ultralytics/YOLOV3源码解读
1、model.py源码解读：不足之处希望大佬指出
import os
import torch.nn.functional as F
from utils.parse_config import *
from utils.utils import *
ONNX_EXPORT = False
def create_modules(module_defs):
"""
Constructs module list of layer blocks from module configuration in module_defs
"""
hyperparams = module_defs.pop(0) #移除返回列表的首个元素
output_filters = [int(hyperparams['channels'])]# 图像通道数
module_list = nn.ModuleList() # module_list用于储存每个block，block对应的是cfg中的一块
yolo_layer_count = 0
for i, module_def in enumerate(module_defs): #返回模块索引i和模块信息module_def
modules = nn.Sequential() #采用nn.Sequential()创建modules，一个modules有多个层
if module_def['type'] == 'convolutional':
bn = int(module_def['batch_normalize']) #在cfg中batch_normalize是1需要加bn层，为0不加
filters = int(module_def['filters']) #卷积核数目
kernel_size = int(module_def['size']) #卷积核尺寸
pad = (kernel_size - 1) // 2 if int(module_def['pad']) else 0 #这个地方看了好久，很妙的公式，在cfg文件中我们看到padding
#全部设置为1，pdd =0该层不用padding，pad=1需要填充；对于3*3的步长为2的降采样卷积核，可以使填充后计算得到1/2降采样效果，对于1*1的卷积核pad=0
#添加卷积层
modules.add_module('conv_%d' % i, nn.Conv2d(in_channels=output_filters[-1],#通道列表的最后一个数
out_channels=filters,
kernel_size=kernel_size,
stride=int(module_def['stride']),
padding=pad,
bias=not bn))
if bn:
modules.add_module('batch_norm_%d' % i, nn.BatchNorm2d(filters))
if module_def['activation'] == 'leaky':
modules.add_module('leaky_%d' % i, nn.LeakyReLU(0.1, inplace=True))
# 池化层
elif module_def['type'] == 'maxpool':
kernel_size = int(module_def['size'])
stride = int(module_def['stride'])
if kernel_size == 2 and stride == 1:
modules.add_module('_debug_padding_%d' % i, nn.ZeroPad2d((0, 1, 0, 1)))
maxpool = nn.MaxPool2d(kernel_size=kernel_size, stride=stride, padding=int((kernel_size - 1) // 2))
modules.add_module('maxpool_%d' % i, maxpool)
#上采样
elif module_def['type'] == 'upsample':
# upsample = nn.Upsample(scale_factor=int(module_def['stride']), mode='nearest') # WARNING: deprecated
upsample = Upsample(scale_factor=int(module_def['stride'])) #步长为2，最近邻插值
modules.add_module('upsample_%d' % i, upsample)
elif module_def['type'] == 'route':
layers = [int(x) for x in module_def['layers'].split(',')]
filters = sum([output_filters[i + 1 if i > 0 else i] for i in layers])#卷积核数目为两层的加和
modules.add_module('route_%d' % i, EmptyLayer()) #初始化为空,占位层
#shortcut层
elif module_def['type'] == 'shortcut':
filters = output_filters[int(module_def['from'])]
modules.add_module('shortcut_%d' % i, EmptyLayer())
#yolo层
elif module_def['type'] == 'yolo':
anchor_idxs = [int(x) for x in module_def['mask'].split(',')]
# Extract anchors
anchors = [float(x) for x in module_def['anchors'].split(',')] #通过设置决定每次使用哪三个anchor
anchors = [(anchors[i], anchors[i + 1]) for i in range(0, len(anchors), 2)]
anchors = [anchors[i] for i in anchor_idxs]
nc = int(module_def['classes']) # number of classes
img_size = hyperparams['height'] ##默认w=h，只提取一个就行
# Define detection layer
yolo_layer = YOLOLayer(anchors, nc, img_size, yolo_layer_count, cfg=hyperparams['cfg'])
modules.add_module('yolo_%d' % i, yolo_layer)
yolo_layer_count += 1
# Register module list and number of output filters
module_list.append(modules) #模型放在这个list
output_filters.append(filters)#将每一层的输出通道数记录
return hyperparams, module_list
class EmptyLayer(nn.Module):
"""Placeholder for 'route' and 'shortcut' layers"""
def __init__(self):
super(EmptyLayer, self).__init__()
def forward(self, x):
return x
class Upsample(nn.Module):
# Custom Upsample layer (nn.Upsample gives deprecated warning message)
def __init__(self, scale_factor=1, mode='nearest'):
super(Upsample, self).__init__()
self.scale_factor = scale_factor
self.mode = mode
def forward(self, x):
return F.interpolate(x, scale_factor=self.scale_factor, mode=self.mode) #根据给定输出size或倍率上采样或下采样数据
class YOLOLayer(nn.Module):
def __init__(self, anchors, nc, img_size, yolo_layer, cfg):
super(YOLOLayer, self).__init__()
self.anchors = torch.Tensor(anchors)
self.na = len(anchors) # number of anchors (3)
self.nc = nc # number of classes (80)
self.nx = 0 # initialize number of x gridpoints
self.ny = 0 # initialize number of y gridpoints
if ONNX_EXPORT: # grids must be computed in __init__
stride = [32, 16, 8][yolo_layer] # stride of this layer
nx = int(img_size[1] / stride) # number x grid points
ny = int(img_size[0] / stride) # number y grid points
create_grids(self, max(img_size), (nx, ny))
def forward(self, p, img_size, var=None):#p为传入的预测值，torch.Size([bs, 255, 13, 13])
if ONNX_EXPORT:
bs = 1 # batch size
else:
bs, ny, nx = p.shape[0], p.shape[-2], p.shape[-1]
if (self.nx, self.ny) != (nx, ny):
create_grids(self, img_size, (nx, ny), p.device)
# p.view(bs, 255, 13, 13) -- > (bs, 3, 13, 13, 85) # (bs, anchors, grid, grid, classes + xywh)
p = p.view(bs, self.na, self.nc + 5, self.ny, self.nx).permute(0, 1, 3, 4, 2).contiguous() # prediction
# 把255的列向量信息，加一个维度分开成3个box
# p的shape为：torch.Size([1, 3, 13, 13, 85]) ，(bs, anchors, grid, grid, xywh+conf+class)
# 注意，当前YOLO层分配了3个anchor
if self.training:
return p
elif ONNX_EXPORT:
# Constants CAN NOT BE BROADCAST, ensure correct shape!
ngu = self.ng.repeat((1, self.na * self.nx * self.ny, 1))
grid_xy = self.grid_xy.repeat((1, self.na, 1, 1, 1)).view((1, -1, 2))
anchor_wh = self.anchor_wh.repeat((1, 1, self.nx, self.ny, 1)).view((1, -1, 2)) / ngu
# p = p.view(-1, 5 + self.nc)
# xy = torch.sigmoid(p[..., 0:2]) + grid_xy[0] # x, y
# wh = torch.exp(p[..., 2:4]) * anchor_wh[0] # width, height
# p_conf = torch.sigmoid(p[:, 4:5]) # Conf
# p_cls = F.softmax(p[:, 5:85], 1) * p_conf # SSD-like conf
# return torch.cat((xy / ngu[0], wh, p_conf, p_cls), 1).t()
p = p.view(1, -1, 5 + self.nc)
xy = torch.sigmoid(p[..., 0:2]) + grid_xy # x, y
wh = torch.exp(p[..., 2:4]) * anchor_wh # width, height
p_conf = torch.sigmoid(p[..., 4:5]) # Conf
p_cls = p[..., 5:5 + self.nc]
# Broadcasting only supported on first dimension in CoreML. See onnx-coreml/_operators.py
# p_cls = F.softmax(p_cls, 2) * p_conf # SSD-like conf
p_cls = torch.exp(p_cls).permute((2, 1, 0))
p_cls = p_cls / p_cls.sum(0).unsqueeze(0) * p_conf.permute((2, 1, 0)) # F.softmax() equivalent
p_cls = p_cls.permute(2, 1, 0)
return torch.cat((xy / ngu, wh, p_conf, p_cls), 2).squeeze().t()
else: # inference
io = p.clone() # inference output
io[..., 0:2] = torch.sigmoid(io[..., 0:2]) + self.grid_xy # xy
io[..., 2:4] = torch.exp(io[..., 2:4]) * self.anchor_wh # wh yolo method
# io[..., 2:4] = ((torch.sigmoid(io[..., 2:4]) * 2) ** 3) * self.anchor_wh # wh power method
io[..., 4:] = torch.sigmoid(io[..., 4:]) # p_conf, p_cls
# io[..., 5:] = F.softmax(io[..., 5:], dim=4) # p_cls
io[..., :4] *= self.stride
if self.nc == 1:
io[..., 5] = 1 # single-class model https://github.com/ultralytics/yolov3/issues/235
# reshape from [1, 3, 13, 13, 85] to [1, 507, 85]
return io.view(bs, -1, 5 + self.nc), p
class Darknet(nn.Module):
"""YOLOv3 object detection model"""
def __init__(self, cfg, img_size=(416, 416)):
super(Darknet, self).__init__() #继承父类的构造函数
self.module_defs = parse_model_cfg(cfg) #返回模型的参数信息
self.module_defs[0]['cfg'] = cfg
self.module_defs[0]['height'] = img_size
self.hyperparams, self.module_list = create_modules(self.module_defs)#获取net网络配置参数和模型结构
self.yolo_layers = get_yolo_layers(self)
# Needed to write header when saving weights
self.header_info = np.zeros(5, dtype=np.int32) # First five are header values
self.seen = self.header_info[3] # number of images seen during training
def forward(self, x, var=None):
img_size = max(x.shape[-2:])
layer_outputs = [] #储存每层输出
output = [] #存储YOLO层的检测输出
for i, (module_def, module) in enumerate(zip(self.module_defs, self.module_list)):
mtype = module_def['type']
if mtype in ['convolutional', 'upsample', 'maxpool']:
x = module(x) #计算前向输出
elif mtype == 'route':
layer_i = [int(x) for x in module_def['layers'].split(',')]
if len(layer_i) == 1:
x = layer_outputs[layer_i[0]]
else:
x = torch.cat([layer_outputs[i] for i in layer_i], 1)#按第一维度融合
elif mtype == 'shortcut':
layer_i = int(module_def['from'])
x = layer_outputs[-1] + layer_outputs[layer_i]#加和其上一层-1，当前层不是-1执行最后才append）和往上数第三层(-3)
elif mtype == 'yolo':
x = module[0](x, img_size) #训练需要计算损失函数
output.append(x) #存储yolo层前向传播结果
layer_outputs.append(x) #所有层输出
if self.training:
return output
elif ONNX_EXPORT:
output = torch.cat(output, 1) # cat 3 layers 85 x (507, 2028, 8112) to 85 x 10647
nc = self.module_list[self.yolo_layers[0]][0].nc # number of classes
return output[5:5 + nc].t(), output[:4].t() # ONNX scores, boxes
else:
io, p = list(zip(*output)) # inference output, training output
return torch.cat(io, 1), p
def fuse(self):
# Fuse Conv2d + BatchNorm2d layers throughout model
fused_list = nn.ModuleList()
for a in list(self.children())[0]:
for i, b in enumerate(a):
if isinstance(b, nn.modules.batchnorm.BatchNorm2d):
# fuse this bn layer with the previous conv2d layer
conv = a[i - 1]
fused = torch_utils.fuse_conv_and_bn(conv, b)
a = nn.Sequential(fused, *list(a.children())[i + 1:])
break
fused_list.append(a)
self.module_list = fused_list
# model_info(self) # yolov3-spp reduced from 225 to 152 layers
def get_yolo_layers(model):
a = [module_def['type'] == 'yolo' for module_def in model.module_defs]
return [i for i, x in enumerate(a) if x] # [82, 94, 106] for yolov3
def create_grids(self, img_size=416, ng=(13, 13), device='cpu'):
nx, ny = ng # x and y grid size
self.img_size = img_size
self.stride = img_size / max(ng)
# build xy offsets
yv, xv = torch.meshgrid([torch.arange(ny), torch.arange(nx)])
self.grid_xy = torch.stack((xv, yv), 2).to(device).float().view((1, 1, ny, nx, 2))
# build wh gains
self.anchor_vec = self.anchors.to(device) / self.stride
self.anchor_wh = self.anchor_vec.view(1, self.na, 1, 1, 2).to(device)
self.ng = torch.Tensor(ng).to(device)
self.nx = nx
self.ny = ny
def load_darknet_weights(self, weights, cutoff=-1):
# Parses and loads the weights stored in 'weights'
# cutoff: save layers between 0 and cutoff (if cutoff = -1 all are saved)
weights_file = weights.split(os.sep)[-1]
# Try to download weights if not available locally
if not os.path.isfile(weights):
try:
os.system('wget https://pjreddie.com/media/files/' + weights_file + ' -O ' + weights)
except IOError:
print(weights + ' not found.\nTry https://drive.google.com/drive/folders/1uxgUBemJVw9wZsdpboYbzUN4bcRhsuAI')
# Establish cutoffs
if weights_file == 'darknet53.conv.74':
cutoff = 75
elif weights_file == 'yolov3-tiny.conv.15':
cutoff = 15
# Open the weights file
with open(weights, 'rb') as f:
header = np.fromfile(f, dtype=np.int32, count=5) # First five are header values
# Needed to write header when saving weights
self.header_info = header
self.seen = header[3] # number of images seen during training
weights = np.fromfile(f, dtype=np.float32) # The rest are weights
ptr = 0
for i, (module_def, module) in enumerate(zip(self.module_defs[:cutoff], self.module_list[:cutoff])):
if module_def['type'] == 'convolutional':
conv_layer = module[0]
if module_def['batch_normalize']:
# Load BN bias, weights, running mean and running variance
bn_layer = module[1]
num_b = bn_layer.bias.numel() # Number of biases
# Bias
bn_b = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(bn_layer.bias)
bn_layer.bias.data.copy_(bn_b)
ptr += num_b
# Weight
bn_w = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(bn_layer.weight)
bn_layer.weight.data.copy_(bn_w)
ptr += num_b
# Running Mean
bn_rm = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(bn_layer.running_mean)
bn_layer.running_mean.data.copy_(bn_rm)
ptr += num_b
# Running Var
bn_rv = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(bn_layer.running_var)
bn_layer.running_var.data.copy_(bn_rv)
ptr += num_b
else:
# Load conv. bias
num_b = conv_layer.bias.numel()
conv_b = torch.from_numpy(weights[ptr:ptr + num_b]).view_as(conv_layer.bias)
conv_layer.bias.data.copy_(conv_b)
ptr += num_b
# Load conv. weights
num_w = conv_layer.weight.numel()
conv_w = torch.from_numpy(weights[ptr:ptr + num_w]).view_as(conv_layer.weight)
conv_layer.weight.data.copy_(conv_w)
ptr += num_w
return cutoff
def save_weights(self, path='model.weights', cutoff=-1):
# Converts a PyTorch model to Darket format (*.pt to *.weights)
# Note: Does not work if model.fuse() is applied
with open(path, 'wb') as f:
self.header_info[3] = self.seen # number of images seen during training
self.header_info.tofile(f)
# Iterate through layers
for i, (module_def, module) in enumerate(zip(self.module_defs[:cutoff], self.module_list[:cutoff])):
if module_def['type'] == 'convolutional':
conv_layer = module[0]
# If batch norm, load bn first
if module_def['batch_normalize']:
bn_layer = module[1]
bn_layer.bias.data.cpu().numpy().tofile(f)
bn_layer.weight.data.cpu().numpy().tofile(f)
bn_layer.running_mean.data.cpu().numpy().tofile(f)
bn_layer.running_var.data.cpu().numpy().tofile(f)
# Load conv bias
else:
conv_layer.bias.data.cpu().numpy().tofile(f)
# Load conv weights
conv_layer.weight.data.cpu().numpy().tofile(f)
def convert(cfg='cfg/yolov3-spp.cfg', weights='weights/yolov3-spp.weights'):
# Converts between PyTorch and Darknet format per extension (i.e. *.weights convert to *.pt and vice versa)
# from models import *; convert('cfg/yolov3-spp.cfg', 'weights/yolov3-spp.weights')
# Initialize model
model = Darknet(cfg)
# Load weights and save
if weights.endswith('.pt'): # if PyTorch format
model.load_state_dict(torch.load(weights, map_location='cpu')['model'])
save_weights(model, path='converted.weights', cutoff=-1)
print("Success: converted '%s' to 'converted.weights'" % weights)
elif weights.endswith('.weights'): # darknet format
_ = load_darknet_weights(model, weights)
chkpt = {'epoch': -1, 'best_loss': None, 'model': model.state_dict(), 'optimizer': None}
torch.save(chkpt, 'converted.pt')
print("Success: converted '%s' to 'converted.pt'" % weights)
else:
print('Error: extension not supported.')
