百度飞浆ResNet50大模型微调实现十二种猫图像分类

要求很简单，给train和test集，训练模型实现图像分类。

这里使用的是残差连接模型，这个平台有预训练好的模型，可以直接拿来主义。

训练十几个迭代，每个批次60左右，准确率达到90%以上

一、导入库，解压文件

python 复制代码

import os
import zipfile
import random
import json
import cv2
import numpy as np
from PIL import Image

import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
import paddle
import paddle.nn as nn
from paddle.io import Dataset,DataLoader
from paddle.nn import \
                    Layer, \
                    Conv2D, Linear, \
                    Embedding, MaxPool2D, \
                    BatchNorm2D, ReLU
                    
import paddle.vision.transforms as transforms
from paddle.vision.models import resnet50
from paddle.metric import Accuracy

train_parameters = {
    "input_size": [3, 224, 224],                     # 输入图片的shape
    "class_dim": 12,                                 # 分类数
    "src_path":"data/data10954/cat_12_train.zip",   # 原始数据集路径
    "src_test_path":"data/data10954/cat_12_test.zip",   # 原始数据集路径
    "target_path":"/home/aistudio/data/dataset",     # 要解压的路径 
    "train_list_path": "./train.txt",                # train_data.txt路径
    "eval_list_path": "./eval.txt",                  # eval_data.txt路径
    "label_dict":{},                                 # 标签字典
    "readme_path": "/home/aistudio/data/readme.json",# readme.json路径
    "num_epochs":6,                                 # 训练轮数
    "train_batch_size": 16,                          # 批次的大小
    "learning_strategy": {                           # 优化函数相关的配置
        "lr": 0.0005                                  # 超参数学习率
    } 
}


scr_path=train_parameters['src_path']
target_path=train_parameters['target_path']
src_test_path=train_parameters["src_test_path"]
z = zipfile.ZipFile(scr_path, 'r')
z.extractall(path=target_path)
z = zipfile.ZipFile(src_test_path, 'r')
z.extractall(path=target_path)
z.close()
for imgpath in os.listdir(target_path + '/cat_12_train'):
    src = os.path.join(target_path + '/cat_12_train/', imgpath)
    img = Image.open(src)
    if img.mode != 'RGB':
        img = img.convert('RGB')
        img.save(src)

for imgpath in os.listdir(target_path + '/cat_12_test'):
    src = os.path.join(target_path + '/cat_12_test/', imgpath)
    img = Image.open(src)
    if img.mode != 'RGB':
        img = img.convert('RGB')
        img.save(src)

解压后将所有图像变为RGB图像

二、加载训练集，进行预处理、数据增强、格式变换

python 复制代码

transform = transforms.Compose([
    transforms.Resize(size=224),
    transforms.ColorJitter(0.2, 0.2, 0.2, 0.2),
    transforms.RandomHorizontalFlip(),
    transforms.RandomRotation(15),
    transforms.RandomResizedCrop(size=224, scale=(0.8, 1.0)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

x_train,x_eval,y_train=[],[],[]#获取训练图像和标签、测试图像和标签
contents=[]
with open('data/data10954/train_list.txt')as f:
    contents=f.read().split('\n')

for item in contents:
    if item=='':
        continue
    path='data/dataset/'+item.split('\t')[0]
    data=np.array(Image.open(path).convert('RGB'))


    data=np.array(transform(data))
    x_train.append(data)
    y_train.append(int(item.split('\t')[-1]))

contetns=os.listdir('data/dataset/cat_12_test')
for item in contetns:
    path='data/dataset/cat_12_test/'+item
    data=np.array(Image.open(path).convert('RGB'))
    data=np.array(transform(data))

    x_eval.append(data)

重点是transforms变换的预处理

三、划分训练集和测试集

python 复制代码

x_train=np.array(x_train)

y_train=np.array(y_train)

x_eval=np.array(x_eval)



x_train,x_test,y_train,y_test=train_test_split(x_train,y_train,test_size=0.2,random_state=42,stratify=y_train)

x_train=paddle.to_tensor(x_train,dtype='float32')
y_train=paddle.to_tensor(y_train,dtype='int64')
x_test=paddle.to_tensor(x_test,dtype='float32')
y_test=paddle.to_tensor(y_test,dtype='int64')
x_eval=paddle.to_tensor(x_eval,dtype='float32')

这是必要的，可以随时利用测试集查看准确率

四、加载预训练模型，选择损失函数和优化器

python 复制代码

learning_rate=0.001
epochs =5  # 迭代轮数
batch_size = 50  # 批次大小
weight_decay=1e-5
num_class=12

cnn=resnet50(pretrained=True)
checkpoint=paddle.load('checkpoint.pdparams')

for param in cnn.parameters():
    param.requires_grad=False
cnn.fc = nn.Linear(2048, num_class)
cnn.set_dict(checkpoint['cnn_state_dict'])
criterion=nn.CrossEntropyLoss()
optimizer = paddle.optimizer.Adam(learning_rate=learning_rate, parameters=cnn.fc.parameters(),weight_decay=weight_decay)

第一次训练把加载模型注释掉即可，优化器包含最后一层全连接的参数

五、模型训练

python 复制代码

if x_train.shape[3]==3:
    x_train=paddle.transpose(x_train,perm=(0,3,1,2))

dataset = paddle.io.TensorDataset([x_train, y_train])
data_loader = DataLoader(dataset, batch_size=batch_size, shuffle=True)
for epoch in range(epochs):

    for batch_data, batch_labels in data_loader:
        outputs = cnn(batch_data)
        loss = criterion(outputs, batch_labels)
        print(epoch)
        loss.backward()
        optimizer.step()
        optimizer.clear_grad()

    print(f"Epoch [{epoch+1}/{epochs}], Loss: {loss.numpy()[0]}")#保存参数
paddle.save({
    'cnn_state_dict': cnn.state_dict(),

}, 'checkpoint.pdparams')

使用批处理，这个很重要，不然平台分分钟炸了

六、测试集准确率

python 复制代码

num_class=12
batch_size=64
cnn=resnet50(pretrained=True)
checkpoint=paddle.load('checkpoint.pdparams')

for param in cnn.parameters():
    param.requires_grad=False
cnn.fc = nn.Linear(2048, num_class)
cnn.set_dict(checkpoint['cnn_state_dict'])

cnn.eval()

if x_test.shape[3]==3:
        x_test=paddle.transpose(x_test,perm=(0,3,1,2))
dataset = paddle.io.TensorDataset([x_test, y_test])
data_loader = DataLoader(dataset, batch_size=batch_size, shuffle=True)

with paddle.no_grad():
    score=0
    for batch_data, batch_labels in data_loader:
    
        predictions = cnn(batch_data)
        predicted_probabilities = paddle.nn.functional.softmax(predictions, axis=1)
        predicted_labels = paddle.argmax(predicted_probabilities, axis=1) 
        print(predicted_labels)
        
        for i in range(len(predicted_labels)):
            if predicted_labels[i].numpy()==batch_labels[i]:
                score+=1
    print(score/len(y_test))

设置eval模式，使用批处理测试准确率