一、本文介绍
这次给大家带来的改进机制是EfficientNetV1主干,用其替换我们YOLOv8的特征提取网络,其主要思想是通过均衡地缩放网络的深度、宽度和分辨率,以提高卷积神经网络的性能。这种方法采用了一个简单但有效的复合系数,统一调整所有维度。经过我的实验该主干网络确实能够涨点在大中小三种物体检测上,同时该主干网络提供多种版本,大家可以在源代码中进行修改版本的使用。本文通过介绍其主要框架原理,然后教大家如何添加该网络结构到网络模型中。
推荐指数:⭐⭐⭐⭐⭐
涨点效果:⭐⭐⭐⭐⭐
专栏回顾:YOLOv8改进系列专栏——本专栏持续复习各种顶会内容——科研必备
训练结果对比图->
目录
一、本文介绍
二、EfficientNetV1的框架原理
三、EfficientNetV1的核心代码
四、手把手教你添加EfficientNetV1机制
修改一
修改二
修改三
修改四
修改五
修改六
修改七
修改八
五、EfficientNetV1的yaml文件
六、成功运行记录
七、本文总结
二、EfficientNetV1的框架原理
官方论文地址: 官方论文地址点击即可跳转
官方代码地址: 官方代码地址点击即可跳转
EfficientNetV1的主要思想是通过均衡地缩放网络的深度、宽度和分辨率,以提高卷积神经网络的性能。这种方法采用了一个简单但有效的复合系数,统一调整所有维度。EfficientNet在多个方面优于现有的ConvNets,特别是在ImageNet数据集上,EfficientNet-B7模型在保持较小的大小和更快的推理速度的同时,达到了84.3%的顶级准确率。此外,EfficientNet还在CIFAR-100和Flowers等其他数据集上展示了出色的迁移学习性能,参数数量大大减少。
总结:EfficientNetV1的主要创新为提出了一种新的模型缩放方法,该方法使用一个复合系数来统一地缩放网络的深度、宽度和分辨率,实现更均衡的网络扩展
这张图展示了EfficientNet提出的模型缩放方法。图中(a)表示基线网络,而图(b)-(d)表示传统的缩放方法,只增加网络的一个维度:宽度、深度或分辨率。图(e)展示了EfficientNet的创新之处,即复合缩放方法,它使用固定比例同时均匀地缩放网络的所有三个维度。
三、EfficientNetV1的核心代码
import re
import math
import collections
from functools import partial
import torch
from torch import nn
from torch.nn import functional as F
from torch.utils import model_zoo# Parameters for the entire model (stem, all blocks, and head)
GlobalParams = collections.namedtuple('GlobalParams', ['width_coefficient', 'depth_coefficient', 'image_size', 'dropout_rate','num_classes', 'batch_norm_momentum', 'batch_norm_epsilon','drop_connect_rate', 'depth_divisor', 'min_depth', 'include_top'])# Parameters for an individual model block
BlockArgs = collections.namedtuple('BlockArgs', ['num_repeat', 'kernel_size', 'stride', 'expand_ratio','input_filters', 'output_filters', 'se_ratio', 'id_skip'])# Set GlobalParams and BlockArgs's defaults
GlobalParams.__new__.__defaults__ = (None,) * len(GlobalParams._fields)
BlockArgs.__new__.__defaults__ = (None,) * len(BlockArgs._fields)# Swish activation function
if hasattr(nn, 'SiLU'):Swish = nn.SiLU
else:# For compatibility with old PyTorch versionsclass Swish(nn.Module):def forward(self, x):return x * torch.sigmoid(x)# A memory-efficient implementation of Swish function
class SwishImplementation(torch.autograd.Function):@staticmethoddef forward(ctx, i):result = i * torch.sigmoid(i)ctx.save_for_backward(i)return result@staticmethoddef backward(ctx, grad_output):i = ctx.saved_tensors[0]sigmoid_i = torch.sigmoid(i)return grad_output * (sigmoid_i * (1 + i * (1 - sigmoid_i)))class MemoryEfficientSwish(nn.Module):def forward(self, x):return SwishImplementation.apply(x)def round_filters(filters, global_params):"""Calculate and round number of filters based on width multiplier.Use width_coefficient, depth_divisor and min_depth of global_params.Args:filters (int): Filters number to be calculated.global_params (namedtuple): Global params of the model.Returns:new_filters: New filters number after calculating."""multiplier = global_params.width_coefficientif not multiplier:return filters# TODO: modify the params names.# maybe the names (width_divisor,min_width)# are more suitable than (depth_divisor,min_depth).divisor = global_params.depth_divisormin_depth = global_params.min_depthfilters *= multipliermin_depth = min_depth or divisor # pay attention to this line when using min_depth# follow the formula transferred from official TensorFlow implementationnew_filters = max(min_depth, int(filters + divisor / 2) // divisor * divisor)if new_filters < 0.9 * filters: # prevent rounding by more than 10%new_filters += divisorreturn int(new_filters)def round_repeats(repeats, global_params):"""Calculate module's repeat number of a block based on depth multiplier.Use depth_coefficient of global_params.Args:repeats (int): num_repeat to be calculated.global_params (namedtuple): Global params of the model.Returns:new repeat: New repeat number after calculating."""multiplier = global_params.depth_coefficientif not multiplier:return repeats# follow the formula transferred from official TensorFlow implementationreturn int(math.ceil(multiplier * repeats))def drop_connect(inputs, p, training):"""Drop connect.Args:input (tensor: BCWH): Input of this structure.p (float: 0.0~1.0): Probability of drop connection.training (bool): The running mode.Returns:output: Output after drop connection."""assert 0 <= p <= 1, 'p must be in range of [0,1]'if not training:return inputsbatch_size = inputs.shape[0]keep_prob = 1 - p# generate binary_tensor mask according to probability (p for 0, 1-p for 1)random_tensor = keep_probrandom_tensor += torch.rand([batch_size, 1, 1, 1], dtype=inputs.dtype, device=inputs.device)binary_tensor = torch.floor(random_tensor)output = inputs / keep_prob * binary_tensorreturn outputdef get_width_and_height_from_size(x):"""Obtain height and width from x.Args:x (int, tuple or list): Data size.Returns:size: A tuple or list (H,W)."""if isinstance(x, int):return x, xif isinstance(x, list) or isinstance(x, tuple):return xelse:raise TypeError()def calculate_output_image_size(input_image_size, stride):"""Calculates the output image size when using Conv2dSamePadding with a stride.Necessary for static padding. Thanks to mannatsingh for pointing this out.Args:input_image_size (int, tuple or list): Size of input image.stride (int, tuple or list): Conv2d operation's stride.Returns:output_image_size: A list [H,W]."""if input_image_size is None:return Noneimage_height, image_width = get_width_and_height_from_size(input_image_size)stride = stride if isinstance(stride, int) else stride[0]image_height = int(math.ceil(image_height / stride))image_width = int(math.ceil(image_width / stride))return [image_height, image_width]# Note:
# The following 'SamePadding' functions make output size equal ceil(input size/stride).
# Only when stride equals 1, can the output size be the same as input size.
# Don't be confused by their function names ! ! !def get_same_padding_conv2d(image_size=None):"""Chooses static padding if you have specified an image size, and dynamic padding otherwise.Static padding is necessary for ONNX exporting of models.Args:image_size (int or tuple): Size of the image.Returns:Conv2dDynamicSamePadding or Conv2dStaticSamePadding."""if image_size is None:return Conv2dDynamicSamePaddingelse:return partial(Conv2dStaticSamePadding, image_size=image_size)class Conv2dDynamicSamePadding(nn.Conv2d):"""2D Convolutions like TensorFlow, for a dynamic image size.The padding is operated in forward function by calculating dynamically."""# Tips for 'SAME' mode padding.# Given the following:# i: width or height# s: stride# k: kernel size# d: dilation# p: padding# Output after Conv2d:# o = floor((i+p-((k-1)*d+1))/s+1)# If o equals i, i = floor((i+p-((k-1)*d+1))/s+1),# => p = (i-1)*s+((k-1)*d+1)-idef __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, groups=1, bias=True):super().__init__(in_channels, out_channels, kernel_size, stride, 0, dilation, groups, bias)self.stride = self.stride if len(self.stride) == 2 else [self.stride[0]] * 2def forward(self, x):ih, iw = x.size()[-2:]kh, kw = self.weight.size()[-2:]sh, sw = self.strideoh, ow = math.ceil(ih / sh), math.ceil(iw / sw) # change the output size according to stride ! ! !pad_h = max((oh - 1) * self.stride[0] + (kh - 1) * self.dilation[0] + 1 - ih, 0)pad_w = max((ow - 1) * self.stride[1] + (kw - 1) * self.dilation[1] + 1 - iw, 0)if pad_h > 0 or pad_w > 0:x = F.pad(x, [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2])return F.conv2d(x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups)class Conv2dStaticSamePadding(nn.Conv2d):"""2D Convolutions like TensorFlow's 'SAME' mode, with the given input image size.The padding mudule is calculated in construction function, then used in forward."""# With the same calculation as Conv2dDynamicSamePaddingdef __init__(self, in_channels, out_channels, kernel_size, stride=1, image_size=None, **kwargs):super().__init__(in_channels, out_channels, kernel_size, stride, **kwargs)self.stride = self.stride if len(self.stride) == 2 else [self.stride[0]] * 2# Calculate padding based on image size and save itassert image_size is not Noneih, iw = (image_size, image_size) if isinstance(image_size, int) else image_sizekh, kw = self.weight.size()[-2:]sh, sw = self.strideoh, ow = math.ceil(ih / sh), math.ceil(iw / sw)pad_h = max((oh - 1) * self.stride[0] + (kh - 1) * self.dilation[0] + 1 - ih, 0)pad_w = max((ow - 1) * self.stride[1] + (kw - 1) * self.dilation[1] + 1 - iw, 0)if pad_h > 0 or pad_w > 0:self.static_padding = nn.ZeroPad2d((pad_w // 2, pad_w - pad_w // 2,pad_h // 2, pad_h - pad_h // 2))else:self.static_padding = nn.Identity()def forward(self, x):x = self.static_padding(x)x = F.conv2d(x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups)return xdef get_same_padding_maxPool2d(image_size=None):"""Chooses static padding if you have specified an image size, and dynamic padding otherwise.Static padding is necessary for ONNX exporting of models.Args:image_size (int or tuple): Size of the image.Returns:MaxPool2dDynamicSamePadding or MaxPool2dStaticSamePadding."""if image_size is None:return MaxPool2dDynamicSamePaddingelse:return partial(MaxPool2dStaticSamePadding, image_size=image_size)class MaxPool2dDynamicSamePadding(nn.MaxPool2d):"""2D MaxPooling like TensorFlow's 'SAME' mode, with a dynamic image size.The padding is operated in forward function by calculating dynamically."""def __init__(self, kernel_size, stride, padding=0, dilation=1, return_indices=False, ceil_mode=False):super().__init__(kernel_size, stride, padding, dilation, return_indices, ceil_mode)self.stride = [self.stride] * 2 if isinstance(self.stride, int) else self.strideself.kernel_size = [self.kernel_size] * 2 if isinstance(self.kernel_size, int) else self.kernel_sizeself.dilation = [self.dilation] * 2 if isinstance(self.dilation, int) else self.dilationdef forward(self, x):ih, iw = x.size()[-2:]kh, kw = self.kernel_sizesh, sw = self.strideoh, ow = math.ceil(ih / sh), math.ceil(iw / sw)pad_h = max((oh - 1) * self.stride[0] + (kh - 1) * self.dilation[0] + 1 - ih, 0)pad_w = max((ow - 1) * self.stride[1] + (kw - 1) * self.dilation[1] + 1 - iw, 0)if pad_h > 0 or pad_w > 0:x = F.pad(x, [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2])return F.max_pool2d(x, self.kernel_size, self.stride, self.padding,self.dilation, self.ceil_mode, self.return_indices)class MaxPool2dStaticSamePadding(nn.MaxPool2d):"""2D MaxPooling like TensorFlow's 'SAME' mode, with the given input image size.The padding mudule is calculated in construction function, then used in forward."""def __init__(self, kernel_size, stride, image_size=None, **kwargs):super().__init__(kernel_size, stride, **kwargs)self.stride = [self.stride] * 2 if isinstance(self.stride, int) else self.strideself.kernel_size = [self.kernel_size] * 2 if isinstance(self.kernel_size, int) else self.kernel_sizeself.dilation = [self.dilation] * 2 if isinstance(self.dilation, int) else self.dilation# Calculate padding based on image size and save itassert image_size is not Noneih, iw = (image_size, image_size) if isinstance(image_size, int) else image_sizekh, kw = self.kernel_sizesh, sw = self.strideoh, ow = math.ceil(ih / sh), math.ceil(iw / sw)pad_h = max((oh - 1) * self.stride[0] + (kh - 1) * self.dilation[0] + 1 - ih, 0)pad_w = max((ow - 1) * self.stride[1] + (kw - 1) * self.dilation[1] + 1 - iw, 0)if pad_h > 0 or pad_w > 0:self.static_padding = nn.ZeroPad2d((pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2))else:self.static_padding = nn.Identity()def forward(self, x):x = self.static_padding(x)x = F.max_pool2d(x, self.kernel_size, self.stride, self.padding,self.dilation, self.ceil_mode, self.return_indices)return x################################################################################
# Helper functions for loading model params
################################################################################# BlockDecoder: A Class for encoding and decoding BlockArgs
# efficientnet_params: A function to query compound coefficient
# get_model_params and efficientnet:
# Functions to get BlockArgs and GlobalParams for efficientnet
# url_map and url_map_advprop: Dicts of url_map for pretrained weights
# load_pretrained_weights: A function to load pretrained weightsclass BlockDecoder(object):"""Block Decoder for readability,straight from the official TensorFlow repository."""@staticmethoddef _decode_block_string(block_string):"""Get a block through a string notation of arguments.Args:block_string (str): A string notation of arguments.Examples: 'r1_k3_s11_e1_i32_o16_se0.25_noskip'.Returns:BlockArgs: The namedtuple defined at the top of this file."""assert isinstance(block_string, str)ops = block_string.split('_')options = {}for op in ops:splits = re.split(r'(\d.*)', op)if len(splits) >= 2:key, value = splits[:2]options[key] = value# Check strideassert (('s' in options and len(options['s']) == 1) or(len(options['s']) == 2 and options['s'][0] == options['s'][1]))return BlockArgs(num_repeat=int(options['r']),kernel_size=int(options['k']),stride=[int(options['s'][0])],expand_ratio=int(options['e']),input_filters=int(options['i']),output_filters=int(options['o']),se_ratio=float(options['se']) if 'se' in options else None,id_skip=('noskip' not in block_string))@staticmethoddef _encode_block_string(block):"""Encode a block to a string.Args:block (namedtuple): A BlockArgs type argument.Returns:block_string: A String form of BlockArgs."""args = ['r%d' % block.num_repeat,'k%d' % block.kernel_size,'s%d%d' % (block.strides[0], block.strides[1]),'e%s' % block.expand_ratio,'i%d' % block.input_filters,'o%d' % block.output_filters]if 0 < block.se_ratio <= 1:args.append('se%s' % block.se_ratio)if block.id_skip is False:args.append('noskip')return '_'.join(args)@staticmethoddef decode(string_list):"""Decode a list of string notations to specify blocks inside the network.Args:string_list (list[str]): A list of strings, each string is a notation of block.Returns:blocks_args: A list of BlockArgs namedtuples of block args."""assert isinstance(string_list, list)blocks_args = []for block_string in string_list:blocks_args.append(BlockDecoder._decode_block_string(block_string))return blocks_args@staticmethoddef encode(blocks_args):"""Encode a list of BlockArgs to a list of strings.Args:blocks_args (list[namedtuples]): A list of BlockArgs namedtuples of block args.Returns:block_strings: A list of strings, each string is a notation of block."""block_strings = []for block in blocks_args:block_strings.append(BlockDecoder._encode_block_string(block))return block_stringsdef efficientnet_params(model_name):"""Map EfficientNet model name to parameter coefficients.Args:model_name (str): Model name to be queried.Returns:params_dict[model_name]: A (width,depth,res,dropout) tuple."""params_dict = {# Coefficients: width,depth,res,dropout'efficientnet-b0': (1.0, 1.0, 224, 0.2),'efficientnet-b1': (1.0, 1.1, 240, 0.2),'efficientnet-b2': (1.1, 1.2, 260, 0.3),'efficientnet-b3': (1.2, 1.4, 300, 0.3),'efficientnet-b4': (1.4, 1.8, 380, 0.4),'efficientnet-b5': (1.6, 2.2, 456, 0.4),'efficientnet-b6': (1.8, 2.6, 528, 0.5),'efficientnet-b7': (2.0, 3.1, 600, 0.5),'efficientnet-b8': (2.2, 3.6, 672, 0.5),'efficientnet-l2': (4.3, 5.3, 800, 0.5),}return params_dict[model_name]def efficientnet(width_coefficient=None, depth_coefficient=None, image_size=None,dropout_rate=0.2, drop_connect_rate=0.2, num_classes=1000, include_top=True):"""Create BlockArgs and GlobalParams for efficientnet model.Args:width_coefficient (float)depth_coefficient (float)image_size (int)dropout_rate (float)drop_connect_rate (float)num_classes (int)Meaning as the name suggests.Returns:blocks_args, global_params."""# Blocks args for the whole model(efficientnet-b0 by default)# It will be modified in the construction of EfficientNet Class according to modelblocks_args = ['r1_k3_s11_e1_i32_o16_se0.25','r2_k3_s22_e6_i16_o24_se0.25','r2_k5_s22_e6_i24_o40_se0.25','r3_k3_s22_e6_i40_o80_se0.25','r3_k5_s11_e6_i80_o112_se0.25','r4_k5_s22_e6_i112_o192_se0.25','r1_k3_s11_e6_i192_o320_se0.25',]blocks_args = BlockDecoder.decode(blocks_args)global_params = GlobalParams(width_coefficient=width_coefficient,depth_coefficient=depth_coefficient,image_size=image_size,dropout_rate=dropout_rate,num_classes=num_classes,batch_norm_momentum=0.99,batch_norm_epsilon=1e-3,drop_connect_rate=drop_connect_rate,depth_divisor=8,min_depth=None,include_top=include_top,)return blocks_args, global_paramsdef get_model_params(model_name, override_params):"""Get the block args and global params for a given model name.Args:model_name (str): Model's name.override_params (dict): A dict to modify global_params.Returns:blocks_args, global_params"""if model_name.startswith('efficientnet'):w, d, s, p = efficientnet_params(model_name)# note: all models have drop connect rate = 0.2blocks_args, global_params = efficientnet(width_coefficient=w, depth_coefficient=d, dropout_rate=p, image_size=s)else:raise NotImplementedError('model name is not pre-defined: {}'.format(model_name))if override_params:# ValueError will be raised here if override_params has fields not included in global_params.global_params = global_params._replace(**override_params)return blocks_args, global_params# train with Standard methods
# check more details in paper(EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks)
url_map = {'efficientnet-b0': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b0-355c32eb.pth','efficientnet-b1': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b1-f1951068.pth','efficientnet-b2': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b2-8bb594d6.pth','efficientnet-b3': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b3-5fb5a3c3.pth','efficientnet-b4': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b4-6ed6700e.pth','efficientnet-b5': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b5-b6417697.pth','efficientnet-b6': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b6-c76e70fd.pth','efficientnet-b7': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b7-dcc49843.pth',
}# train with Adversarial Examples(AdvProp)
# check more details in paper(Adversarial Examples Improve Image Recognition)
url_map_advprop = {'efficientnet-b0': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b0-b64d5a18.pth','efficientnet-b1': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b1-0f3ce85a.pth','efficientnet-b2': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b2-6e9d97e5.pth','efficientnet-b3': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b3-cdd7c0f4.pth','efficientnet-b4': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b4-44fb3a87.pth','efficientnet-b5': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b5-86493f6b.pth','efficientnet-b6': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b6-ac80338e.pth','efficientnet-b7': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b7-4652b6dd.pth','efficientnet-b8': 'https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/adv-efficientnet-b8-22a8fe65.pth',
}# TODO: add the petrained weights url map of 'efficientnet-l2'def load_pretrained_weights(model, model_name, weights_path=None, load_fc=True, advprop=False, verbose=True):"""Loads pretrained weights from weights path or download using url.Args:model (Module): The whole model of efficientnet.model_name (str): Model name of efficientnet.weights_path (None or str):str: path to pretrained weights file on the local disk.None: use pretrained weights downloaded from the Internet.load_fc (bool): Whether to load pretrained weights for fc layer at the end of the model.advprop (bool): Whether to load pretrained weightstrained with advprop (valid when weights_path is None)."""if isinstance(weights_path, str):state_dict = torch.load(weights_path)else:# AutoAugment or Advprop (different preprocessing)url_map_ = url_map_advprop if advprop else url_mapstate_dict = model_zoo.load_url(url_map_[model_name])if load_fc:ret = model.load_state_dict(state_dict, strict=False)assert not ret.missing_keys, 'Missing keys when loading pretrained weights: {}'.format(ret.missing_keys)else:state_dict.pop('_fc.weight')state_dict.pop('_fc.bias')ret = model.load_state_dict(state_dict, strict=False)assert set(ret.missing_keys) == set(['_fc.weight', '_fc.bias']), 'Missing keys when loading pretrained weights: {}'.format(ret.missing_keys)assert not ret.unexpected_keys, 'Missing keys when loading pretrained weights: {}'.format(ret.unexpected_keys)if verbose:print('Loaded pretrained weights for {}'.format(model_name))VALID_MODELS = ('efficientnet-b0', 'efficientnet-b1', 'efficientnet-b2', 'efficientnet-b3','efficientnet-b4', 'efficientnet-b5', 'efficientnet-b6', 'efficientnet-b7','efficientnet-b8',# Support the construction of 'efficientnet-l2' without pretrained weights'efficientnet-l2'
)class MBConvBlock(nn.Module):"""Mobile Inverted Residual Bottleneck Block.Args:block_args (namedtuple): BlockArgs, defined in utils.py.global_params (namedtuple): GlobalParam, defined in utils.py.image_size (tuple or list): [image_height, image_width].References:[1] https://arxiv.org/abs/1704.04861 (MobileNet v1)[2] https://arxiv.org/abs/1801.04381 (MobileNet v2)[3] https://arxiv.org/abs/1905.02244 (MobileNet v3)"""def __init__(self, block_args, global_params, image_size=None):super().__init__()self._block_args = block_argsself._bn_mom = 1 - global_params.batch_norm_momentum # pytorch's difference from tensorflowself._bn_eps = global_params.batch_norm_epsilonself.has_se = (self._block_args.se_ratio is not None) and (0 < self._block_args.se_ratio <= 1)self.id_skip = block_args.id_skip # whether to use skip connection and drop connect# Expansion phase (Inverted Bottleneck)inp = self._block_args.input_filters # number of input channelsoup = self._block_args.input_filters * self._block_args.expand_ratio # number of output channelsif self._block_args.expand_ratio != 1:Conv2d = get_same_padding_conv2d(image_size=image_size)self._expand_conv = Conv2d(in_channels=inp, out_channels=oup, kernel_size=1, bias=False)self._bn0 = nn.BatchNorm2d(num_features=oup, momentum=self._bn_mom, eps=self._bn_eps)# image_size = calculate_output_image_size(image_size, 1) <-- this wouldn't modify image_size# Depthwise convolution phasek = self._block_args.kernel_sizes = self._block_args.strideConv2d = get_same_padding_conv2d(image_size=image_size)self._depthwise_conv = Conv2d(in_channels=oup, out_channels=oup, groups=oup, # groups makes it depthwisekernel_size=k, stride=s, bias=False)self._bn1 = nn.BatchNorm2d(num_features=oup, momentum=self._bn_mom, eps=self._bn_eps)image_size = calculate_output_image_size(image_size, s)# Squeeze and Excitation layer, if desiredif self.has_se:Conv2d = get_same_padding_conv2d(image_size=(1, 1))num_squeezed_channels = max(1, int(self._block_args.input_filters * self._block_args.se_ratio))self._se_reduce = Conv2d(in_channels=oup, out_channels=num_squeezed_channels, kernel_size=1)self._se_expand = Conv2d(in_channels=num_squeezed_channels, out_channels=oup, kernel_size=1)# Pointwise convolution phasefinal_oup = self._block_args.output_filtersConv2d = get_same_padding_conv2d(image_size=image_size)self._project_conv = Conv2d(in_channels=oup, out_channels=final_oup, kernel_size=1, bias=False)self._bn2 = nn.BatchNorm2d(num_features=final_oup, momentum=self._bn_mom, eps=self._bn_eps)self._swish = MemoryEfficientSwish()def forward(self, inputs, drop_connect_rate=None):"""MBConvBlock's forward function.Args:inputs (tensor): Input tensor.drop_connect_rate (bool): Drop connect rate (float, between 0 and 1).Returns:Output of this block after processing."""# Expansion and Depthwise Convolutionx = inputsif self._block_args.expand_ratio != 1:x = self._expand_conv(inputs)x = self._bn0(x)x = self._swish(x)x = self._depthwise_conv(x)x = self._bn1(x)x = self._swish(x)# Squeeze and Excitationif self.has_se:x_squeezed = F.adaptive_avg_pool2d(x, 1)x_squeezed = self._se_reduce(x_squeezed)x_squeezed = self._swish(x_squeezed)x_squeezed = self._se_expand(x_squeezed)x = torch.sigmoid(x_squeezed) * x# Pointwise Convolutionx = self._project_conv(x)x = self._bn2(x)# Skip connection and drop connectinput_filters, output_filters = self._block_args.input_filters, self._block_args.output_filtersif self.id_skip and self._block_args.stride == 1 and input_filters == output_filters:# The combination of skip connection and drop connect brings about stochastic depth.if drop_connect_rate:x = drop_connect(x, p=drop_connect_rate, training=self.training)x = x + inputs # skip connectionreturn xdef set_swish(self, memory_efficient=True):"""Sets swish function as memory efficient (for training) or standard (for export).Args:memory_efficient (bool): Whether to use memory-efficient version of swish."""self._swish = MemoryEfficientSwish() if memory_efficient else Swish()class EfficientNet(nn.Module):def __init__(self, blocks_args=None, global_params=None):super().__init__()assert isinstance(blocks_args, list), 'blocks_args should be a list'assert len(blocks_args) > 0, 'block args must be greater than 0'self._global_params = global_paramsself._blocks_args = blocks_args# Batch norm parametersbn_mom = 1 - self._global_params.batch_norm_momentumbn_eps = self._global_params.batch_norm_epsilon# Get stem static or dynamic convolution depending on image sizeimage_size = global_params.image_sizeConv2d = get_same_padding_conv2d(image_size=image_size)# Stemin_channels = 3 # rgbout_channels = round_filters(32, self._global_params) # number of output channelsself._conv_stem = Conv2d(in_channels, out_channels, kernel_size=3, stride=2, bias=False)self._bn0 = nn.BatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)image_size = calculate_output_image_size(image_size, 2)# Build blocksself._blocks = nn.ModuleList([])for block_args in self._blocks_args:# Update block input and output filters based on depth multiplier.block_args = block_args._replace(input_filters=round_filters(block_args.input_filters, self._global_params),output_filters=round_filters(block_args.output_filters, self._global_params),num_repeat=round_repeats(block_args.num_repeat, self._global_params))# The first block needs to take care of stride and filter size increase.self._blocks.append(MBConvBlock(block_args, self._global_params, image_size=image_size))image_size = calculate_output_image_size(image_size, block_args.stride)if block_args.num_repeat > 1: # modify block_args to keep same output sizeblock_args = block_args._replace(input_filters=block_args.output_filters, stride=1)for _ in range(block_args.num_repeat - 1):self._blocks.append(MBConvBlock(block_args, self._global_params, image_size=image_size))# image_size = calculate_output_image_size(image_size, block_args.stride) # stride = 1# Headin_channels = block_args.output_filters # output of final blockout_channels = round_filters(1280, self._global_params)Conv2d = get_same_padding_conv2d(image_size=image_size)self._conv_head = Conv2d(in_channels, out_channels, kernel_size=1, bias=False)self._bn1 = nn.BatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)# Final linear layerself._avg_pooling = nn.AdaptiveAvgPool2d(1)if self._global_params.include_top:self._dropout = nn.Dropout(self._global_params.dropout_rate)self._fc = nn.Linear(out_channels, self._global_params.num_classes)# set activation to memory efficient swish by defaultself._swish = MemoryEfficientSwish()self.width_list = [i.size(1) for i in self.forward(torch.randn(1, 3, 640, 640))]def set_swish(self, memory_efficient=True):"""Sets swish function as memory efficient (for training) or standard (for export).Args:memory_efficient (bool): Whether to use memory-efficient version of swish."""self._swish = MemoryEfficientSwish() if memory_efficient else Swish()for block in self._blocks:block.set_swish(memory_efficient)def extract_endpoints(self, inputs):# """Use convolution layer to extract features# from reduction levels i in [1, 2, 3, 4, 5].## Args:# inputs (tensor): Input tensor.## Returns:# Dictionary of last intermediate features# with reduction levels i in [1, 2, 3, 4, 5].# Example:# >>> import torch# >>> from efficientnet.model import EfficientNet# >>> inputs = torch.rand(1, 3, 224, 224)# >>> model = EfficientNet.from_pretrained('efficientnet-b0')# >>> endpoints = model.extract_endpoints(inputs)# >>> print(endpoints['reduction_1'].shape) # torch.Size([1, 16, 112, 112])# >>> print(endpoints['reduction_2'].shape) # torch.Size([1, 24, 56, 56])# >>> print(endpoints['reduction_3'].shape) # torch.Size([1, 40, 28, 28])# >>> print(endpoints['reduction_4'].shape) # torch.Size([1, 112, 14, 14])# >>> print(endpoints['reduction_5'].shape) # torch.Size([1, 320, 7, 7])# >>> print(endpoints['reduction_6'].shape) # torch.Size([1, 1280, 7, 7])# """endpoints = dict()# Stemx = self._swish(self._bn0(self._conv_stem(inputs)))prev_x = x# Blocksfor idx, block in enumerate(self._blocks):drop_connect_rate = self._global_params.drop_connect_rateif drop_connect_rate:drop_connect_rate *= float(idx) / len(self._blocks) # scale drop connect_ratex = block(x, drop_connect_rate=drop_connect_rate)if prev_x.size(2) > x.size(2):endpoints['reduction_{}'.format(len(endpoints) + 1)] = prev_xelif idx == len(self._blocks) - 1:endpoints['reduction_{}'.format(len(endpoints) + 1)] = xprev_x = x# Headx = self._swish(self._bn1(self._conv_head(x)))endpoints['reduction_{}'.format(len(endpoints) + 1)] = xreturn endpointsdef forward(self, inputs):"""use convolution layer to extract feature .Args:inputs (tensor): Input tensor.Returns:Output of the final convolutionlayer in the efficientnet model."""# Stemx = self._swish(self._bn0(self._conv_stem(inputs)))unique_tensors = {}# Blocksfor idx, block in enumerate(self._blocks):drop_connect_rate = self._global_params.drop_connect_rateif drop_connect_rate:drop_connect_rate *= float(idx) / len(self._blocks) # scale drop connect_ratex = block(x, drop_connect_rate=drop_connect_rate)width, height = x.shape[2], x.shape[3]unique_tensors[(width, height)] = xresult_list = list(unique_tensors.values())[-4:]# Headreturn result_list@classmethoddef from_name(cls, model_name, in_channels=3, **override_params):"""Create an efficientnet model according to name.Args:model_name (str): Name for efficientnet.in_channels (int): Input data's channel number.override_params (other key word params):Params to override model's global_params.Optional key:'width_coefficient', 'depth_coefficient','image_size', 'dropout_rate','num_classes', 'batch_norm_momentum','batch_norm_epsilon', 'drop_connect_rate','depth_divisor', 'min_depth'Returns:An efficientnet model."""cls._check_model_name_is_valid(model_name)blocks_args, global_params = get_model_params(model_name, override_params)model = cls(blocks_args, global_params)model._change_in_channels(in_channels)return model@classmethoddef from_pretrained(cls, model_name, weights_path=None, advprop=False,in_channels=3, num_classes=1000, **override_params):"""Create an efficientnet model according to name.Args:model_name (str): Name for efficientnet.weights_path (None or str):str: path to pretrained weights file on the local disk.None: use pretrained weights downloaded from the Internet.advprop (bool):Whether to load pretrained weightstrained with advprop (valid when weights_path is None).in_channels (int): Input data's channel number.num_classes (int):Number of categories for classification.It controls the output size for final linear layer.override_params (other key word params):Params to override model's global_params.Optional key:'width_coefficient', 'depth_coefficient','image_size', 'dropout_rate','batch_norm_momentum','batch_norm_epsilon', 'drop_connect_rate','depth_divisor', 'min_depth'Returns:A pretrained efficientnet model."""model = cls.from_name(model_name, num_classes=num_classes, **override_params)load_pretrained_weights(model, model_name, weights_path=weights_path,load_fc=(num_classes == 1000), advprop=advprop)model._change_in_channels(in_channels)return model@classmethoddef get_image_size(cls, model_name):"""Get the input image size for a given efficientnet model.Args:model_name (str): Name for efficientnet.Returns:Input image size (resolution)."""cls._check_model_name_is_valid(model_name)_, _, res, _ = efficientnet_params(model_name)return res@classmethoddef _check_model_name_is_valid(cls, model_name):"""Validates model name.Args:model_name (str): Name for efficientnet.Returns:bool: Is a valid name or not."""if model_name not in VALID_MODELS:raise ValueError('model_name should be one of: ' + ', '.join(VALID_MODELS))def _change_in_channels(self, in_channels):"""Adjust model's first convolution layer to in_channels, if in_channels not equals 3.Args:in_channels (int): Input data's channel number."""if in_channels != 3:Conv2d = get_same_padding_conv2d(image_size=self._global_params.image_size)out_channels = round_filters(32, self._global_params)self._conv_stem = Conv2d(in_channels, out_channels, kernel_size=3, stride=2, bias=False)def efficient(model_name='efficientnet-b0', pretrained=False):if pretrained:model = EfficientNet.from_pretrained('{}'.format(model_name))else:model = EfficientNet.from_name('{}'.format(model_name))return modelif __name__ == "__main__":# VALID_MODELS = (# 'efficientnet-b0', 'efficientnet-b1', 'efficientnet-b2', 'efficientnet-b3',# 'efficientnet-b4', 'efficientnet-b5', 'efficientnet-b6', 'efficientnet-b7',# 'efficientnet-b8',# # Support the construction of 'efficientnet-l2' without pretrained weights# 'efficientnet-l2'# )# Generating Sample imageimage_size = (1, 3, 640, 640)image = torch.rand(*image_size)# Modelmodel = efficient('efficientnet-b0')out = model(image)print(len(out))
四、手把手教你添加EfficientNetV1机制
这个主干的网络结构添加起来算是所有的改进机制里最麻烦的了,因为有一些网略结构可以用yaml文件搭建出来,有一些网络结构其中的一些细节根本没有办法用yaml文件去搭建,用yaml文件去搭建会损失一些细节部分(而且一个网络结构设计很多细节的结构修改方式都不一样,一个一个去修改大家难免会出错),所以这里让网络直接返回整个网络,然后修改部分 yolo代码以后就都以这种形式添加了,以后我提出的网络模型基本上都会通过这种方式修改,我也会进行一些模型细节改进。创新出新的网络结构大家直接拿来用就可以的。下面开始添加教程->
(同时每一个后面都有代码,大家拿来复制粘贴替换即可,但是要看好了不要复制粘贴替换多了)
修改一
我们复制网络结构代码到“ultralytics/nn/modules”目录下创建一个py文件复制粘贴进去 ,我这里起的名字是EfficientNetV1
修改二
找到如下的文件"ultralytics/nn/tasks.py" 在开始的部分导入我们的模型如下图。
from .modules.EfficientV1 import efficient
修改三
添加如下两行代码!!!
修改四
找到七百多行大概把具体看图片,按照图片来修改就行,添加红框内的部分,注意没有()只是函数名,我这里只添加了部分的版本,大家有兴趣这个EfficientNetV1还有更多的版本可以添加,看我给的代码函数头即可。
elif m in {自行添加对应的模型即可,下面都是一样的}:m = m()c2 = m.width_list # 返回通道列表backbone = True
修改五
下面的两个红框内都是需要改动的。
if isinstance(c2, list):m_ = mm_.backbone = Trueelse:m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args) # modulet = str(m)[8:-2].replace('__main__.', '') # module typem.np = sum(x.numel() for x in m_.parameters()) # number paramsm_.i, m_.f, m_.type = i + 4 if backbone else i, f, t # attach index, 'from' index, type
修改六
如下的也需要修改,全部按照我的来。
代码如下把原先的代码替换了即可。
if verbose:LOGGER.info(f'{i:>3}{str(f):>20}{n_:>3}{m.np:10.0f} {t:<45}{str(args):<30}') # printsave.extend(x % (i + 4 if backbone else i) for x in ([f] if isinstance(f, int) else f) if x != -1) # append to savelistlayers.append(m_)if i == 0:ch = []if isinstance(c2, list):ch.extend(c2)if len(c2) != 5:ch.insert(0, 0)else:ch.append(c2)
修改七
修改七和前面的都不太一样,需要修改前向传播中的一个部分, 已经离开了parse_model方法了。
可以在图片中开代码行数,没有离开task.py文件都是同一个文件。 同时这个部分有好几个前向传播都很相似,大家不要看错了,是70多行左右的!!!,同时我后面提供了代码,大家直接复制粘贴即可,有时间我针对这里会出一个视频。
代码如下->
def _predict_once(self, x, profile=False, visualize=False):"""Perform a forward pass through the network.Args:x (torch.Tensor): The input tensor to the model.profile (bool): Print the computation time of each layer if True, defaults to False.visualize (bool): Save the feature maps of the model if True, defaults to False.Returns:(torch.Tensor): The last output of the model."""y, dt = [], [] # outputsfor m in self.model:if m.f != -1: # if not from previous layerx = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f] # from earlier layersif profile:self._profile_one_layer(m, x, dt)if hasattr(m, 'backbone'):x = m(x)if len(x) != 5: # 0 - 5x.insert(0, None)for index, i in enumerate(x):if index in self.save:y.append(i)else:y.append(None)x = x[-1] # 最后一个输出传给下一层else:x = m(x) # runy.append(x if m.i in self.save else None) # save outputif visualize:feature_visualization(x, m.type, m.i, save_dir=visualize)return x
到这里就完成了修改部分,但是这里面细节很多,大家千万要注意不要替换多余的代码,导致报错,也不要拉下任何一部,都会导致运行失败,而且报错很难排查!!!很难排查!!!
修改八
我们找到如下文件'ultralytics/utils/torch_utils.py'按照如下的图片进行修改,否则容易打印不出来计算量。
五、EfficientNetV1的yaml文件
复制如下yaml文件进行运行!!!
# Ultralytics YOLO 🚀, AGPL-3.0 license
# YOLOv8 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect# Parameters
nc: 80 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov8n.yaml' will call yolov8.yaml with scale 'n'# [depth, width, max_channels]n: [0.33, 0.25, 1024] # YOLOv8n summary: 225 layers, 3157200 parameters, 3157184 gradients, 8.9 GFLOPss: [0.33, 0.50, 1024] # YOLOv8s summary: 225 layers, 11166560 parameters, 11166544 gradients, 28.8 GFLOPsm: [0.67, 0.75, 768] # YOLOv8m summary: 295 layers, 25902640 parameters, 25902624 gradients, 79.3 GFLOPsl: [1.00, 1.00, 512] # YOLOv8l summary: 365 layers, 43691520 parameters, 43691504 gradients, 165.7 GFLOPsx: [1.00, 1.25, 512] # YOLOv8x summary: 365 layers, 68229648 parameters, 68229632 gradients, 258.5 GFLOP# YOLOv8.0n backbone
backbone:# [from, repeats, module, args]- [-1, 1, efficient, []] # 4- [-1, 1, SPPF, [1024, 5]] # 5# YOLOv8.0n head
head:- [-1, 1, nn.Upsample, [None, 2, 'nearest']] # 6- [[-1, 3], 1, Concat, [1]] # 7 cat backbone P4- [-1, 3, C2f, [512]] # 8- [-1, 1, nn.Upsample, [None, 2, 'nearest']] # 9- [[-1, 2], 1, Concat, [1]] # 10 cat backbone P3- [-1, 3, C2f, [256]] # 11 (P3/8-small)- [-1, 1, Conv, [256, 3, 2]] # 12- [[-1, 8], 1, Concat, [1]] # 13 cat head P4- [-1, 3, C2f, [512]] # 14 (P4/16-medium)- [-1, 1, Conv, [512, 3, 2]] # 15- [[-1, 5], 1, Concat, [1]] # 16 cat head P5- [-1, 3, C2f, [1024]] # 17 (P5/32-large)- [[11, 14, 17], 1, Detect, [nc]] # Detect(P3, P4, P5)
六、成功运行记录
下面是成功运行的截图,已经完成了有1个epochs的训练,图片太大截不全第2个epochs了。
七、本文总结
到此本文的正式分享内容就结束了,在这里给大家推荐我的YOLOv8改进有效涨点专栏,本专栏目前为新开的平均质量分98分,后期我会根据各种最新的前沿顶会进行论文复现,也会对一些老的改进机制进行补充,目前本专栏免费阅读(暂时,大家尽早关注不迷路~),如果大家觉得本文帮助到你了,订阅本专栏,关注后续更多的更新~
专栏回顾:YOLOv8改进系列专栏——本专栏持续复习各种顶会内容——科研必备