D2L-Kaggle-树叶分类

实战Kaggle比赛：树叶分类

竞赛总结

• 模型为多种ResNet变种

  • DenseNet, ResNeXt, ResNeSt, …
  • EfficientNet

• 优化算法多为Adam和其变种

• 学习率一般为Cosine，或者训练不动的时候往下调整（逐步变小）

• AutoGluon

  • 15行代码，安装+训练：https://www.kaggle.com/code/zhreshold/autogluon-vision-0-96-with-15-lines/notebook
  • 精度96%：
    • 通过定制化可以提升精度
    • 下一个模型将搜索更多的模型超参数
    • AG目前仍然关注工业界应用，而非比赛

代码实现

代码来自simple resnet baseline的解法，菜菜不会写，这里只是阅读 + 笔记，大自然的搬运工

• 引入依赖：

# 首先导入包
import torch
import torch.nn as nn
import pandas as pd
import numpy as np
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
from PIL import Image
import os
import matplotlib.pyplot as plt
import torchvision.models as models
# This is for the progress bar.
from tqdm import tqdm
import seaborn as sns

• 看下数据：

# 看看label文件长啥样
labels_dataframe = pd.read_csv('../input/classify-leaves/train.csv')
labels_dataframe.head(5)

labels_dataframe.describe()

#function to show bar length

def barw(ax): 
    
    for p in ax.patches:
        val = p.get_width() #height of the bar
        x = p.get_x()+ p.get_width() # x- position 
        y = p.get_y() + p.get_height()/2 #y-position
        ax.annotate(round(val,2),(x,y))
        
#finding top leaves

plt.figure(figsize = (15,30))
ax0 =sns.countplot(y=labels_dataframe['label'],order=labels_dataframe['label'].value_counts().index)
barw(ax0)
plt.show()

• label文件排序：

# 把label文件排个序
leaves_labels = sorted(list(set(labels_dataframe['label'])))
n_classes = len(leaves_labels)
print(n_classes)
leaves_labels[:10]

• label转换为对应的数字：

# 把label转成对应的数字
class_to_num = dict(zip(leaves_labels, range(n_classes)))
class_to_num

• 建议映射关系，方便最后预测的时候使用

# 再转换回来，方便最后预测的时候使用
num_to_class = {v : k for k, v in class_to_num.items()}

• 创建自己的Dataset

# 继承pytorch的dataset，创建自己的
class LeavesData(Dataset):
    def __init__(self, csv_path, file_path, mode='train', valid_ratio=0.2, resize_height=256, resize_width=256):
        """
        Args:
            csv_path (string): csv 文件路径
            img_path (string): 图像文件所在路径
            mode (string): 训练模式还是测试模式
            valid_ratio (float): 验证集比例
        """
        
        # 需要调整后的照片尺寸，我这里每张图片的大小尺寸不一致#
        self.resize_height = resize_height
        self.resize_width = resize_width

        self.file_path = file_path
        self.mode = mode

        # 读取 csv 文件
        # 利用pandas读取csv文件
        self.data_info = pd.read_csv(csv_path, header=None)  #header=None是去掉表头部分
        # 计算 length
        self.data_len = len(self.data_info.index) - 1
        self.train_len = int(self.data_len * (1 - valid_ratio))
        
        if mode == 'train':
            # 第一列包含图像文件的名称
            self.train_image = np.asarray(self.data_info.iloc[1:self.train_len, 0])  #self.data_info.iloc[1:,0]表示读取第一列，从第二行开始到train_len
            # 第二列是图像的 label
            self.train_label = np.asarray(self.data_info.iloc[1:self.train_len, 1])
            self.image_arr = self.train_image 
            self.label_arr = self.train_label
        elif mode == 'valid':
            self.valid_image = np.asarray(self.data_info.iloc[self.train_len:, 0])  
            self.valid_label = np.asarray(self.data_info.iloc[self.train_len:, 1])
            self.image_arr = self.valid_image
            self.label_arr = self.valid_label
        elif mode == 'test':
            self.test_image = np.asarray(self.data_info.iloc[1:, 0])
            self.image_arr = self.test_image
            
        self.real_len = len(self.image_arr)

        print('Finished reading the {} set of Leaves Dataset ({} samples found)'
              .format(mode, self.real_len))

    def __getitem__(self, index):
        # 从 image_arr中得到索引对应的文件名
        single_image_name = self.image_arr[index]

        # 读取图像文件
        img_as_img = Image.open(self.file_path + single_image_name)

        #如果需要将RGB三通道的图片转换成灰度图片可参考下面两行
#         if img_as_img.mode != 'L':
#             img_as_img = img_as_img.convert('L')
		
  			#设置好需要转换的变量，还可以包括一系列的nomarlize等等操作
        if self.mode == 'train':
            transform = transforms.Compose([
                transforms.Resize((224, 224)),
                transforms.RandomHorizontalFlip(p=0.5),   #随机水平翻转 选择一个概率
                transforms.ToTensor()
            ])
        else:
            # valid和test不做数据增强
            transform = transforms.Compose([
                transforms.Resize((224, 224)),
                transforms.ToTensor()
            ])
        
        img_as_img = transform(img_as_img)
        
        if self.mode == 'test':
            return img_as_img
        else:
            # 得到图像的 string label
            label = self.label_arr[index]
            # number label
            number_label = class_to_num[label]

            return img_as_img, number_label  #返回每一个index对应的图片数据和对应的label

    def __len__(self):
        return self.real_len

• 将数据变成自定义的数据集：

train_path = '../input/classify-leaves/train.csv'
test_path = '../input/classify-leaves/test.csv'
# csv文件中已经images的路径了，因此这里只到上一级目录
img_path = '../input/classify-leaves/'

train_dataset = LeavesData(train_path, img_path, mode='train')
val_dataset = LeavesData(train_path, img_path, mode='valid')
test_dataset = LeavesData(test_path, img_path, mode='test')
print(train_dataset)
print(val_dataset)
print(test_dataset)

• Data Loader定义

# 定义data loader
train_loader = torch.utils.data.DataLoader(
        dataset=train_dataset,
        batch_size=8, 
        shuffle=False,
        num_workers=5
    )

val_loader = torch.utils.data.DataLoader(
        dataset=val_dataset,
        batch_size=8, 
        shuffle=False,
        num_workers=5
    )
test_loader = torch.utils.data.DataLoader(
        dataset=test_dataset,
        batch_size=8, 
        shuffle=False,
        num_workers=5
    )

• 展示数据

# 给大家展示一下数据长啥样
def im_convert(tensor):
    """ 展示数据"""
    
    image = tensor.to("cpu").clone().detach()
    image = image.numpy().squeeze()
    image = image.transpose(1,2,0)
    image = image.clip(0, 1)

    return image

fig=plt.figure(figsize=(20, 12))
columns = 4
rows = 2

dataiter = iter(val_loader)
inputs, classes = dataiter.next()

for idx in range (columns*rows):
    ax = fig.add_subplot(rows, columns, idx+1, xticks=[], yticks=[])
    ax.set_title(num_to_class[int(classes[idx])])
    plt.imshow(im_convert(inputs[idx]))
plt.show()

• 采用训练的是CPU还是GPU

# 看一下是在cpu还是GPU上
def get_device():
    return 'cuda' if torch.cuda.is_available() else 'cpu'

device = get_device()
print(device)

• resnet模型使用：

# 是否要冻住模型的前面一些层
def set_parameter_requires_grad(model, feature_extracting):
    if feature_extracting:
        model = model
        for param in model.parameters():
            param.requires_grad = False
# resnet34模型
def res_model(num_classes, feature_extract = False, use_pretrained=True):

    model_ft = models.resnet34(pretrained=use_pretrained)
    set_parameter_requires_grad(model_ft, feature_extract)
    num_ftrs = model_ft.fc.in_features
    model_ft.fc = nn.Sequential(nn.Linear(num_ftrs, num_classes))

    return model_ft

• 超参数设置：

# 超参数
learning_rate = 3e-4
weight_decay = 1e-3
num_epoch = 50
model_path = './pre_res_model.ckpt'

• 初始化模型，放到GPU/CPU上

# Initialize a model, and put it on the device specified.
model = res_model(176)
model = model.to(device)
model.device = device
# For the classification task, we use cross-entropy as the measurement of performance.
criterion = nn.CrossEntropyLoss()

# Initialize optimizer, you may fine-tune some hyperparameters such as learning rate on your own.
optimizer = torch.optim.Adam(model.parameters(), lr = learning_rate, weight_decay=weight_decay)

# The number of training epochs.
n_epochs = num_epoch

best_acc = 0.0
for epoch in range(n_epochs):
    # ---------- Training ----------
    # Make sure the model is in train mode before training.
    model.train() 
    # These are used to record information in training.
    train_loss = []
    train_accs = []
    # Iterate the training set by batches.
    for batch in tqdm(train_loader):
        # A batch consists of image data and corresponding labels.
        imgs, labels = batch
        imgs = imgs.to(device)
        labels = labels.to(device)
        # Forward the data. (Make sure data and model are on the same device.)
        logits = model(imgs)
        # Calculate the cross-entropy loss.

        # We don't need to apply softmax before computing cross-entropy as it is done automatically.
        loss = criterion(logits, labels)
        
        # Gradients stored in the parameters in the previous step should be cleared out first.
        optimizer.zero_grad()
        # Compute the gradients for parameters.
        loss.backward()
        # Update the parameters with computed gradients.
        optimizer.step()
        
        # Compute the accuracy for current batch.
        acc = (logits.argmax(dim=-1) == labels).float().mean()

        # Record the loss and accuracy.
        train_loss.append(loss.item())
        train_accs.append(acc)
        
    # The average loss and accuracy of the training set is the average of the recorded values.
    train_loss = sum(train_loss) / len(train_loss)
    train_acc = sum(train_accs) / len(train_accs)

    # Print the information.
    print(f"[ Train | {epoch + 1:03d}/{n_epochs:03d} ] loss = {train_loss:.5f}, acc = {train_acc:.5f}")
    
    # ---------- Validation ----------
    # Make sure the model is in eval mode so that some modules like dropout are disabled and work normally.
    model.eval()
    # These are used to record information in validation.
    valid_loss = []
    valid_accs = []
    
    # Iterate the validation set by batches.
    for batch in tqdm(val_loader):
        imgs, labels = batch
        # We don't need gradient in validation.
        # Using torch.no_grad() accelerates the forward process.
        with torch.no_grad():
            logits = model(imgs.to(device))
            
        # We can still compute the loss (but not the gradient).
        loss = criterion(logits, labels.to(device))

        # Compute the accuracy for current batch.
        acc = (logits.argmax(dim=-1) == labels.to(device)).float().mean()

        # Record the loss and accuracy.
        valid_loss.append(loss.item())
        valid_accs.append(acc)
        
    # The average loss and accuracy for entire validation set is the average of the recorded values.
    valid_loss = sum(valid_loss) / len(valid_loss)
    valid_acc = sum(valid_accs) / len(valid_accs)

    # Print the information.
    print(f"[ Valid | {epoch + 1:03d}/{n_epochs:03d} ] loss = {valid_loss:.5f}, acc = {valid_acc:.5f}")
    
    # if the model improves, save a checkpoint at this epoch
    if valid_acc > best_acc:
        best_acc = valid_acc
        torch.save(model.state_dict(), model_path)
        print('saving model with acc {:.3f}'.format(best_acc))

• 把结果存起来：

saveFileName = './submission.csv'

## predict
model = res_model(176)

# create model and load weights from checkpoint
model = model.to(device)
model.load_state_dict(torch.load(model_path))

# Make sure the model is in eval mode.
# Some modules like Dropout or BatchNorm affect if the model is in training mode.
model.eval()

# Initialize a list to store the predictions.
predictions = []
# Iterate the testing set by batches.
for batch in tqdm(test_loader):
    
    imgs = batch
    with torch.no_grad():
        logits = model(imgs.to(device))
    
    # Take the class with greatest logit as prediction and record it.
    predictions.extend(logits.argmax(dim=-1).cpu().numpy().tolist())

preds = []
for i in predictions:
    preds.append(num_to_class[i])

test_data = pd.read_csv(test_path)
test_data['label'] = pd.Series(preds)
submission = pd.concat([test_data['image'], test_data['label']], axis=1)
submission.to_csv(saveFileName, index=False)
print("Done!!!!!!!!!!!!!!!!!!!!!!!!!!!")

总结

• 提升精度思路：根据数据挑选增强，使用新模型、新优化算法，多个模型融合,测试时使用增强
• 数据相对简单，排名有相对随机性
• 在工业界应用中:
  • 少使用模型融合和测试时增强，计算代价过高
  • 通常固定模型超参数，而将精力主要花在提升数据质量

References

1. 树叶分类kaggle
2. Kaggle竞赛代码阅读
3. 沐神总结，里面有关于DenseNet，EfficientNet等相关的介绍，还有Adam等内容