计算机视觉CV-Pytorch 天池计算机视觉入门赛:SVHN数据集实战
这里我们以datawhale和天池合作的天池计算机视觉入门赛为例,通过案例实战来进一步巩固本章所介绍的图像分类知识。
比赛简介与赛题分析
该比赛以SVHN街道字符为赛题数据,数据集报名后可见并可下载,该数据来自收集的SVHN街道字符,并进行了匿名采样处理,详细的介绍见赛事官网。
注:以下代码均默认已将比赛数据的根文件夹命名为tianchi_SVHN
并放置于Dive-into-CV-PyTorch/dataset/tianchi_SVHN
下
我们要做的就是识别图片中的数字串,赛题给定的数据图片中不同图片中包含的字符数量不等,如下图所示。
看起来好像有点棘手,和本章介绍的图像分类还并不一样。这里我们利用一个巧妙的思路,将赛题转化为一个分类问题来解。
赛题数据集大部分图像中字符个数为2-4个,最多的字符个数为6个。因此可以对于所有的图像都抽象为6个字
符的定长字符识别问题,少于6位的部分填充为X。
例如字符23填充为23XXXX,字符231填充为231XXX。
于是相当于将赛题转化为了分别对6个数字进行的分类问题,每个数字预测0-9/X。
注:这种思路显然不是本次比赛以及SVHN数据集的最佳解法,但却十分巧妙,这样设计的实战项目可以巩固对于本章图像分类知识的学习,考察灵活掌握的程度。
环境安装
本节所介绍的实战代码对环境没有特殊依赖,理论上Python2/3,Pytorch1.x版本均可以跑通。此外,由于数据集较小,有无GPU都是可以的,动手开始实战吧~
下面给出 python3.7 + torch1.3.1-gpu版本的环境安装示例
注:假设你已经安装了Anaconda及CUDA10.0,如果你对环境安装不太了解,请阅读第一章的环境安装部分教程。
首先在Anaconda中创建一个专门用于本次天池联系赛的虚拟环境。
$ conda create -n py37_torch131 python=3.7
然后激活环境,并安装gpu版本Pytorch
$ source activate py37_torch131
$ conda install pytorch=1.3.1 torchvision cudatoolkit=10.0
最后通过下面的命令一键完成其他依赖库的安装。
$ pip install tqdm pandas matplotlib opencv-python jupyter
首先导入必要的库
# -*- coding: utf-8 -*-
import os, sys, glob, shutil, json
os.environ["CUDA_VISIBLE_DEVICES"] = '0'
import cv2
import numpy as np
from PIL import Image
from tqdm import tqdm, tqdm_notebook
%pylab inline
import torch
torch.manual_seed(0)
torch.backends.cudnn.deterministic = False
torch.backends.cudnn.benchmark = True
import torchvision.models as models
import torchvision.transforms as transforms
import torchvision.datasets as datasets
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.autograd import Variable
from torch.utils.data.dataset import Dataset
定义读取数据集
class SVHNDataset(Dataset):
def __init__(self, img_path, img_label, transform=None):
self.img_path = img_path
self.img_label = img_label
if transform is not None:
self.transform = transform
else:
self.transform = None
def __getitem__(self, index):
img = Image.open(self.img_path[index]).convert('RGB')
if self.transform is not None:
img = self.transform(img)
lbl = np.array(self.img_label[index], dtype=np.int)
lbl = list(lbl) + (5 - len(lbl)) * [10]
return img, torch.from_numpy(np.array(lbl[:5]))
def __len__(self):
return len(self.img_path)
定义读取数据dataloader
train_path = glob.glob('../../../dataset/tianchi_SVHN/train/*.png')
train_path.sort()
train_json = json.load(open('../../../dataset/tianchi_SVHN/train.json'))
train_label = [train_json[x]['label'] for x in train_json]
print(len(train_path), len(train_label))
train_loader = torch.utils.data.DataLoader(
SVHNDataset(train_path, train_label,
transforms.Compose([
transforms.Resize((64, 128)),
transforms.RandomCrop((60, 120)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])),
batch_size=40,
shuffle=True,
num_workers=10,
)
val_path = glob.glob('../../../dataset/tianchi_SVHN/val/*.png')
val_path.sort()
val_json = json.load(open('../../../dataset/tianchi_SVHN/val.json'))
val_label = [val_json[x]['label'] for x in val_json]
print(len(val_path), len(val_label))
val_loader = torch.utils.data.DataLoader(
SVHNDataset(val_path, val_label,
transforms.Compose([
transforms.Resize((60, 120)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])),
batch_size=40,
shuffle=False,
num_workers=10,
)
30000 30000
10000 10000
通过上述代码, 定义了赛题图像数据和对应标签的读取器dataloader。后面实际进行训练时,dataloader会根据我们代码中的定义,进行在线的数据増广,数据扩增的效果如下所示:
注:这里仅为一个示例图,上面的代码并没有使用旋转和颜色变换的数据增强
定义分类模型
这里使用ResNet18模型进行特征提取
class SVHN_Model1(nn.Module):
def __init__(self):
super(SVHN_Model1, self).__init__()
model_conv = models.resnet18(pretrained=True)
model_conv.avgpool = nn.AdaptiveAvgPool2d(1)
model_conv = nn.Sequential(*list(model_conv.children())[:-1])
self.cnn = model_conv
self.fc1 = nn.Linear(512, 11)
self.fc2 = nn.Linear(512, 11)
self.fc3 = nn.Linear(512, 11)
self.fc4 = nn.Linear(512, 11)
self.fc5 = nn.Linear(512, 11)
def forward(self, img):
feat = self.cnn(img)
# print(feat.shape)
feat = feat.view(feat.shape[0], -1)
c1 = self.fc1(feat)
c2 = self.fc2(feat)
c3 = self.fc3(feat)
c4 = self.fc4(feat)
c5 = self.fc5(feat)
return c1, c2, c3, c4, c5
def train(train_loader, model, criterion, optimizer):
# 切换模型为训练模式
model.train()
train_loss = []
for i, (input, target) in enumerate(train_loader):
if use_cuda:
input = input.cuda()
target = target.cuda()
c0, c1, c2, c3, c4 = model(input)
loss = criterion(c0, target[:, 0]) + \
criterion(c1, target[:, 1]) + \
criterion(c2, target[:, 2]) + \
criterion(c3, target[:, 3]) + \
criterion(c4, target[:, 4])
optimizer.zero_grad()
loss.backward()
optimizer.step()
train_loss.append(loss.item())
return np.mean(train_loss)
def validate(val_loader, model, criterion):
# 切换模型为预测模型
model.eval()
val_loss = []
# 不记录模型梯度信息
with torch.no_grad():
for i, (input, target) in enumerate(val_loader):
if use_cuda:
input = input.cuda()
target = target.cuda()
c0, c1, c2, c3, c4 = model(input)
loss = criterion(c0, target[:, 0]) + \
criterion(c1, target[:, 1]) + \
criterion(c2, target[:, 2]) + \
criterion(c3, target[:, 3]) + \
criterion(c4, target[:, 4])
val_loss.append(loss.item())
return np.mean(val_loss)
def predict(test_loader, model, tta=10):
model.eval()
test_pred_tta = None
# TTA 次数
for _ in range(tta):
test_pred = []
with torch.no_grad():
for i, (input, target) in enumerate(test_loader):
if use_cuda:
input = input.cuda()
c0, c1, c2, c3, c4 = model(input)
if use_cuda:
output = np.concatenate([
c0.data.cpu().numpy(),
c1.data.cpu().numpy(),
c2.data.cpu().numpy(),
c3.data.cpu().numpy(),
c4.data.cpu().numpy()], axis=1)
else:
output = np.concatenate([
c0.data.numpy(),
c1.data.numpy(),
c2.data.numpy(),
c3.data.numpy(),
c4.data.numpy()], axis=1)
test_pred.append(output)
test_pred = np.vstack(test_pred)
if test_pred_tta is None:
test_pred_tta = test_pred
else:
test_pred_tta += test_pred
return test_pred_tta
训练与验证
model = SVHN_Model1()
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), 0.001)
use_cuda = True
if use_cuda:
model = model.cuda()
best_loss = 1000.0
for epoch in range(3):
train_loss = train(train_loader, model, criterion, optimizer)
val_loss = validate(val_loader, model, criterion)
val_label = [''.join(map(str, x)) for x in val_loader.dataset.img_label]
val_predict_label = predict(val_loader, model, 1)
val_predict_label = np.vstack([
val_predict_label[:, :11].argmax(1),
val_predict_label[:, 11:22].argmax(1),
val_predict_label[:, 22:33].argmax(1),
val_predict_label[:, 33:44].argmax(1),
val_predict_label[:, 44:55].argmax(1),
]).T
val_label_pred = []
for x in val_predict_label:
val_label_pred.append(''.join(map(str, x[x!=10])))
val_char_acc = np.mean(np.array(val_label_pred) == np.array(val_label))
print('Epoch: {0}, Train loss: {1} \t Val loss: {2} \t Val Acc: {3}'.format(epoch, train_loss, val_loss, val_char_acc))
# 记录下验证集精度
if val_loss < best_loss:
best_loss = val_loss
# print('Find better model in Epoch {0}, saving model.'.format(epoch))
torch.save(model.state_dict(), './model.pt')
Epoch: 0, Train loss: 3.2353286933898926 Val loss: 3.518701043128967 Val Acc: 0.3391
Epoch: 1, Train loss: 1.99197505068779 Val loss: 2.9245959606170655 Val Acc: 0.4436
...
...
预测并生成提交文件
test_path = glob.glob('../../../dataset/tianchi_SVHN/test_a/*.png')
test_path.sort()
test_label = [[1]] * len(test_path)
print(len(test_path), len(test_label))
test_loader = torch.utils.data.DataLoader(
SVHNDataset(test_path, test_label,
transforms.Compose([
transforms.Resize((70, 140)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])),
batch_size=40,
shuffle=False,
num_workers=10,
)
40000 40000
# 加载保存的最优模型生成提交文件
model.load_state_dict(torch.load('model.pt'))
test_predict_label = predict(test_loader, model, 1)
print(test_predict_label.shape)
test_label = [''.join(map(str, x)) for x in test_loader.dataset.img_label]
test_predict_label = np.vstack([
test_predict_label[:, :11].argmax(1),
test_predict_label[:, 11:22].argmax(1),
test_predict_label[:, 22:33].argmax(1),
test_predict_label[:, 33:44].argmax(1),
test_predict_label[:, 44:55].argmax(1),
]).T
test_label_pred = []
for x in test_predict_label:
test_label_pred.append(''.join(map(str, x[x!=10])))
import pandas as pd
df_submit = pd.read_csv('../../../dataset/tianchi_SVHN/test_A_sample_submit.csv')
df_submit['file_code'] = test_label_pred
df_submit.to_csv('submit.csv', index=None)
(40000, 55)
调参实战
首先让我们快速确定baseline的几个基础参数
学习率调整
首先快速确定初始学习率,以及学习率调整策略
初始学习率判断标准:训练初期loss快速下降
我们先以0.01作为初始学习率,训练几个epoch
Epoch: 0, Train loss: 3.201288181463877 Val loss: 3.6346988077163696 Val Acc: 0.3234
Epoch: 1, Train loss: 1.9820853635470073 Val loss: 3.010792998790741 Val Acc: 0.44
Epoch: 2, Train loss: 1.6244886504809062 Val loss: 2.765939730644226 Val Acc: 0.4758
Epoch: 3, Train loss: 1.4178793375492096 Val loss: 2.6497371077537535 Val Acc: 0.5069
Epoch: 4, Train loss: 1.2603404329617818 Val loss: 2.716396602153778 Val Acc: 0.4849
Epoch: 5, Train loss: 1.1504622252782186 Val loss: 2.5561904258728028 Val Acc: 0.5357
Epoch: 6, Train loss: 1.04843408370018 Val loss: 2.5777493786811827 Val Acc: 0.5415
Epoch: 7, Train loss: 0.9496217265923818 Val loss: 2.524811834573746 Val Acc: 0.5326
Epoch: 8, Train loss: 0.8733580025633176 Val loss: 2.6108505029678346 Val Acc: 0.5505
Epoch: 9, Train loss: 0.7865098938941956 Val loss: 2.607123359680176 Val Acc: 0.5523
best val acc: 0.5523
可以看到,训练初期loss快速下降,所以0.01作为初始学习率是合适的。
训练中,学习率是可以进行调整的,一种常用的方法是阶段性学习率衰减策略。
在上面尝试性的训练过程中,可以看到,验证集loss在10个epoch左右下降趋势已不明显,因此可以尝试在第10个epoch,将学习率衰减为原来的10%。按照这样的学习率调整策略,我们训练20个epoch看下:
Epoch: 0, Train loss: 3.217955897013346 Val loss: 3.4973887462615965 Val Acc: 0.3428
Epoch: 1, Train loss: 1.9865643223921459 Val loss: 3.041480797290802 Val Acc: 0.4252
Epoch: 2, Train loss: 1.6422564171155294 Val loss: 2.734321162700653 Val Acc: 0.4845
Epoch: 3, Train loss: 1.4109662581682205 Val loss: 2.6289015769958497 Val Acc: 0.5124
Epoch: 4, Train loss: 1.268737798611323 Val loss: 2.6607061581611635 Val Acc: 0.5025
Epoch: 5, Train loss: 1.149802592118581 Val loss: 2.693304219722748 Val Acc: 0.5122
Epoch: 6, Train loss: 1.0440267440478006 Val loss: 2.5291405525207518 Val Acc: 0.5363
Epoch: 7, Train loss: 0.9582021673123042 Val loss: 2.64508363032341 Val Acc: 0.5167
Epoch: 8, Train loss: 0.8617052629590034 Val loss: 2.5092941007614136 Val Acc: 0.5473
Epoch: 9, Train loss: 0.7965546736717224 Val loss: 2.4935685346126557 Val Acc: 0.5429
Epoch: 10, Train loss: 0.47227135149141153 Val loss: 2.3851090211868287 Val Acc: 0.5987
Epoch: 11, Train loss: 0.356143202851216 Val loss: 2.458720991849899 Val Acc: 0.603
Epoch: 12, Train loss: 0.3016581511025627 Val loss: 2.5575384349822996 Val Acc: 0.6012
Epoch: 13, Train loss: 0.26212659483402967 Val loss: 2.7137050013542177 Val Acc: 0.5975
Epoch: 14, Train loss: 0.22160522095113994 Val loss: 2.834437639474869 Val Acc: 0.6031
Epoch: 15, Train loss: 0.1934196574985981 Val loss: 3.015086229324341 Val Acc: 0.5971
Epoch: 16, Train loss: 0.16129128922770422 Val loss: 3.131588038921356 Val Acc: 0.601
Epoch: 17, Train loss: 0.14016953054318826 Val loss: 3.417837708234787 Val Acc: 0.5923
Epoch: 18, Train loss: 0.12422899308552345 Val loss: 3.5185752115249636 Val Acc: 0.5991
Epoch: 19, Train loss: 0.09977891861026486 Val loss: 3.5448452734947207 Val Acc: 0.6027
best val acc: 0.6031
可以看到,这样的策略将验证集准确率大幅提高到了0.6031 🚀
核心是要抓主要矛盾。
目前的问题是验证集和测试集的准确率存在很大的误差,通常这说明验证集和测试集存在一定的差异。
由于这是个比赛,数据集都是官方提供的,那这种差异是不是我们自己代码的隐性bug带来的呢?
我们观察代码,会发现,训练和测试的dataloader中resize部分是存在不一致的
train_loader = torch.utils.data.DataLoader(
SVHNDataset(train_path, train_label,
transforms.Compose([
transforms.Resize((64, 128)),
transforms.RandomCrop((60, 120)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])),
batch_size=40,
shuffle=True,
num_workers=10,
)
...
test_loader = torch.utils.data.DataLoader(
SVHNDataset(test_path, test_label,
transforms.Compose([
transforms.Resize((70, 140)),
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])),
batch_size=40,
shuffle=False,
num_workers=10,
)
图像的输入尺寸是不一致的,而且差异不小,相当于人为引入了验证集和测试集之间的差异。
因此我们将test_loader
调整为
test_loader = torch.utils.data.DataLoader(
SVHNDataset(test_path, test_label,
transforms.Compose([
transforms.Resize((60, 120)), # TODO: modify here
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])),
batch_size=40,
shuffle=False,
num_workers=10,
)
这个bug修复后,成绩一下就能达到0.7左右的水平,我提交的分数是0.7319 🚀
数据增强策略
我们观察上面的训练日志可以发现,训练集loss可以达到非常低的水平,但是验证集做不到,因此目前的主要矛盾变成了过拟合。
...
Epoch: 17, Train loss: 0.14016953054318826 Val loss: 3.417837708234787 Val Acc: 0.5923
Epoch: 18, Train loss: 0.12422899308552345 Val loss: 3.5185752115249636 Val Acc: 0.5991
Epoch: 19, Train loss: 0.09977891861026486 Val loss: 3.5448452734947207 Val Acc: 0.6027
best val acc: 0.6031
于是数据增强就成了目前最有可能提分的武器之一。
上面的baseline中我们在训练过程中仅仅使用了RandomCrop作为数据增强,下面我们尝试在训练集中使用更多的数据增强方法,并验证效果。
我们尝试加上一些颜色空间的变换
train_loader = torch.utils.data.DataLoader(
SVHNDataset(train_path, train_label,
transforms.Compose([
transforms.Resize((64, 128)),
transforms.RandomCrop((60, 120)),
transforms.ColorJitter(0.3, 0.3, 0.2), # TODO: new add
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])),
batch_size=40,
shuffle=True,
num_workers=10,
)
重新训练并提交后,得到了0.7453的分数,又上涨了一些。🚀
你还可以继续尝试其他的数据增强方法,以及调整其他的参数,这里就不介绍了。
核心就是注意把握两点:
-
抓主要矛盾,确保你现在的实验是在解决目前面临的主要问题,而不是一些无关痛痒的尝试。
-
遵循单一变量原则进行实验。
相信通过本文的介绍,你已经了解了调参的基本方法和思路。
开始自己动手实验吧,Good Luck~