案例背景:视频识别

假设每次输入是8s的灰度视频,视频帧率为25fps,则视频由200帧图像序列构成.每帧是一副单通道的灰度图像,通过pythonb里面的np.stack(深度拼接)可将200帧拼接成200通道的深度数据.进而送到网络里面去训练.

如果输入图像200通道觉得多,可以对视频进行抽帧,针对具体场景可以随机抽帧或等间隔抽帧.比如这里等间隔抽取40帧.则最后输入视频相当于输入一个40通道的图像数据了.

pytorch对超过三通道数据的加载:

读取视频每一帧,转为array格式,然后依次将每一帧进行深度拼接,最后得到一个40通道的array格式的深度数据,保存到pickle里.

对每个视频都进行上述操作,保存到pickle里.

我这里将火的视频深度数据保存在一个.pkl文件中,一共2504个火的视频,即2504个火的深度数据.

将非火的视频深度数据保存在一个.pkl文件中,一共3985个非火的视频,即3985个非火的深度数据.

数据加载

import torch 
from torch.utils import data
import os
from PIL import Image
import numpy as np
import pickle
 
class Fire_Unfire(data.Dataset):
  def __init__(self,fire_path,unfire_path):
    self.pickle_fire = open(fire_path,'rb')
    self.pickle_unfire = open(unfire_path,'rb')
    
  def __getitem__(self,index):
    if index <2504:
      fire = pickle.load(self.pickle_fire)#高*宽*通道
      fire = fire.transpose(2,0,1)#通道*高*宽
      data = torch.from_numpy(fire)
      label = 1
      return data,label
    elif index>=2504 and index<6489:
      unfire = pickle.load(self.pickle_unfire)
      unfire = unfire.transpose(2,0,1)
      data = torch.from_numpy(unfire)
      label = 0
      return data,label
    
  def __len__(self):
    return 6489

root_path = './datasets/train'
dataset = Fire_Unfire(root_path +'/fire_train.pkl',root_path +'/unfire_train.pkl')
 
#转换成pytorch网络输入的格式(批量大小,通道数,高,宽)
from torch.utils.data import DataLoader
fire_dataloader = DataLoader(dataset,batch_size=4,shuffle=True,drop_last = True)

模型训练

import torch
from torch.utils import data
from nets.mobilenet import mobilenet
from config.config import default_config
from torch.autograd import Variable as V
import numpy as np
import sys
import time
 
opt = default_config()
def train():
  #模型定义
  model = mobilenet().cuda()
  if opt.pretrain_model:
    model.load_state_dict(torch.load(opt.pretrain_model))
  
  #损失函数
  criterion = torch.nn.CrossEntropyLoss().cuda()
  
  #学习率
  lr = opt.lr
  
  #优化器
  optimizer = torch.optim.SGD(model.parameters(),lr = lr,weight_decay=opt.weight_decay)
  
  
  pre_loss = 0.0
  #训练
  for epoch in range(opt.max_epoch):
     #训练数据
    train_data = Fire_Unfire(opt.root_path +'/fire_train.pkl',opt.root_path +'/unfire_train.pkl')
    train_dataloader = data.DataLoader(train_data,batch_size=opt.batch_size,shuffle=True,drop_last = True)
    loss_sum = 0.0
    for i,(datas,labels) in enumerate(train_dataloader):
      #print(i,datas.size(),labels)
      #梯度清零
      optimizer.zero_grad()
      #输入
      input = V(datas.cuda()).float()
      #目标
      target = V(labels.cuda()).long()
      #输出
      score = model(input).cuda()
      #损失
      loss = criterion(score,target)
      loss_sum += loss
      #反向传播
      loss.backward()
      #梯度更新
      optimizer.step()      
    print('{}{}{}{}{}'.format('epoch:',epoch,',','loss:',loss))
    torch.save(model.state_dict(),'models/mobilenet_%d.pth'%(epoch+370))