DAY 36 复习日

仔细回顾一下神经网络到目前的内容，没跟上进度的同学补一下进度。

• **作业：**对之前的信贷项目，利用神经网络训练下，尝试用到目前的知识点让代码更加规范和美观。

在这里我对之前的迁移学习项目进行了复现

#smote改进

import glob
import os.path
import random
import numpy as np
import tensorflow as tf
from tensorflow.io.gfile import GFile
# 新增SMOTE所需库
from imblearn.over_sampling import SMOTE
from collections import defaultdict

# -------------------------- 常量定义 --------------------------
BOTTLENECK_TENSOR_SIZE = 2048
BOTTLENECK_TENSOR_NAME = 'pool_3/_reshape:0'
JPEG_DATA_TENSOR_NAME = 'DecodeJpeg/contents:0'

MODEL_DIR = r"E:\项目学习内容\迁移学习\time1数据集\inception-2015-12-05 (1)\\"
os.makedirs(MODEL_DIR, exist_ok=True)
MODEL_FILE = 'tensorflow_inception_graph.pb'

CACHE_DIR = './images/tmp/bottleneck/'
os.makedirs(CACHE_DIR, exist_ok=True)

INPUT_DATA = r"E:\项目学习内容\迁移学习\time1数据集\time11\GC10-DET\data"
VALIDATION_PERCENTAGE = 10
TEST_PERCENTAGE = 10

LEARNING_RATE = 0.01
STEPS = 4000
BATCH = 100

# -------------------------- 数据预处理函数 --------------------------
def create_image_lists(testing_percentage, validation_percentage):
    result = {}
    sub_dirs = [x[0] for x in os.walk(INPUT_DATA)]
    is_root_dir = True
    for sub_dir in sub_dirs:
        if is_root_dir:
            is_root_dir = False
            continue

        extensions = ['jpg', 'jpeg', 'JPG', 'JPEG']
        file_list = []
        dir_name = os.path.basename(sub_dir)
        for extension in extensions:
            file_glob = os.path.join(INPUT_DATA, dir_name, f'*.{extension}')
            file_list.extend(glob.glob(file_glob))
        if not file_list:
            continue

        label_name = dir_name.lower()
        training_images = []
        testing_images = []
        validation_images = []
        for file_name in file_list:
            base_name = os.path.basename(file_name)
            chance = np.random.randint(100)
            if chance < validation_percentage:
                validation_images.append(base_name)
            elif chance < (testing_percentage + validation_percentage):
                testing_images.append(base_name)
            else:
                training_images.append(base_name)

        result[label_name] = {
            'dir': dir_name,
            'training': training_images,
            'testing': testing_images,
            'validation': validation_images
        }
    return result


def get_image_path(image_lists, image_dir, label_name, index, category):
    label_lists = image_lists[label_name]
    category_list = label_lists[category]
    mod_index = index % len(category_list)
    base_name = category_list[mod_index]
    sub_dir = label_lists['dir']
    return os.path.join(image_dir, sub_dir, base_name)


def get_bottleneck_path(image_lists, label_name, index, category):
    return get_image_path(image_lists, CACHE_DIR, label_name, index, category) + '.txt'


# -------------------------- 特征提取函数 --------------------------
def run_bottleneck_on_image(sess, image_data, image_data_tensor, bottleneck_tensor):
    bottleneck_values = sess.run(bottleneck_tensor, feed_dict={image_data_tensor: image_data})
    return np.squeeze(bottleneck_values)


def get_or_create_bottleneck(sess, image_lists, label_name, index, category, jpeg_data_tensor, bottleneck_tensor):
    label_lists = image_lists[label_name]
    sub_dir = label_lists['dir']
    sub_dir_path = os.path.join(CACHE_DIR, sub_dir)
    os.makedirs(sub_dir_path, exist_ok=True)

    bottleneck_path = get_bottleneck_path(image_lists, label_name, index, category)
    if os.path.exists(bottleneck_path):
        with open(bottleneck_path, 'r') as f:
            bottleneck_string = f.read()
        return [float(x) for x in bottleneck_string.split(',')]

    image_path = get_image_path(image_lists, INPUT_DATA, label_name, index, category)
    with GFile(image_path, 'rb') as f:
        image_data = f.read()
    bottleneck_values = run_bottleneck_on_image(sess, image_data, jpeg_data_tensor, bottleneck_tensor)
    bottleneck_string = ','.join(str(x) for x in bottleneck_values)
    with open(bottleneck_path, 'w') as f:
        f.write(bottleneck_string)

    return bottleneck_values


# -------------------------- 训练数据准备（核心修改：SMOTE过采样） --------------------------
def prepare_smote_training_data(sess, image_lists, n_classes, jpeg_data_tensor, bottleneck_tensor):
    """提前收集所有训练集特征和标签，用SMOTE生成平衡样本"""
    all_bottlenecks = []
    all_labels = []
    label_name_list = list(image_lists.keys())
    
    # 收集原始训练数据
    for label_index, label_name in enumerate(label_name_list):
        for index in range(len(image_lists[label_name]['training'])):
            bottleneck = get_or_create_bottleneck(
                sess, image_lists, label_name, index, 'training', jpeg_data_tensor, bottleneck_tensor
            )
            all_bottlenecks.append(bottleneck)
            label = np.zeros(n_classes, dtype=np.float32)
            label[label_index] = 1.0
            all_labels.append(label)
    
    # 转换为numpy数组
    X = np.array(all_bottlenecks)
    y = np.array(all_labels).argmax(axis=1)  # SMOTE需要整数标签
    
    # 应用SMOTE过采样（关键步骤）
    smote = SMOTE(random_state=42)
    X_resampled, y_resampled = smote.fit_resample(X, y)
    
    # 转回独热编码
    y_resampled_onehot = np.zeros((len(y_resampled), n_classes), dtype=np.float32)
    y_resampled_onehot[np.arange(len(y_resampled)), y_resampled] = 1.0
    
    return X_resampled, y_resampled_onehot


def get_smote_batch(smote_features, smote_labels, batch_size):
    """从SMOTE处理后的平衡数据中随机取批次"""
    indices = np.random.choice(len(smote_features), batch_size, replace=False)
    return smote_features[indices], smote_labels[indices]


# -------------------------- 测试数据准备 --------------------------
def get_test_bottlenecks(sess, image_lists, n_classes, jpeg_data_tensor, bottleneck_tensor):
    bottlenecks = []
    ground_truths = []
    label_name_list = list(image_lists.keys())
    for label_index, label_name in enumerate(label_name_list):
        category = 'testing'
        for index in range(len(image_lists[label_name][category])):
            bottleneck = get_or_create_bottleneck(
                sess, image_lists, label_name, index, category, jpeg_data_tensor, bottleneck_tensor
            )
            ground_truth = np.zeros(n_classes, dtype=np.float32)
            ground_truth[label_index] = 1.0
            bottlenecks.append(bottleneck)
            ground_truths.append(ground_truth)
    return np.array(bottlenecks), np.array(ground_truths)


# -------------------------- 主函数 --------------------------
def main(_):
    image_lists = create_image_lists(TEST_PERCENTAGE, VALIDATION_PERCENTAGE)
    n_classes = len(image_lists.keys())
    if n_classes == 0:
        raise ValueError(f"目标域数据集路径 {INPUT_DATA} 下无类别文件夹，请检查！")

    model_path = os.path.join(MODEL_DIR, MODEL_FILE)
    if not os.path.exists(model_path):
        raise FileNotFoundError(f"源域模型 {model_path} 不存在，请下载并放置到该路径！")

    with GFile(model_path, 'rb') as f:
        graph_def = tf.compat.v1.GraphDef()
        graph_def.ParseFromString(f.read())

    with tf.compat.v1.Session() as sess:
        bottleneck_tensor, jpeg_data_tensor = tf.compat.v1.import_graph_def(
            graph_def, return_elements=[BOTTLENECK_TENSOR_NAME, JPEG_DATA_TENSOR_NAME]
        )

        # 定义模型结构
        bottleneck_input = tf.compat.v1.placeholder(tf.float32, [None, BOTTLENECK_TENSOR_SIZE], name='BottleneckInputPlaceholder')
        ground_truth_input = tf.compat.v1.placeholder(tf.float32, [None, n_classes], name='GroundTruthInput')

        with tf.compat.v1.name_scope('final_training_ops'):
            weights = tf.compat.v1.Variable(tf.random.truncated_normal([BOTTLENECK_TENSOR_SIZE, n_classes], stddev=0.001))
            biases = tf.compat.v1.Variable(tf.zeros([n_classes]))
            logits = tf.matmul(bottleneck_input, weights) + biases
            final_tensor = tf.nn.softmax(logits)

        # 定义损失和优化器
        cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=ground_truth_input)
        cross_entropy_mean = tf.reduce_mean(cross_entropy)
        train_step = tf.compat.v1.train.GradientDescentOptimizer(LEARNING_RATE).minimize(cross_entropy_mean)

        # 定义评估指标
        with tf.compat.v1.name_scope('evaluation'):
            correct_prediction = tf.equal(tf.argmax(final_tensor, 1), tf.argmax(ground_truth_input, 1))
            evaluation_step = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

        sess.run(tf.compat.v1.global_variables_initializer())

        # 提前生成SMOTE平衡数据（关键修改）
        smote_features, smote_labels = prepare_smote_training_data(
            sess, image_lists, n_classes, jpeg_data_tensor, bottleneck_tensor
        )

        # 训练过程
        for i in range(STEPS):
            # 使用SMOTE平衡数据训练（关键修改）
            train_bottlenecks, train_ground_truth = get_smote_batch(smote_features, smote_labels, BATCH)
            sess.run(train_step, feed_dict={
                bottleneck_input: train_bottlenecks,
                ground_truth_input: train_ground_truth
            })

            if i % 100 == 0 or i + 1 == STEPS:
                # 验证集仍用原始分布（避免数据泄露）
                validation_bottlenecks = []
                validation_ground_truth = []
                label_name_list = list(image_lists.keys())
                for _ in range(BATCH):
                    label_index = random.randrange(n_classes)
                    label_name = label_name_list[label_index]
                    image_index = random.randrange(65536)
                    bottleneck = get_or_create_bottleneck(
                        sess, image_lists, label_name, image_index, 'validation', jpeg_data_tensor, bottleneck_tensor
                    )
                    ground_truth = np.zeros(n_classes, dtype=np.float32)
                    ground_truth[label_index] = 1.0
                    validation_bottlenecks.append(bottleneck)
                    validation_ground_truth.append(ground_truth)
                validation_acc = sess.run(evaluation_step, feed_dict={
                    bottleneck_input: validation_bottlenecks,
                    ground_truth_input: validation_ground_truth
                })
                print(f"Step {i:4d}: 验证集准确率 = {validation_acc * 100:.1f}%")

        # 测试
        test_bottlenecks, test_ground_truth = get_test_bottlenecks(
            sess, image_lists, n_classes, jpeg_data_tensor, bottleneck_tensor
        )
        test_acc = sess.run(evaluation_step, feed_dict={
            bottleneck_input: test_bottlenecks,
            ground_truth_input: test_ground_truth
        })
        print(f"\n最终测试集准确率 = {test_acc * 100:.1f}%")


if __name__ == '__main__':
    tf.compat.v1.app.run(main=main)

@浙大疏锦行