAdapterSingleStageESOD = None
from models.spconv import SPYOLOv5Head, SPYOLOv6Head
from models.experimental import *
from utils.autoanchor import check_anchor_order
from utils.general import make_divisible, check_file, set_logging, xyxy2xywh
from utils.torch_utils import time_synchronized, fuse_conv_and_bn, model_info, scale_img, initialize_weights, \
    select_device, copy_attr

try:
    import thop  # for FLOPs computation
except ImportError:
    thop = None

class Detect(nn.Module):
    stride = None  # strides computed during build
    onnx_dynamic = False  # ONNX export parameter

    def __init__(self, nc=80, anchors=(), ch=(), inplace=True):  # detection layer
        super(Detect, self).__init__()
        self.nc = nc  # number of classes
        self.no = nc + 5  # number of outputs per anchor
        self.nl = len(anchors)  # number of detection layers
        self.na = len(anchors[0]) // 2  # number of anchors
        self.grid = [torch.zeros(1)] * self.nl  # init grid
        a = torch.tensor(anchors).float().view(self.nl, -1, 2)
        self.register_buffer('anchors', a)  # shape(nl,na,2)
        self.register_buffer('anchor_grid', a.clone().view(self.nl, 1, -1, 1, 1, 2))  # shape(nl,1,na,1,1,2)
        # self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch)  # output conv
        self.m = get_decoupled_heads(ch, self.nc, self.na)  # decoupled head
        self.inplace = inplace  # use in-place ops (e.g. slice assignment)

        self.sparse = False
        self.register_buffer('sparse_gird', torch.zeros(1))

    def set_sparse(self):
        sp_dict = {nn.Conv2d: SPYOLOv5Head, YOLOv6Head: SPYOLOv6Head}
        sp_head = sp_dict[type(self.m[0])]
        self.m = nn.ModuleList(sp_head(m) for m in self.m)
        self.sparse = True
    
    @torch.no_grad()
    def get_indices(self, offsets, mask, thresh=0.3):
        device, dtype = mask.device, mask.dtype
        if torch.max(mask) > 1. or torch.min(mask) < 0.:
            mask = mask.detach().sigmoid()

        patch_w, patch_h = offsets[0, 3:5] - offsets[0, 1:3]
        if not hasattr(self, 'sparse_gird') or self.sparse_gird is None or self.sparse_gird[0].shape != (1,patch_h,patch_w):
            yv, xv = torch.meshgrid([torch.arange(patch_h), torch.arange(patch_w)])
            yv, xv = yv.to(device), xv.to(device)
            self.sparse_gird = torch.stack((torch.zeros_like(yv), yv, xv)).view(3,1,patch_h,patch_w)  # shape(1,ph,pw)
        gb, gy, gx = self.sparse_gird
        ob1, ox1, oy1 = offsets[:, :3].unsqueeze(-1).chunk(3, dim=1)  # shape(n,1,1)
        ob, ox, oy = (ob1 + gb).view(-1), (ox1 + gx).view(-1), (oy1 + gy).view(-1)
        
        maxima = F.max_pool2d(mask, 3, stride=1, padding=1) == mask
        response = mask >= thresh
        indices = (maxima & response).to(dtype)
        indices = F.max_pool2d(indices, 3, stride=1, padding=1)  # expansion  

        slices = indices[ob, 0, oy, ox].view(offsets.shape[0], 1, patch_h, patch_w)

        indices_per_layer = []
        for i in range(self.nl):
            s = 2 ** i

            if i != 0:
                slices_i = F.max_pool2d(slices, s, stride=s, padding=0)
                slices_i = F.max_pool2d(slices_i, 3, stride=1, padding=1)  # expansion
            else:
                slices_i = slices

            indices_per_layer.append(torch.nonzero(slices_i[:, 0, :, :]))

        ###################
        
        # indices_per_layer = []
        # for i in range(self.nl):
        #     s = 2 ** i
        #     if s > 1:
        #         # TODO: size-adaptive?
        #         mask_i = F.avg_pool2d(mask, s, stride=s, padding=0)
        #         # mask_i = F.max_pool2d(mask, s, stride=s, padding=0)
        #     else:
        #         mask_i = mask
            
        #     maxima = F.max_pool2d(mask_i, 3, stride=1, padding=1) == mask_i
        #     response = mask_i > thresh
        #     indices = (maxima & response).float()
        #     indices = F.max_pool2d(indices, 3, stride=1, padding=1)  # expansion for 3x3 conv

        #     sw, sh = patch_w // s, patch_h // s
        #     ob, ox, oy = (ob1 + gb[:,:sh,:sw]).view(-1), (ox1//s + gx[:,:sh,:sw]).view(-1), (oy1//s + gy[:,:sh,:sw]).view(-1)
        #     slices = indices[ob, 0, oy, ox].view(offsets.shape[0], sh, sw)

        #     indices_per_layer.append(torch.nonzero(slices))

        return indices_per_layer

    def forward(self, x):
        # x = x.copy()  # for profiling
        masks, offsets, indices_per_layer = None, None, None
        if isinstance(x, tuple):
            if len(x) == 2:
                x, offsets = x  # offsets(bi,x1,y1,x2,y2)
            else:
                x, offsets, masks = x
                assert len(masks) == 1 and not isinstance(masks, torch.Tensor)
                if offsets is not None and hasattr(self, 'sparse') and self.sparse:
                    indices_per_layer = self.get_indices(offsets, masks[0])
            if offsets is not None:
                img_bs = torch.max(offsets[:, 0]).int().item() + 1
            else:
                img_bs = x[0].shape[0]
        else:
            img_bs = x[0].shape[0]
        
        device = x[0].device
        z = []  # inference output
        patch_offsets = []
        for i in range(self.nl):
            # if len(x) > self.nl:
            #     hid_feat_i = F.max_pool2d(x[self.nl * 2 - 1 - i], kernel_size=8, stride=8, padding=0)  # 这里有一个倒序关系
            #     x[i] = torch.cat((x[i], hid_feat_i), dim=1)
            
            bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
            if offsets is not None:
                r = (2 ** (i - 1)) if self.nl == 4 else 2 ** i
                patch_off = torch.cat((offsets[:, :1], offsets[:, 1:] / r), dim=1)  # TODO: from 4 to 32
                patch_off_xy = patch_off[:, 1:3].view(-1, 1, 1, 1, 2)
                patch_offsets.append(patch_off)
            
            if indices_per_layer is not None:
                sp_x = self.m[i](x[i], indices_per_layer[i])  # sparse conv
                # x[i] = sp_x.dense(channels_first=True) deprecated # for training
            else:
                sp_x = None
                x[i] = self.m[i](x[i])  # conv
                x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()

            if not self.training:  # inference
                if self.grid[i].shape[2:4] != (ny, nx) or self.onnx_dynamic:
                    self.grid[i] = self._make_grid(nx, ny).to(device)

                if sp_x is not None:
                    y = sp_x.features.sigmoid().view(-1, self.na, self.no)
                    bi, yi, xi = sp_x.indices.long().T
                    assert offsets is not None
                    grid_off = self.grid[i][0, 0, yi, xi].view(-1, 1, 2) + patch_off_xy[bi, ...].view(-1, 1, 2)
                    anch_wh = self.anchor_grid[i].view(1, self.na, 2)
                    batch_ind = offsets[bi, 0]  # [num_patches, 5] --> [num_objects, 5], compatible for box concat
                else:
                    y = x[i].sigmoid()
                    anch_wh = self.anchor_grid[i].view(1, self.na, 1, 1, 2)
                    if offsets is not None:
                        grid_off = self.grid[i] + patch_off_xy
                        batch_ind = offsets[:, 0]   
                    else:
                        grid_off = self.grid[i]
                        batch_ind = None

                if self.inplace:
                    y[..., 0:2] = (y[..., 0:2] * 2. - 0.5 + grid_off) * self.stride[i]  # xy
                    y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * anch_wh  # wh
                else:  # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953
                    xy = (y[..., 0:2] * 2. - 0.5 + grid_off) * self.stride[i]  # xy
                    wh = (y[..., 2:4] * 2) ** 2 * anch_wh  # wh
                    y = torch.cat((xy, wh, y[..., 4:]), -1)
                # y[..., 4] = 1.0
                
                if offsets is not None:
                    pbox = []
                    for bi in range(img_bs):
                        pbox_bi = y[batch_ind == bi]
                        np = len(pbox_bi)
                        if np:
                            pbox.append(pbox_bi.view(-1, self.no))
                        else:
                            pbox.append(torch.zeros((0, self.no), device=device))
                    max_pnum = max([len(boxes) for boxes in pbox])
                    z.append(torch.stack(
                        [torch.cat((boxes, torch.zeros((max_pnum - len(boxes), self.no), device=device))) for boxes in pbox]
                    ))
                else:
                    z.append(y.view(bs, -1, self.no))
            
        if offsets is not None:
            x = (x, patch_offsets)
        return x if self.training else (torch.cat(z, 1), x)

    @staticmethod
    def _make_grid(nx=20, ny=20):
        yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
        return torch.stack((xv, yv), 2).view((1, 1, ny, nx, 2)).float()


class Segmenter(nn.Module):
    def __init__(self, nc=10, ch=()):
        super(Segmenter, self).__init__()
        self.m = nn.ModuleList(nn.Conv2d(x, nc, 1) for x in ch)  # output conv
    
    def forward(self, x):
        return [self.m[i](x[i]) for i in range(len(x))]
    

class Center(nn.Module):
    def __init__(self, nc=80, ch=()):  # detection layer
        super(Center, self).__init__()
        self.nc = nc  # number of classes
        self.no = nc + 4  # number of outputs
        self.nl = 1
        self.na = 0
        self.anchors = None
        self.anchor_grid = None
        self.grid = torch.zeros(1)
        self.b = torch.zeros(0)
        self.c = torch.zeros(0)
        self.stride = torch.tensor([4, 32])  # fake

        assert len(ch) == 1
        ch = ch[0]
        self.m = nn.ModuleList([
            nn.Sequential(  # hm
                nn.Conv2d(ch, ch, kernel_size=3, stride=1, padding=1, bias=True),
                nn.ReLU(inplace=True),
                nn.Conv2d(ch, nc, kernel_size=1, stride=1, padding=0, bias=True),
            ),
            nn.Sequential(  # wh
                nn.Conv2d(ch, ch, kernel_size=3, stride=1, padding=1, bias=True),
                nn.ReLU(inplace=True),
                nn.Conv2d(ch, 2, kernel_size=1, stride=1, padding=0, bias=True),
                nn.ReLU(inplace=True),
            ),
            nn.Sequential(  # reg
                nn.Conv2d(ch, ch, kernel_size=3, stride=1, padding=1, bias=True),
                nn.ReLU(inplace=True),
                nn.Conv2d(ch, 2, kernel_size=1, stride=1, padding=0, bias=True),
                nn.ReLU(inplace=True),
            ),
        ])  # output convs
        self.m[0][-1].bias.data.fill_(-2.19)  # hm, expect 0.01
    
    def forward(self, x):
        assert isinstance(x, tuple)
        x, offsets = x  # offsets(bi,x1,y1,x2,y2)
        assert len(x) == 1
        x = x[0]
        hm, wh, reg = [self.m[i](x) for i in range(3)]
        
        # offsets = torch.cat((offsets[:, :1], offsets[:, 1:] * 2.), dim=1)
            
        if not self.training:  # inference
            nb, nc, ny, nx = hm.shape
            device = hm.device
            
            wh_ = wh.permute(0, 2, 3, 1).contiguous()
            reg_ = reg.permute(0, 2, 3, 1).contiguous()
            if self.grid.shape[1:2] != wh_.shape[1:2]:
                self.grid = self._make_grid(nx, ny).to(x.device)
            # [nb, ny, nx, 4], absolute pixel relative to input size
            if offsets is not None:
                offsets_xy = offsets[:, 1:3].view(-1, 1, 1, 2) * 2
                bbox = torch.cat([self.grid + offsets_xy + reg_ - wh_ / 2.,
                                  self.grid + offsets_xy + reg_ + wh_ / 2.], dim=-1)
            else:
                bbox = torch.cat([self.grid + reg_ - wh_ / 2., self.grid + reg_ + wh_ / 2.], dim=-1)
            # no clamp
            # bbox[..., [0, 2]].clamp_(0, nx)
            # bbox[..., [1, 3]].clamp_(0, ny)
            # [nb, nc, ny, nx, 4]
            # print(self.grid.shape, wh_.shape, hm.shape, offsets_xy.shape, bbox.shape)
            bbox = bbox.view(nb, 1, ny, nx, 4).repeat((1, nc, 1, 1, 1)) * self.stride[0]
            
            if self.c.shape[:-1] != hm.shape:
                # [nb, nc, ny, nx, 1]
                self.c = torch.arange(nc).to(x.device).view(1, nc, 1, 1, 1).repeat((nb, 1, ny, nx, 1))
            self.b = offsets[:, :1].view(nb, 1, 1, 1, 1).repeat((1, nc, ny, nx, 1))
            
            hm_ = hm.sigmoid()
            hmax = F.max_pool2d(hm_, 3, stride=1, padding=1)
            maxima = hmax == hm_
            # hmax_cls = torch.argmax(hm_, dim=1, keepdim=True)
            # maxima_class = hmax_cls == self.c[..., 0]
            # maxima &= maxima_class
            
            # [n, 6]
            preds = torch.cat([bbox[maxima], hm_[maxima].view(-1, 1), self.c[maxima]], dim=1)
            bi = self.b[maxima][:, 0]

            pbox = []
            for i in range(torch.max(bi).int().item() + 1):
                pbox_bi = preds[bi == i]
                # topk_indices = torch.argsort(pbox_bi[:, 4], descending=True)[:500]
                if len(pbox_bi):
                    pbox.append(
                        torch.cat((xyxy2xywh(pbox_bi[:, :4]), pbox_bi[:, 4:5],
                                   F.one_hot(pbox_bi[:, 5].long(), self.nc)), dim=1)
                    )
                else:
                    pbox.append(torch.zeros((0, 5 + self.nc), device=device))
            max_pnum = max([len(boxes) for boxes in pbox])
            predictions = torch.stack(
                [torch.cat((boxes, torch.zeros((max_pnum - len(boxes), 5 + self.nc), device=device))) for boxes in pbox]
            )
        else:
            predictions = None

        x = ((hm, wh, reg), [offsets])  # for consistency when testing
            
        return x if self.training else (predictions, x)
    
    @staticmethod
    def _make_grid(nx=20, ny=20):
        yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
        return torch.stack((xv, yv), 2).view((1, ny, nx, 2)).float()
    

class Model(nn.Module):
    def __init__(self, cfg='yolov5s.yaml', ch=3, nc=None, anchors=None):  # model, input channels, number of classes
        super(Model, self).__init__()
        if isinstance(cfg, dict):
            self.yaml = cfg  # model dict
        else:  # is *.yaml
            import yaml  # for torch hub
            self.yaml_file = Path(cfg).name
            with open(cfg) as f:
                self.yaml = yaml.safe_load(f)  # model dict

        # Define model
        ch = self.yaml['ch'] = self.yaml.get('ch', ch)  # input channels
        if nc and nc != self.yaml['nc']:
            logger.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}")
            self.yaml['nc'] = nc  # override yaml value
        if anchors:
            logger.info(f'Overriding model.yaml anchors with anchors={anchors}')
            self.yaml['anchors'] = round(anchors)  # override yaml value
        self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch])  # model, savelist
        self.names = [str(i) for i in range(self.yaml['nc'])]  # default names
        self.inplace = self.yaml.get('inplace', True)
        # logger.info([x.shape for x in self.forward(torch.zeros(1, ch, 64, 64))])

        # Build strides, anchors
        m = self.model[-1]  # Detect()
        if isinstance(m, Detect):
            s = 256  # 2x min stride
            m.inplace = self.inplace
            # TODO
            # m.stride = torch.tensor([ 4., 8., 16., 32.])
            m.stride = torch.tensor([ 8., 16., 32.])
            # m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))])  # forward
            m.anchors /= m.stride.view(-1, 1, 1)
            check_anchor_order(m)
            self.stride = m.stride
            self._initialize_biases()  # only run once
            # logger.info('Strides: %s' % m.stride.tolist())
        elif isinstance(m, Center):
            # m.stride = torch.tensor(m.stride)  # no forward
            self.stride = m.stride
        # elif isinstance(m, Detect2):  deprecated
        #     s = 256  # 2x min stride
        #     m.inplace = self.inplace
        #     # m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))])  # forward
        #     self.forward(torch.zeros(1, ch, s, s))  # forward
        #     m.stride = torch.tensor(m.stride)  # no forward
        #     # m.anchors /= m.stride.view(-1, 1, 1) 这里不归一化
        #     self.stride = m.stride

        # Init weights, biases
        initialize_weights(self)
        try:
            self.info()
        except:
            logger.info('Failed to capture the model info')
        logger.info('')

    def forward(self, x, augment=False, profile=False, hm_only=False):
        if augment:
            return self.forward_augment(x)  # augmented inference, None
        else:
            return self.forward_once(x, profile, hm_only=hm_only)  # single-scale inference, train

    def forward_augment(self, x):
        img_size = x.shape[-2:]  # height, width
        s = [1, 0.83, 0.67]  # scales
        f = [None, 3, None]  # flips (2-ud, 3-lr)
        y = []  # outputs
        for si, fi in zip(s, f):
            xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max()))
            yi = self.forward_once(xi)[0]  # forward
            # cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1])  # save
            yi = self._descale_pred(yi, fi, si, img_size)
            y.append(yi)
        return torch.cat(y, 1), None  # augmented inference, train

    def forward_once(self, x, profile=False, hm_only=False):
        y, dt = [], []  # outputs
        
        masks, pred_masks, offsets = None, None, None
        heatmap = None
        if isinstance(x, tuple):
            x, masks = x  # ground-truth masks
        x0 = x
        B, C, H, W = x.shape
        for mi, m in enumerate(self.model):
            if m.f != -1:  # if not from previous layer
                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else (x0 if j == -100 else y[j]) for j in m.f]  # from earlier layers

            if profile:
                o = thop.profile(m, inputs=(x,), verbose=False)[0] / 1E9 * 2 if thop else 0  # FLOPs
                t = time_synchronized()
                for _ in range(10):
                    _ = m(x)
                dt.append((time_synchronized() - t) * 100)
                if m == self.model[0]:
                    logger.info(f"{'time (ms)':>10s} {'GFLOPs':>10s} {'params':>10s}  {'module'}")
                logger.info(f'{dt[-1]:10.2f} {o:10.2f} {m.np:10.0f}  {m.type}')

            if isinstance(m, HeatMapParser) and masks is not None:
                x = (x[0], masks)
            elif type(m) in [Detect, Center] and offsets is not None:
                x = (x, offsets)
                x = (*x, masks if masks is not None else pred_masks)
            elif isinstance(m, MaskedC3TR):
                x = (x, heatmap)
            elif isinstance(m, Token2Image):
                x = [x, (H, W)]
            
            x = m(x)  # run

            if isinstance(m, Segmenter):
                pred_masks = x
                if hm_only:
                    return (None, None), pred_masks
                if masks is None:
                    masks = pred_masks
            elif isinstance(m, HeatMapParser):
                if isinstance(x, torch.Tensor):
                    offsets = x
                    if offsets.size(0) == 0:
                        return (None, None), pred_masks
                elif isinstance(x[1], torch.Tensor):
                    x, offsets = x
                    if len(x) == 0:
                        return (None, None), pred_masks
                else:
                    x, thresh = x
                    heatmap = pred_masks[0].detach().sigmoid()
                    heatmap = heatmap > thresh
            y.append(x if m.i in self.save else None)  # save output

        if profile:
            logger.info('%.1fms total' % sum(dt))
            
        return x, pred_masks

    def _descale_pred(self, p, flips, scale, img_size):
        # de-scale predictions following augmented inference (inverse operation)
        if self.inplace:
            p[..., :4] /= scale  # de-scale
            if flips == 2:
                p[..., 1] = img_size[0] - p[..., 1]  # de-flip ud
            elif flips == 3:
                p[..., 0] = img_size[1] - p[..., 0]  # de-flip lr
        else:
            x, y, wh = p[..., 0:1] / scale, p[..., 1:2] / scale, p[..., 2:4] / scale  # de-scale
            if flips == 2:
                y = img_size[0] - y  # de-flip ud
            elif flips == 3:
                x = img_size[1] - x  # de-flip lr
            p = torch.cat((x, y, wh, p[..., 4:]), -1)
        return p

    def _initialize_biases(self, cf=None):  # initialize biases into Detect(), cf is class frequency
        # https://arxiv.org/abs/1708.02002 section 3.3
        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
        m = self.model[-1]  # Detect() module
        try:
            for mi, s in zip(m.m, m.stride):  # from YOLOv5 head
                b = mi.bias.view(m.na, -1)  # conv.bias(255) to (3,85)
                b.data[:, 4] += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)
                b.data[:, 5:] += math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum())  # cls
                mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
        except AttributeError:
            for mi, s in zip(m.m, m.stride):  # from decoupled head
                mi.obj_pred.bias.data.fill_(math.log(8 / (640 / s) ** 2))  # obj (8 objects per 640 image)
                mi.cls_pred.bias.data.fill_(math.log(0.6 / (m.nc - 0.99)) if cf is None else torch.log(cf / cf.sum()))

        for m_ in self.model:
            if str(m_.type) == 'models.yolo.Segmenter':  # stupid
                for mi in m_.m:
                    b = mi.bias.view(-1)
                    b.data += math.log(0.6 / (m.nc - 0.99) if cf is None else torch.log(cf / cf.sum()))  # cls
                    mi.bias = torch.nn.Parameter(b, requires_grad=True)
                break

    def _print_biases(self):
        m = self.model[-1]  # Detect() module
        for mi in m.m:  # from
            b = mi.bias.detach().view(m.na, -1).T  # conv.bias(255) to (3,85)
            logger.info(
                ('%6g Conv2d.bias:' + '%10.3g' * 6) % (mi.weight.shape[1], *b[:5].mean(1).tolist(), b[5:].mean()))

    # def _print_weights(self):
    #     for m in self.model.modules():
    #         if type(m) is Bottleneck:
    #             logger.info('%10.3g' % (m.w.detach().sigmoid() * 2))  # shortcut weights

    def fuse(self):  # fuse model Conv2d() + BatchNorm2d() layers
        logger.info('Fusing layers... ')
        for m in self.model.modules():
            if type(m) is Conv and hasattr(m, 'bn'):
                m.conv = fuse_conv_and_bn(m.conv, m.bn)  # update conv
                delattr(m, 'bn')  # remove batchnorm
                m.forward = m.fuseforward  # update forward
        try:
            self.info()
        except:
            print('Failed to capture the model info')
        return self

    def nms(self, mode=True):  # add or remove NMS module
        present = type(self.model[-1]) is NMS  # last layer is NMS
        if mode and not present:
            logger.info('Adding NMS... ')
            m = NMS()  # module
            m.f = -1  # from
            m.i = self.model[-1].i + 1  # index
            self.model.add_module(name='%s' % m.i, module=m)  # add
            self.eval()
        elif not mode and present:
            logger.info('Removing NMS... ')
            self.model = self.model[:-1]  # remove
        return self

    def autoshape(self):  # add AutoShape module
        logger.info('Adding AutoShape... ')
        m = AutoShape(self)  # wrap model
        copy_attr(m, self, include=('yaml', 'nc', 'hyp', 'names', 'stride'), exclude=())  # copy attributes
        return m

    def info(self, verbose=False, img_size=640):  # print model information
        model_info(self, verbose, img_size)


def parse_model(d, ch):  # model_dict, input_channels(3)
    logger.info('\n%3s%18s%3s%10s  %-40s%-30s' % ('', 'from', 'n', 'params', 'module', 'arguments'))
    anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple']
    na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors  # number of anchors
    no = na * (nc + 5)  # number of outputs = anchors * (classes + 5)

    layers, save, c2 = [], [], ch[-1]  # layers, savelist, ch out
    for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']):  # from, number, module, args
        m = eval(m) if isinstance(m, str) else m  # eval strings
        for j, a in enumerate(args):
            try:
                args[j] = eval(a) if isinstance(a, str) else a  # eval strings
            except:
                pass

        n = max(round(n * gd), 1) if n > 1 else n  # depth gain
        if m in [nn.Conv2d, Conv, GhostConv, Bottleneck, GhostBottleneck, SPP, ASPP, SPPF,
                 DWConv, DCN, RepLKConv, MixConv2d, Focus, Blur, CrossConv,
                 BottleneckCSP, C3, C3TR, MaskedC3TR, C2f, ResBlockLayer, RTMDetCSPLayer,
                 HeatMapParser]:
            c2 = args[0]
            if c2 != no and 'GPViTAdapterSingleStageESOD' not in [_x[2] for _x in d['backbone']]:  # if not output
                c2 = make_divisible(c2 * gw, 8)

            if m is HeatMapParser:
                args = [c2, *args[1:]]
            else:
                args = [ch[f], c2, *args[1:]]
                if m in [BottleneckCSP, C3, C3TR, MaskedC3TR, C2f, RTMDetCSPLayer]:
                    args.insert(2, n)  # number of repeats
                    n = 1
        elif m is nn.BatchNorm2d:
            args = [ch[f]]
        elif m is Concat:
            c2 = sum([ch[x] for x in f])
        elif m in [Add, nn.Identity]:
            pass
        elif m in [Detect, Segmenter]:  # Detect2 deprecated
            args.append([ch[x] for x in f])
            if len(args) > 1 and isinstance(args[1], int):  # number of anchors
                args[1] = [list(range(args[1] * 2))] * len(f)
        elif m is Center:
            args.append([ch[x] for x in f])
        elif m is SpatialPriorModule:
            c2 = args[0] * 2
        elif m is GPViTAdapterSingleStageESOD:
            pass
        elif m is Indexer:
            c2 = args[0]
            args = args[1:]
        elif m is Contract:
            c2 = ch[f] * args[0] ** 2
        elif m is Expand:
            c2 = ch[f] // args[0] ** 2
        else:
            c2 = ch[f]

        m_ = nn.Sequential(*[m(*args) for _ in range(n)]) if n > 1 else m(*args)  # module
        t = str(m)[8:-2].replace('__main__.', '')  # module type
        np = sum([x.numel() for x in m_.parameters()])  # number params
        m_.i, m_.f, m_.type, m_.np = i, f, t, np  # attach index, 'from' index, type, number params
        logger.info('%3s%18s%3s%10.0f  %-40s%-30s' % (i, f, n, np, t, args))  # print
        save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x not in [-1, -100])  # append to savelist
        layers.append(m_)
        if i == 0:
            ch = []
        ch.append(c2)
    return nn.Sequential(*layers), sorted(save)


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--cfg', type=str, default='yolov5s.yaml', help='model.yaml')
    parser.add_argument('--device', default='', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')
    opt = parser.parse_args()
    opt.cfg = check_file(opt.cfg)  # check file
    set_logging()
    device = select_device(opt.device)

    # Create model
    model = Model(opt.cfg).to(device)
    model.train()

    # Profile
    # img = torch.rand(8 if torch.cuda.is_available() else 1, 3, 320, 320).to(device)
    # y = model(img, profile=True)

    # Tensorboard (not working https://github.com/ultralytics/yolov5/issues/2898)
    # from torch.utils.tensorboard import SummaryWriter
    # tb_writer = SummaryWriter('.')
    # logger.info("Run 'tensorboard --logdir=models' to view tensorboard at http://localhost:6006/")
    # tb_writer.add_graph(torch.jit.trace(model, img, strict=False), [])  # add model graph
    # tb_writer.add_image('test', img[0], dataformats='CWH')  # add model to tensorboard