Python45对青春有你2进行二分类和图像预测

一、青春有你2进行二分类

"""
数据准备在进行Finetune前&＃xff0c;我们可以设置一些运行时的配置&＃xff0c;例如如下代码中的配置&＃xff0c;表示&＃xff1a;use_cuda&＃xff1a;设置为False表示使用CPU进行训练。如果您本机支持GPU&＃xff0c;且安装的是GPU版本的PaddlePaddle&＃xff0c;我们建议您将这个选项设置为True&＃xff1b;epoch&＃xff1a;迭代轮数&＃xff1b;batch_size&＃xff1a;每次训练的时候&＃xff0c;给模型输入的每批数据大小为32&＃xff0c;模型训练时能够并行处理批数据&＃xff0c;因此batch_size越大&＃xff0c;训练的效率越高&＃xff0c;但是同时带来了内存的负荷&＃xff0c;过大的batch_size可能导致内存不足而无法训练&＃xff0c;因此选择一个合适的batch_size是很重要的一步&＃xff1b;log_interval&＃xff1a;每隔10 step打印一次训练日志&＃xff1b;eval_interval&＃xff1a;每隔50 step在验证集上进行一次性能评估&＃xff1b;checkpoint_dir&＃xff1a;将训练的参数和数据保存到cv_finetune_turtorial_demo目录中&＃xff1b;strategy&＃xff1a;使用DefaultFinetuneStrategy策略进行finetune
"""
import paddlehub as hub
import paddle
paddle.enable_static()from paddlehub.dataset.base_cv_dataset import BaseCVDataset
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
module&＃61;hub.Module(name&＃61;"resnet_v2_50_imagenet")
import matplotlib
matplotlib.use("TkAgg")class DemoDataSet(BaseCVDataset):def __init__(self):self.dataset_dir&＃61;"data"super(DemoDataSet,self).__init__(base_path&＃61;self.dataset_dir,train_list_file&＃61;"train_list.txt",validate_list_file&＃61;"validate_list.txt",test_list_file&＃61;"test_list.txt",label_list_file&＃61;"label_list.txt")
dataset &＃61; DemoDataSet()
img&＃61;mpimg.imread("data/xu/1.jpg")
img1&＃61;mpimg.imread("data/yu/1.jpg")
plt.figure(figsize&＃61;(10,10))
plt.subplot(1,2,1)
plt.imshow(img)
plt.axis("off")
plt.subplot(1,2,2)
plt.imshow(img1)
plt.axis("off")
plt.show()
"""
生成数据读取器
"""
data_reader&＃61;hub.reader.ImageClassificationReader(image_width&＃61;module.get_expected_image_width(),image_height&＃61;module.get_expected_image_height(),images_mean&＃61;module.get_pretrained_images_mean(),images_std&＃61;module.get_pretrained_images_std(),dataset&＃61;dataset
)config &＃61; hub.RunConfig(use_cuda&＃61;True, #是否使用GPU训练&＃xff0c;默认为False&＃xff1b;num_epoch&＃61;3, #Fine-tune的轮数&＃xff1b;checkpoint_dir&＃61;"cv_finetune_turtorial_demo",#模型checkpoint保存路径, 若用户没有指定&＃xff0c;程序会自动生成&＃xff1b;batch_size&＃61;3, #训练的批大小&＃xff0c;如果使用GPU&＃xff0c;请根据实际情况调整batch_size&＃xff1b;eval_interval&＃61;10, #模型评估的间隔&＃xff0c;默认每100个step评估一次验证集&＃xff1b;strategy&＃61;hub.finetune.strategy.DefaultFinetuneStrategy())input_dict, output_dict, program &＃61; module.context(trainable&＃61;True)
img &＃61; input_dict["image"]
feature_map &＃61; output_dict["feature_map"]
feed_list &＃61; [img.name]task &＃61; hub.ImageClassifierTask(data_reader&＃61;data_reader,feed_list&＃61;feed_list,feature&＃61;feature_map,num_classes&＃61;dataset.num_labels,config&＃61;config)import numpy as np
import matplotlib.pyplot as plt
import matplotlib.image as mpimgdata &＃61; ["data/xu/16.jpg","data/yu/16.jpg"]label_map &＃61; dataset.label_dict()
index &＃61; 0
run_states &＃61; task.predict(data&＃61;data)
results &＃61; [run_state.run_results for run_state in run_states]for batch_result in results:print(batch_result)batch_result &＃61; np.argmax(batch_result, axis&＃61;2)[0]print(batch_result)for result in batch_result:index &＃43;&＃61; 1result &＃61; label_map[result]print("input %i is %s, and the predict result is %s" %(index, data[index - 1], result))img &＃61; mpimg.imread(data[0])
img1 &＃61; mpimg.imread(data[1])
plt.figure(figsize&＃61;(10,10))
plt.subplot(1,2,1)
plt.imshow(img)
plt.axis(&＃39;off&＃39;)
plt.subplot(1,2,2)
plt.imshow(img1)
plt.axis(&＃39;off&＃39;)
plt.show()


二、对青春有你2进行图像预测

# 作者&＃xff1a;Kerwin Wan
# 开发时间&＃xff1a;2022/9/29 15:27"""
图像预测
"""
import paddlehub as hub
import paddle
paddle.enable_static()
# 加载预训练模型
module&＃61;hub.Module(name&＃61;"resnet_v2_50_imagenet")
from paddlehub.dataset.base_cv_dataset import BaseCVDataset
class DemoDataSet(BaseCVDataset):def __init__(self):self.dataset_dir&＃61;"data"super(DemoDataSet, self).__init__(base_path&＃61;self.dataset_dir,train_list_file&＃61;"train_list.txt",validate_list_file&＃61;"validate_list.txt",label_list_file&＃61;"label_list.txt")
dataset&＃61;DemoDataSet()import matplotlib.pyplot as plt
import matplotlib.image as mpimg
img&＃61;mpimg.imread("data/xu/1.jpg")
img1&＃61;mpimg.imread("data/yu/1.jpg")
plt.figure(figsize&＃61;(10,10))
plt.subplot(1,2,1)
plt.imshow(img)
plt.axis("off")
plt.subplot(1,2,2)
plt.imshow(img1)
plt.axis("off")
plt.show()# 生成数据读取器
data_reader&＃61;hub.reader.ImageClassificationReader(image_width&＃61;module.get_expected_image_width(),image_height&＃61;module.get_expected_image_height(),images_mean&＃61;module.get_pretrained_images_mean(),images_std&＃61;module.get_pretrained_images_std(),dataset&＃61;dataset
)
# 配置策略
config&＃61;hub.RunConfig(use_cuda&＃61;True,num_epoch&＃61;3,checkpoint_dir&＃61;"cv_finetune_turtorial_demo",batch_size&＃61;3,eval_interval&＃61;10,strategy&＃61;hub.finetune.strategy.DefaultFinetuneStrategy(),
)
# 组建Finetune Task
input_dict,output_dict,program&＃61;module.context(trainable&＃61;True)
img &＃61; input_dict["image"]
feature_map &＃61; output_dict["feature_map"]
feed_list &＃61; [img.name]task &＃61; hub.ImageClassifierTask(data_reader&＃61;data_reader,feed_list&＃61;feed_list,feature&＃61;feature_map,num_classes&＃61;dataset.num_labels,config&＃61;config)
run_states &＃61; task.finetune_and_eval()import numpy as np
import matplotlib.pyplot as plt
import matplotlib.image as mpimgdata &＃61; ["data/xu/16.jpg","data/yu/16.jpg"]label_map &＃61; dataset.label_dict()
index &＃61; 0
run_states &＃61; task.predict(data&＃61;data)
results &＃61; [run_state.run_results for run_state in run_states]for batch_result in results:print(batch_result)batch_result &＃61; np.argmax(batch_result, axis&＃61;2)[0]print(batch_result)for result in batch_result:index &＃43;&＃61; 1result &＃61; label_map[result]print("input %i is %s, and the predict result is %s" %(index, data[index - 1], result))img &＃61; mpimg.imread(data[0])
img1 &＃61; mpimg.imread(data[1])
plt.figure(figsize&＃61;(10,10))
plt.subplot(1,2,1)
plt.imshow(img)
plt.axis(&＃39;off&＃39;)
plt.subplot(1,2,2)
plt.imshow(img1)
plt.axis(&＃39;off&＃39;)
plt.show()
print(" input1 许佳琪 input2 虞书欣")

三、训练预测

# 作者&＃xff1a;Kerwin Wan
# 开发时间&＃xff1a;2022/9/29 16:43#导入需要的包
import numpy as np
import paddle as paddle
import paddle.fluid as fluid
from PIL import Image
import matplotlib.pyplot as plt
import os
import paddle
paddle.enable_static()train_reader &＃61; paddle.batch(paddle.reader.shuffle(paddle.dataset.mnist.train(),buf_size&＃61;512),batch_size&＃61;128)
test_reader &＃61; paddle.batch(paddle.dataset.mnist.test(),batch_size&＃61;128)temp_reader &＃61; paddle.batch(paddle.dataset.mnist.train(),batch_size&＃61;1)
temp_data&＃61;next(temp_reader())
print(temp_data)def multilayer_perceptron(input):# 第一个全连接层&＃xff0c;激活函数为ReLUhidden1 &＃61; fluid.layers.fc(input&＃61;input, size&＃61;100, act&＃61;&＃39;relu&＃39;)# 第二个全连接层&＃xff0c;激活函数为ReLUhidden2 &＃61; fluid.layers.fc(input&＃61;hidden1, size&＃61;100, act&＃61;&＃39;relu&＃39;)# 以softmax为激活函数的全连接输出层&＃xff0c;大小为10prediction &＃61; fluid.layers.fc(input&＃61;hidden2, size&＃61;10, act&＃61;&＃39;softmax&＃39;)return prediction
image &＃61; fluid.layers.data(name&＃61;&＃39;image&＃39;, shape&＃61;[1, 28, 28], dtype&＃61;&＃39;float32&＃39;) #单通道&＃xff0c;28*28像素值
label &＃61; fluid.layers.data(name&＃61;&＃39;label&＃39;, shape&＃61;[1], dtype&＃61;&＃39;int64&＃39;) #图片标签
model &＃61; multilayer_perceptron(image)
# 获取损失函数和准确率函数
cost &＃61; fluid.layers.cross_entropy(input&＃61;model, label&＃61;label) #使用交叉熵损失函数,描述真实样本标签和预测概率之间的差值
avg_cost &＃61; fluid.layers.mean(cost)
acc &＃61; fluid.layers.accuracy(input&＃61;model, label&＃61;label)# 定义优化方法
optimizer &＃61; fluid.optimizer.AdamOptimizer(learning_rate&＃61;0.001) #使用Adam算法进行优化
opts &＃61; optimizer.minimize(avg_cost)# 定义一个使用CPU的解析器
place &＃61; fluid.CPUPlace()
exe &＃61; fluid.Executor(place)
# 进行参数初始化
exe.run(fluid.default_startup_program())# 定义输入数据维度
feeder &＃61; fluid.DataFeeder(place&＃61;place, feed_list&＃61;[image, label])# 开始训练和测试
for pass_id in range(5):# 进行训练for batch_id, data in enumerate(train_reader()): # 遍历train_readertrain_cost, train_acc &＃61; exe.run(program&＃61;fluid.default_main_program(), # 运行主程序feed&＃61;feeder.feed(data), # 给模型喂入数据fetch_list&＃61;[avg_cost, acc]) # fetch 误差、准确率# 每100个batch打印一次信息 误差、准确率if batch_id % 100 &＃61;&＃61; 0:print(&＃39;Pass:%d, Batch:%d, Cost:%0.5f, Accuracy:%0.5f&＃39; %(pass_id, batch_id, train_cost[0], train_acc[0]))# 进行测试test_accs &＃61; []test_costs &＃61; []# 每训练一轮 进行一次测试for batch_id, data in enumerate(test_reader()): # 遍历test_readertest_cost, test_acc &＃61; exe.run(program&＃61;fluid.default_main_program(), # 执行训练程序feed&＃61;feeder.feed(data), # 喂入数据fetch_list&＃61;[avg_cost, acc]) # fetch 误差、准确率test_accs.append(test_acc[0]) # 每个batch的准确率test_costs.append(test_cost[0]) # 每个batch的误差# 求测试结果的平均值test_cost &＃61; (sum(test_costs) / len(test_costs)) # 每轮的平均误差test_acc &＃61; (sum(test_accs) / len(test_accs)) # 每轮的平均准确率print(&＃39;Test:%d, Cost:%0.5f, Accuracy:%0.5f&＃39; % (pass_id, test_cost, test_acc))# 保存模型model_save_dir &＃61; "/home/aistudio/data/hand.inference.model"# 如果保存路径不存在就创建if not os.path.exists(model_save_dir):os.makedirs(model_save_dir)print(&＃39;save models to %s&＃39; % (model_save_dir))fluid.io.save_inference_model(model_save_dir, # 保存推理model的路径[&＃39;image&＃39;], # 推理&＃xff08;inference&＃xff09;需要 feed 的数据[model], # 保存推理&＃xff08;inference&＃xff09;结果的 Variablesexe) # executor 保存 inference model# 对图片进行预处理def load_image(file):im &＃61; Image.open(file).convert(&＃39;L&＃39;) # 将RGB转化为灰度图像&＃xff0c;L代表灰度图像&＃xff0c;灰度图像的像素值在0~255之间im &＃61; im.resize((28, 28), Image.ANTIALIAS) # resize image with high-quality 图像大小为28*28im &＃61; np.array(im).reshape(1, 1, 28, 28).astype(np.float32) # 返回新形状的数组,把它变成一个 numpy 数组以匹配数据馈送格式。# print(im)im &＃61; im / 255.0 * 2.0 - 1.0 # 归一化到【-1~1】之间print(im)return imimg &＃61; Image.open(&＃39;data/xu/6.jpg&＃39;)plt.imshow(img) # 根据数组绘制图像plt.show() # 显示图像infer_exe &＃61; fluid.Executor(place)inference_scope &＃61; fluid.core.Scope()# 加载数据并开始预测with fluid.scope_guard(inference_scope):# 获取训练好的模型# 从指定目录中加载 推理model(inference model)[inference_program, # 推理Programfeed_target_names, # 是一个str列表&＃xff0c;它包含需要在推理 Program 中提供数据的变量的名称。fetch_targets] &＃61; fluid.io.load_inference_model(model_save_dir,# fetch_targets&＃xff1a;是一个 Variable 列表&＃xff0c;从中我们可以得到推断结果。model_save_dir&＃xff1a;模型保存的路径infer_exe) # infer_exe: 运行 inference model的 executorimg &＃61; load_image(&＃39;data/yu/6.jpg&＃39;)results &＃61; exe.run(program&＃61;inference_program, # 运行推测程序feed&＃61;{feed_target_names[0]: img}, # 喂入要预测的imgfetch_list&＃61;fetch_targets) # 得到推测结果,# 获取概率最大的labellab &＃61; np.argsort(results) # argsort函数返回的是result数组值从小到大的索引值# print(lab)print("该图片的预测结果的label为: %d" % lab[0][0][-1]) # -1代表读取数组中倒数第一列