姿态估计1-06：FSA-Net(头部姿态估算)-源码无死角讲解（1）-训练代码总览

以下链接是个人关于FSA-Net(头部姿态估算) 所有见解，如有错误欢迎大家指出，我会第一时间纠正。有兴趣的朋友可以加微信：17575010159 相互讨论技术。若是帮助到了你什么，一定要记得点赞！因为这是对我最大的鼓励。文末附带\color{blue}{文末附带}文末附带公众号−\color{blue}{公众号 -}公众号−海量资源。\color{blue}{ 海量资源}。海量资源。

姿态估计1-00：FSA-Net(头部姿态估算)-目录-史上最新无死角讲解

分析前言

我相信，大家跟到这里来，说明你以及看完论文了，既然如此，我们在来看看training_and_testing/run_fsanet_train.sh文件，这是我们训练的脚本内容如下：

.......

是的，我这里是空的，因为我觉得复制出来太臃肿，显得我的博客不够帅气，所以就不复制出来了，你看自己源码的即可。该文件可以看到如下字样的注释：

# Train on protocal 1
# SSRNET_MT
# FSANET_Capsule
# FSANET_Netvlad
# FSANET_Metric# Train on protocal 2
# SSRNET_MT
# FSANET_Capsule
# FSANET_Netvlad
# FSANET_Metric# Fine-tuned on BIWI with synhead pre-trained model

看完了论文的朋友应该就比较熟悉了，因为在论文中存在如下图示：

也就是说，作者需要执行多次训练脚本，全是为了为了完成实验的对比。那么我们在分析的时候，当然是选择效果最好的哪个进行分析。说得简单，但是现在我也没办法一眼就知道哪个效果最好，不过通过观察可以知道，他们主要的差别就是在于–model_type的参数不一样，其可以选择1到10之间。那么在分析源码的时候，着重分析其处理过程即可。

源码解析

源码注释：

import os
import sys
sys.path.append('..')
import logging
import argparse
import pandas as pd
import numpy as npfrom lib.FSANET_model import *
from lib.SSRNET_model import *import TYY_callbacks
from TYY_generators import *from keras.utils import np_utils
from keras.utils import plot_model
from keras.optimizers import SGD, Adam
from keras.preprocessing.image import ImageDataGenerator
from keras.callbacks import LearningRateScheduler, ModelCheckpointlogging.basicConfig(level=logging.DEBUG)def load_data_npz(npz_path):d = np.load(npz_path)return d["image"], d["pose"]def mk_dir(dir):try:os.mkdir( dir )except OSError:passdef get_args():parser = argparse.ArgumentParser(description="This script trains the CNN model for head pose estimation.",formatter_class=argparse.ArgumentDefaultsHelpFormatter)parser.add_argument("--batch_size", type=int, default=16,help="batch size")parser.add_argument("--nb_epochs", type=int, default=90,help="number of epochs")parser.add_argument("--validation_split", type=float, default=0.2,help="validation split ratio")parser.add_argument("--model_type", type=int, default=3,help="type of model")parser.add_argument("--db_name", type=str, default='300W_LP',help="type of model")args = parser.parse_args()return argsdef main():# 解析并且赋值相关参数args = get_args()db_name = args.db_namebatch_size = args.batch_sizenb_epochs = args.nb_epochsvalidation_split = args.validation_splitmodel_type = args.model_typeimage_size = 64logging.debug("Loading data...")# 如果训练的数据集为300W_LPif db_name == '300W_LP':# 获得对应的npz文件db_list = ['AFW.npz','AFW_Flip.npz','HELEN.npz','HELEN_Flip.npz','IBUG.npz','IBUG_Flip.npz','LFPW.npz','LFPW_Flip.npz']# 用于保存像素image = []# 用于保存姿态pose = []# 循环加入所有的图片像素，以及对应的姿态for i in range(0,len(db_list)):image_temp, pose_temp = load_data_npz('../data/type1/'+db_list[i])image.append(image_temp)pose.append(pose_temp)# 把链表转化为np数组格式.# 加载完数据之后为[122450, 64, 64, 3]image = np.concatenate(image,0)# 加载完数据之后为[122450, 3]pose = np.concatenate(pose,0)# 对于其角度不在[-99,99]之间的数据，全部剔除掉# we only care the angle between [-99,99] and filter other anglesx_data = []y_data = []print(image.shape)print(pose.shape)for i in range(0,pose.shape[0]):temp_pose = pose[i,:]if np.max(temp_pose)<=99.0 and np.min(temp_pose)>=-99.0:x_data.append(image[i,:,:,:])y_data.append(pose[i,:])x_data = np.array(x_data)y_data = np.array(y_data)print(x_data.shape)print(y_data.shape)elif db_name == 'synhead_noBIWI':image, pose = load_data_npz('../data/synhead/media/jinweig/Data2/synhead2_release/synhead_noBIWI.npz')x_data = imagey_data = pose# 如果训练的数据集为BIWIelif db_name == 'BIWI':image, pose = load_data_npz('../data/BIWI_train.npz')x_train = imagey_train = poseimage_test, pose_test = load_data_npz('../data/BIWI_test.npz')x_test = image_testy_test = pose_testelse:print('db_name is wrong!!!')return# 训练到30ep和60ep会进行学习率衰减start_decay_epoch = [30,60]#优化器optMethod = Adam()# 论文中Stage的数目stage_num = [3,3,3]lambda_d = 1# 输出姿态为yaw, pitch, rollnum_classes = 3# 是否使用最好的方法isFine = False#根据model_type参数 进行模型构建if model_type == 0:model = SSR_net_ori_MT(image_size, num_classes, stage_num, lambda_d)()save_name = 'ssrnet_ori_mt'elif model_type == 1:model = SSR_net_MT(image_size, num_classes, stage_num, lambda_d)()save_name = 'ssrnet_mt'elif model_type == 2:num_capsule = 3dim_capsule = 16routings = 2num_primcaps = 7*3m_dim = 5S_set = [num_capsule, dim_capsule, routings, num_primcaps, m_dim]str_S_set = ''.join('_'+str(x) for x in S_set)model = FSA_net_Capsule(image_size, num_classes, stage_num, lambda_d, S_set)()save_name = 'fsanet_capsule'+str_S_setelif model_type == 3:#num_capsule = 3dim_capsule = 16routings = 2num_primcaps = 7*3m_dim = 5S_set = [num_capsule, dim_capsule, routings, num_primcaps, m_dim]str_S_set = ''.join('_'+str(x) for x in S_set)model = FSA_net_Var_Capsule(image_size, num_classes, stage_num, lambda_d, S_set)()save_name = 'fsanet_var_capsule'+str_S_setelif model_type == 4:num_capsule = 3dim_capsule = 16routings = 2num_primcaps = 8*8*3m_dim = 5S_set = [num_capsule, dim_capsule, routings, num_primcaps, m_dim]str_S_set = ''.join('_'+str(x) for x in S_set)model = FSA_net_noS_Capsule(image_size, num_classes, stage_num, lambda_d, S_set)()save_name = 'fsanet_noS_capsule'+str_S_setelif model_type == 5:num_capsule = 3dim_capsule = 16routings = 2num_primcaps = 7*3m_dim = 5S_set = [num_capsule, dim_capsule, routings, num_primcaps, m_dim]str_S_set = ''.join('_'+str(x) for x in S_set)model = FSA_net_NetVLAD(image_size, num_classes, stage_num, lambda_d, S_set)()save_name = 'fsanet_netvlad'+str_S_setelif model_type == 6:num_capsule = 3dim_capsule = 16routings = 2num_primcaps = 7*3m_dim = 5S_set = [num_capsule, dim_capsule, routings, num_primcaps, m_dim]str_S_set = ''.join('_'+str(x) for x in S_set)model = FSA_net_Var_NetVLAD(image_size, num_classes, stage_num, lambda_d, S_set)()save_name = 'fsanet_var_netvlad'+str_S_setelif model_type == 7:num_capsule = 3dim_capsule = 16routings = 2num_primcaps = 8*8*3m_dim = 5S_set = [num_capsule, dim_capsule, routings, num_primcaps, m_dim]str_S_set = ''.join('_'+str(x) for x in S_set)model = FSA_net_noS_NetVLAD(image_size, num_classes, stage_num, lambda_d, S_set)()save_name = 'fsanet_noS_netvlad'+str_S_setelif model_type == 8:# 论文中num_capsule = 3# 论文中的c’=16dim_capsule = 16# 论文中的stream数目routings = 2# 论文中的n'=7，num_primcaps = 7*3# 论文中的m=5m_dim = 5S_set = [num_capsule, dim_capsule, routings, num_primcaps, m_dim]str_S_set = ''.join('_'+str(x) for x in S_set)model = FSA_net_Metric(image_size, num_classes, stage_num, lambda_d, S_set)()save_name = 'fsanet_metric'+str_S_setelif model_type == 9:num_capsule = 3dim_capsule = 16routings = 2num_primcaps = 7*3m_dim = 5S_set = [num_capsule, dim_capsule, routings, num_primcaps, m_dim]str_S_set = ''.join('_'+str(x) for x in S_set)model = FSA_net_Var_Metric(image_size, num_classes, stage_num, lambda_d, S_set)()save_name = 'fsanet_var_metric'+str_S_setelif model_type == 10:num_capsule = 3dim_capsule = 16routings = 2num_primcaps = 8*8*3m_dim = 5S_set = [num_capsule, dim_capsule, routings, num_primcaps, m_dim]str_S_set = ''.join('_'+str(x) for x in S_set)model = FSA_net_noS_Metric(image_size, num_classes, stage_num, lambda_d, S_set)()save_name = 'fsanet_noS_metric'+str_S_set# 指定模型的优化方法，以及loss（均值绝对误差）计算方式，model.compile(optimizer=optMethod, loss=["mae"],loss_weights=[1])logging.debug("Model summary...")# 计算模型参数，打印模型结构model.count_params()model.summary()logging.debug("Saving model...")# 创建必要的目录，如保存模型的路径等等mk_dir(db_name+"_models")mk_dir(db_name+"_models/"+save_name)mk_dir(db_name+"_checkpoints")# 把模型绘画成图，便于分析（总体结构）plot_model(model, to_file=db_name+"_models/"+save_name+"/"+save_name+".png")# 绘画网络模型的细致结构for i_L,layer in enumerate(model.layers):if i_L >0 and i_L< len(model.layers)-1:if 'pred' not in layer.name and 'caps' != layer.name and 'merge' not in layer.name and 'model' in layer.name:plot_model(layer, to_file=db_name+"_models/"+save_name+"/"+layer.name+".png")# 迭代到指定次数，进行学习率衰减decaylearningrate = TYY_callbacks.DecayLearningRate(start_decay_epoch)# 查看指定路径下的模型知否存在，存在则自动加载该目录下的模型callbacks = [ModelCheckpoint(db_name+"_checkpoints/weights.{epoch:02d}-{val_loss:.2f}.hdf5",monitor="val_loss",verbose=1,save_best_only=True,mode="auto"), decaylearningrate]logging.debug("Running training...")# 如果为'BIWI'数据集,则进行测试集和训练集的划分if db_name != 'BIWI':data_num = len(x_data)indexes = np.arange(data_num)np.random.shuffle(indexes)x_data = x_data[indexes]y_data = y_data[indexes]train_num = int(data_num * (1 - validation_split))x_train = x_data[:train_num]x_test = x_data[train_num:]y_train = y_data[:train_num]y_test = y_data[train_num:]elif db_name == 'BIWI':train_num = np.shape(x_train)[0]# 为模型绑定训练数据，测试数据，并且进行训练（真的是个讨厌的框架，用了pytorch之后，其他的框架越看越难受）hist = model.fit_generator(generator=data_generator_pose(X=x_train, Y=y_train, batch_size=batch_size),steps_per_epoch=train_num // batch_size,validation_data=(x_test, y_test),epochs=nb_epochs, verbose=1,callbacks=callbacks)logging.debug("Saving weights...")model.save_weights(os.path.join(db_name+"_models/"+save_name, save_name+'.h5'), overwrite=True)pd.DataFrame(hist.history).to_hdf(os.path.join(db_name+"_models/"+save_name, 'history_'+save_name+'.h5'), "history")if __name__ == '__main__':main()

360行，行行出状元，总得先有钱。三千大道归一大法，无非就是，加载数据，构建模型，训练数据，保存模型。没了，就这么通俗的总结一下。

通过代码的浏览，可以清楚地知道其复杂的地方，是模型的构建过程，具体细节下篇博客进行讲解。