知识点回顾
- 图像数据的格式：灰度和彩色数据
- 模型的定义
- 显存占用的4种地方
- 模型参数+梯度参数
- 优化器参数
- 数据批量所占显存
- 神经元输出中间状态
在这里，我对之前复现的项目进行了改进
# 导入必要库：文件操作、随机数、数值计算、TensorFlow
import glob
import os.path
import random
import numpy as np
import tensorflow as tf
from tensorflow.io.gfile import GFile
# -------------------------- 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
# 焦点损失参数
GAMMA = 2.0 # 调节因子
ALPHA = 0.25 # 平衡因子
# -------------------------- 数据预处理函数（添加过采样） --------------------------
def create_image_lists(testing_percentage, validation_percentage):
"""功能：读取数据集并按比例划分，添加过采样解决不平衡问题"""
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
}
# 添加过采样：平衡训练集
max_training_count = max(len(v['training']) for v in result.values())
for label_name, data in result.items():
training_count = len(data['training'])
if training_count < max_training_count:
# 计算需要过采样的数量
oversample_count = max_training_count - training_count
# 随机复制现有样本（过采样）
data['training'].extend(random.choices(data['training'], k=oversample_count))
# 打印各类别样本数（调试用）
print("\n数据集类别分布（过采样后）：")
for label, data in result.items():
print(f"类别 {label}: 训练集={len(data['training'])}张, 验证集={len(data['validation'])}张, 测试集={len(data['testing'])}张")
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'
# -------------------------- 迁移学习核心 --------------------------
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
def get_random_cached_bottlenecks(sess, n_classes, image_lists, how_many, category, jpeg_data_tensor, bottleneck_tensor):
bottlenecks = []
ground_truths = []
for _ in range(how_many):
label_index = random.randrange(n_classes)
label_name = list(image_lists.keys())[label_index]
image_index = random.randrange(65536)
bottleneck = get_or_create_bottleneck(sess, image_lists, label_name, image_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)
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)
# -------------------------- 焦点损失函数 --------------------------
def focal_loss(labels, logits, class_weights, gamma=GAMMA, alpha=ALPHA):
"""焦点损失实现，解决类别不平衡问题"""
# 计算softmax概率
probs = tf.nn.softmax(logits)
# 计算焦点因子
p_t = tf.reduce_sum(labels * probs, axis=1)
modulating_factor = tf.pow(1.0 - p_t, gamma)
# 计算类别权重
alpha_factor = tf.reduce_sum(labels * class_weights, axis=1)
# 计算交叉熵
ce = tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits)
# 组合焦点损失
focal_loss = alpha_factor * modulating_factor * ce
return tf.reduce_mean(focal_loss)
# -------------------------- 主函数 --------------------------
def main(_):
# 1. 加载数据集（已添加过采样）
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！")
# 2. 计算类权重（用于焦点损失）
class_counts = {label: len(image_lists[label]['training']) for label in image_lists}
total_samples = sum(class_counts.values())
class_weights = [total_samples / (n_classes * class_counts[label]) for label in image_lists]
class_weights_tensor = tf.constant([class_weights], dtype=tf.float32)
print("\n类权重（用于焦点损失）：")
for label, weight in zip(image_lists.keys(), class_weights):
print(f"类别 {label}: 权重 = {weight:.2f}")
# 3. 加载预训练模型
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:
# 导入源域模型
bottleneck_tensor, jpeg_data_tensor = tf.compat.v1.import_graph_def(
graph_def,
return_elements=[BOTTLENECK_TENSOR_NAME, JPEG_DATA_TENSOR_NAME]
)
# 4. 定义目标域模型
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'
)
# 模型结构
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)
# 5. 使用焦点损失代替标准交叉熵
cross_entropy_mean = focal_loss(
ground_truth_input,
logits,
class_weights_tensor,
gamma=GAMMA,
alpha=ALPHA
)
# 6. 动态学习率（指数衰减）
global_step = tf.compat.v1.train.get_or_create_global_step()
learning_rate = tf.compat.v1.train.exponential_decay(
LEARNING_RATE,
global_step,
1000, # 每1000步衰减一次
0.96, # 衰减率
staircase=True # 阶梯式衰减
)
# 使用Adam优化器
optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate)
train_step = optimizer.minimize(cross_entropy_mean, global_step=global_step)
# 7. 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))
# 初始化变量
sess.run(tf.compat.v1.global_variables_initializer())
# 8. 训练循环
print("\n开始训练...")
best_validation_acc = 0.0
for i in range(STEPS):
# 获取训练批次
train_bottlenecks, train_ground_truth = get_random_cached_bottlenecks(
sess, n_classes, image_lists, BATCH, 'training',
jpeg_data_tensor, bottleneck_tensor
)
# 执行训练
_, current_lr = sess.run([train_step, learning_rate], feed_dict={
bottleneck_input: train_bottlenecks,
ground_truth_input: train_ground_truth
})
# 定期评估验证集
if i % 100 == 0 or i + 1 == STEPS:
validation_bottlenecks, validation_ground_truth = get_random_cached_bottlenecks(
sess, n_classes, image_lists, BATCH, 'validation',
jpeg_data_tensor, bottleneck_tensor
)
validation_acc = sess.run(evaluation_step, feed_dict={
bottleneck_input: validation_bottlenecks,
ground_truth_input: validation_ground_truth
})
# 保存最佳模型（可选）
if validation_acc > best_validation_acc:
best_validation_acc = validation_acc
print(f"Step {i:4d}: 验证集准确率 = {validation_acc*100:.1f}% | "
f"学习率 = {current_lr:.6f} | "
f"最佳验证准确率 = {best_validation_acc*100:.1f}%")
# 9. 最终测试
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}%")
# -------------------------- 程序入口 --------------------------
if __name__ == '__main__':
tf.compat.v1.app.run(main=main)@浙大疏锦行