深度學習——性別識別

import torch

import math

import torch.nn as nn

from torch.autograd import Variable

from torchvision import transforms, models

import argparse

import os

from torch.utils.data import DataLoader

from torch.utils.data import Dataset

import random

class sexnet(nn.Module):

     def __init__(self):

         super(sexnet, self).__init__()

         self.dense = nn.Sequential(

             nn.Linear(2, 2),

         )

     def forward(self, x):

         out = self.dense(x)

         return out

class SexDataset(Dataset):

     def __init__(self, txt, transform=None, target_transform=None):

         fh = open(txt, 'r')

         data = []

         for line in fh:

             line = line.strip('\n')

             line = line.rstrip()

             words = line.split()

             data.append((float(words[0]) / 2.0, float(words[1]) / 80.0, int(words[2])))

         random.shuffle(data)

         self.data = data

     def __getitem__(self, index):

         return torch.FloatTensor([self.data[index][0], self.data[index][1]]), self.data[index][2]

     def __len__(self):

         return len(self.data)

def train():

     os.makedirs('./output', exist_ok=True)

     batchsize = 10

     train_data = SexDataset(txt='sex_train.txt')

     val_data = SexDataset(txt='sex_val.txt')

     train_loader = DataLoader(dataset=train_data, batch_size=batchsize, shuffle=True)

     val_loader = DataLoader(dataset=val_data, batch_size=batchsize)

     model = sexnet()

     optimizer = torch.optim.Adam(model.parameters(), lr=0.01, weight_decay=1e-3)

     scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, [10, 20], 0.1)

     loss_func = nn.CrossEntropyLoss()

     epochs = 100

     for epoch in range(epochs):

         # training-----------------------------------

         model.train()

         train_loss = 0

         train_acc = 0

         for batch, (batch_x, batch_y) in enumerate(train_loader):

             batch_x, batch_y = Variable(batch_x), Variable(batch_y)

             out = 外匯跟單gendan5.commodel(batch_x)

             loss = loss_func(out, batch_y)

             train_loss += loss.item()

             pred = torch.max(out, 1)[1]

             train_correct = (pred == batch_y).sum()

             train_acc += train_correct.item()

             print('epoch: %2d/%d batch %3d/%d  Train Loss: %.3f, Acc: %.3f'

                   % (epoch + 1, epochs, batch, math.ceil(len(train_data) / batchsize),

                      loss.item(), train_correct.item() / len(batch_x)))

             optimizer.zero_grad()

             loss.backward()

             optimizer.step()

         scheduler.step()  # 更新 learning rate

         print('Train Loss: %.6f, Acc: %.3f' % (train_loss / (math.ceil(len(train_data) / batchsize)),

                                                train_acc / (len(train_data))))

         # evaluation--------------------------------

         model.eval()

         eval_loss = 0

         eval_acc = 0

         for batch_x, batch_y in val_loader:

             batch_x, batch_y = Variable(batch_x), Variable(batch_y)

             out = model(batch_x)

             loss = loss_func(out, batch_y)

             eval_loss += loss.item()

             pred = torch.max(out, 1)[1]

             num_correct = (pred == batch_y).sum()

             eval_acc += num_correct.item()

         print('Val Loss: %.6f, Acc: %.3f' % (eval_loss / (math.ceil(len(val_data) / batchsize)),

                                              eval_acc / (len(val_data))))

         # save model --------------------------------

         if (epoch + 1) % 1 == 0:

             torch.save(model.state_dict(), 'output/params_' + str(epoch + 1) + '.pth')

if __name__ == '__main__':

     train()

     print('finished')