import os
from datetime import datetime
import tensorflow as tf
import keras
from tensorflow.keras.optimizers import Adam # 使用TensorFlow的Keras
os.environ["SM_FRAMEWORK"] = "tf.keras" # 显式设置后端
os.environ['TF_ENABLE_ONEDNN_OPTS'] = '0'
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '1'
os.environ['TF_GPU_ALLOCATOR'] = 'cuda_malloc_async'
os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices=false'
# 导入 segmentation_models 并设置框架
import segmentation_models as sm
sm.set_framework('tf.keras') # 确保使用 tf.keras 后端
import cv2
import numpy as np
from matplotlib import pyplot as plt
import matplotlib
from keras.utils.np_utils import to_categorical
from tensorflow.keras.utils import Sequence
from sklearn.metrics import f1_score, precision_score, recall_score, roc_auc_score, matthews_corrcoef
import pandas as pd
from tqdm import tqdm
import gc
from keras.callbacks import EarlyStopping
# 设置 matplotlib 支持中文
plt.rcParams['font.sans-serif'] = ['SimHei', 'Microsoft YaHei', 'Arial Unicode MS']
plt.rcParams['axes.unicode_minus'] = False
# 设置环境变量以优化 GPU 性能
os.environ["SM_FRAMEWORK"] = "tf.keras"
os.environ['TF_ENABLE_ONEDNN_OPTS'] = '0'
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '1'
os.environ['TF_GPU_ALLOCATOR'] = 'cuda_malloc_async'
os.environ['TF_XLA_FLAGS'] = '--tf_xla_enable_xla_devices=false'
# 设置全精度计算
tf.keras.mixed_precision.set_global_policy('float32')
# 设置 GPU 内存增长
gpus = tf.config.list_physical_devices('GPU')
if gpus:
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
else:
print("警告：未检测到 GPU，将使用 CPU 运行")
# 打印环境信息
print("TensorFlow 版本:", tf.__version__)
print("segmentation_models 版本:", sm.__version__)
print("GPU 设备:", tf.config.list_physical_devices('GPU'))
print("CUDA 可用:", tf.test.is_built_with_cuda())
print("cuDNN 可用:", tf.test.is_built_with_gpu_support())
# 设置随机种子
np.random.seed(0)
tf.random.set_seed(0)
# 配置参数
n_classes = 2
SIZE = 1024 # 使用第一份代码的尺寸
base_directory = 'D:/pycharm/spark/data/data/data_1024_1024/' # 划分后的数据集路径
train_directory = os.path.join(base_directory, 'train_data')
val_directory = os.path.join(base_directory, 'val_data')
test_directory = os.path.join(base_directory, 'test_data')
result_directory = 'D:/pycharm/spark/data/data/result3/'
checkpoint_directory = 'D:/pycharm/spark/data/data/duststorm/'
os.makedirs(result_directory, exist_ok=True)
os.makedirs(checkpoint_directory, exist_ok=True)
# 检查目录
for directory in [train_directory, val_directory, test_directory]:
img_dir = os.path.join(directory, 'mdgm1')
mask_dir = os.path.join(directory, 'mask1')
if not os.path.exists(img_dir) or not os.path.exists(mask_dir):
print(f"错误：目录 {img_dir} 或 {mask_dir} 不存在")
exit()
# 数据生成器类
class DataGenerator(Sequence):
def __init__(self, image_paths, mask_paths, preprocess_input, batch_size=4, size=SIZE, n_classes=2):
self.image_paths = image_paths
self.mask_paths = mask_paths
self.preprocess_input = preprocess_input
self.batch_size = batch_size
self.size = size
self.n_classes = n_classes
def __len__(self):
return max(1, int(np.ceil(len(self.image_paths) / self.batch_size)))
def __getitem__(self, idx):
batch_image_paths = self.image_paths[idx * self.batch_size:(idx + 1) * self.batch_size]
batch_mask_paths = self.mask_paths[idx * self.batch_size:(idx + 1) * self.batch_size]
images = np.empty((len(batch_image_paths), self.size, self.size, 3), dtype=np.float32)
masks = np.empty((len(batch_image_paths), self.size, self.size), dtype=np.uint8)
for i, (img_path, mask_path) in enumerate(zip(batch_image_paths, batch_mask_paths)):
img = cv2.imread(img_path, cv2.IMREAD_COLOR)
if img is None:
print(f"警告：无法加载图像 {img_path}")
continue
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = cv2.resize(img, (self.size, self.size), interpolation=cv2.INTER_NEAREST)
images[i] = img
mask = cv2.imread(mask_path, cv2.IMREAD_GRAYSCALE)
if mask is None:
print(f"警告：无法加载掩码 {mask_path}")
continue
mask = cv2.resize(mask, (self.size, self.size), interpolation=cv2.INTER_NEAREST)
mask = np.where(mask > 0, 1, 0).astype(np.uint8)
masks[i] = mask
if not images.size or not masks.size:
print(f"警告：批次 {idx} 为空，返回占位数据")
return np.zeros((1, self.size, self.size, 3), dtype=np.float32), \
np.zeros((1, self.size, self.size, self.n_classes), dtype=np.float32)
images = self.preprocess_input(images)
masks = to_categorical(masks, num_classes=self.n_classes).astype(np.float32)
return images, masks
# 过滤图像文件
def is_image_file(filename):
return filename.lower().endswith(('.png', '.jpg', '.jpeg', '.bmp'))
# 加载文件路径
def load_paths(directory):
img_dir = os.path.join(directory, 'mdgm1')
mask_dir = os.path.join(directory, 'mask1')
image_files = sorted([f for f in os.listdir(img_dir) if is_image_file(f)])
mask_files = sorted([f for f in os.listdir(mask_dir) if is_image_file(f)])
common_files = sorted(set(image_files) & set(mask_files))
image_paths = [os.path.join(img_dir, fname) for fname in common_files]
mask_paths = [os.path.join(mask_dir, fname) for fname in common_files]
print(f"{directory}：找到 {len(common_files)} 个匹配文件对")
return image_paths, mask_paths, common_files
# 加载数据集
train_image_paths, train_mask_paths, train_names = load_paths(train_directory)
val_image_paths, val_mask_paths, val_names = load_paths(val_directory)
test_image_paths, test_mask_paths, test_names = load_paths(test_directory)
# 创建数据生成器
BACKBONE = 'inceptionv3'
preprocess_input = sm.get_preprocessing(BACKBONE)
train_generator = DataGenerator(train_image_paths, train_mask_paths, preprocess_input, batch_size=4)
val_generator = DataGenerator(val_image_paths, val_mask_paths, preprocess_input, batch_size=4)
test_generator = DataGenerator(test_image_paths, test_mask_paths, preprocess_input, batch_size=4)
# 模型参数
activation = 'sigmoid'
LR = 1e-5 # 使用第二份代码的学习率
optim = Adam(LR, clipnorm=1.0)
metrics = [sm.metrics.IOUScore(threshold=0.5), sm.metrics.FScore(threshold=0.5)]
loss = sm.losses.BinaryFocalLoss(alpha=0.25, gamma=2.0) + sm.losses.DiceLoss(class_weights=[0.2, 0.8])
# 加载权重
weights_path = 'D:/pycharm/spark/data/data/inceptionv3/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5'
try:
if not os.path.exists(weights_path):
print(f"错误：权重文件 {weights_path} 不存在，使用 ImageNet 权重")
model = sm.Unet(BACKBONE, encoder_weights='imagenet', classes=n_classes, activation=activation)
else:
model = sm.Unet(BACKBONE, encoder_weights=weights_path, classes=n_classes, activation=activation)
except Exception as e:
print(f"加载权重失败: {e}，使用随机初始化")
model = sm.Unet(BACKBONE, encoder_weights=None, classes=n_classes, activation=activation)
# 编译模型
with tf.device('/GPU:0'):
model.compile(optim, loss=loss, metrics=metrics)
print(model.summary())
# 设置回调
timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
checkpoint_filepath = f'{checkpoint_directory}/train_MY30_31_inceptionv3_{timestamp}.keras'
callbacks = [
tf.keras.callbacks.ModelCheckpoint(
filepath=checkpoint_filepath,
monitor='val_iou_score',
mode='max',
save_best_only=True,
verbose=1
),
EarlyStopping(
monitor='val_iou_score',
mode='max',
patience=5,
restore_best_weights=True,
verbose=1
),
tf.keras.callbacks.ReduceLROnPlateau(
monitor='val_loss',
factor=0.5,
patience=3,
min_lr=1e-6,
verbose=1
)
]
# 训练模型
try:
with tf.device('/GPU:0'):
history = model.fit(
train_generator,
epochs=50,
validation_data=val_generator,
callbacks=callbacks,
verbose=1
)
except Exception as e:
print(f"训练失败: {e}")
exit()
# 绘制训练和验证曲线
def plot_metrics(history, metric_name, title, ylabel):
metric = history.history[metric_name]
val_metric = history.history[f'val_{metric_name}']
epochs = range(1, len(metric) + 1)
plt.figure(figsize=(10, 5))
plt.plot(epochs, metric, 'y', label=f'训练 {ylabel}')
plt.plot(epochs, val_metric, 'r', label=f'验证 {ylabel}')
plt.title(title)
plt.xlabel('轮次')
plt.ylabel(ylabel)
plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left')
plt.tight_layout()
plt.show()
plot_metrics(history, 'loss', '训练和验证损失 train_MY30_31_inceptionv3', '损失')
plot_metrics(history, 'iou_score', '训练和验证 IoU train_MY30_31_inceptionv3', 'IoU')
# 保存训练指标
with open(f'{checkpoint_directory}/train_MY30_31_inceptionv3_{timestamp}.csv', 'w', newline='') as f:
writer = csv.writer(f)
writer.writerow(['训练损失', '验证损失', '训练 IoU', '验证 IoU'])
for row in zip(history.history['loss'], history.history['val_loss'],
history.history['iou_score'], history.history['val_iou_score']):
writer.writerow(row)
# 测试集评估
def calculate_iou(pred_mask, gt_mask):
intersection = np.logical_and(pred_mask, gt_mask).sum()
union = np.logical_or(pred_mask, gt_mask).sum()
return intersection / union if union != 0 else 1.0
def pixel_accuracy(y_true, y_pred):
y_pred = tf.round(y_pred)
correct_pixels = tf.reduce_sum(tf.cast(tf.equal(y_true, y_pred), tf.float32))
total_pixels = tf.cast(tf.size(y_true), tf.float32)
return (correct_pixels / total_pixels).numpy()
def pixel_accuracy_for_class(y_true, y_pred, class_id):
y_pred_rounded = tf.cast(tf.round(y_pred), tf.int32)
true_mask = tf.equal(y_true, class_id)
pred_mask = tf.equal(y_pred_rounded, class_id)
correct_pixels_class = tf.reduce_sum(tf.cast(tf.logical_and(true_mask, pred_mask), tf.float32))
total_pixels_class = tf.reduce_sum(tf.cast(true_mask, tf.float32))
return (correct_pixels_class / total_pixels_class).numpy() if total_pixels_class != 0 else 0.0
# 测试集预测和评估
metrics_dict = {
'文件名': [], 'IoU 分数': [], 'AUC 分数': [], 'F1 分数': [],
'F1 微平均': [], 'F1 宏平均': [], '精确率 (宏平均)': [],
'召回率 (宏平均)': [], '像素精度 (整体)': [], '像素精度 (尘暴)': [], 'MCC': []
}
for i in tqdm(range(len(test_image_paths)), desc="测试集预测"):
test_img = cv2.imread(test_image_paths[i], cv2.IMREAD_COLOR)
if test_img is None:
print(f"警告：无法加载测试图像 {test_image_paths[i]}")
continue
test_img = cv2.cvtColor(test_img, cv2.COLOR_BGR2RGB)
test_img = cv2.resize(test_img, (SIZE, SIZE), interpolation=cv2.INTER_NEAREST)
test_img_input = preprocess_input(np.expand_dims(test_img, 0))
test_mask = cv2.imread(test_mask_paths[i], cv2.IMREAD_GRAYSCALE)
if test_mask is None:
print(f"警告：无法加载测试掩码 {test_mask_paths[i]}")
continue
test_mask = cv2.resize(test_mask, (SIZE, SIZE), interpolation=cv2.INTER_NEAREST)
test_mask = np.where(test_mask > 0, 1, 0).astype(np.uint8)
with tf.device('/GPU:0'):
test_pred = model.predict(test_img_input, verbose=0)
test_prediction = np.argmax(test_pred, axis=3)[0, :, :]
iou = calculate_iou(test_prediction, test_mask)
auc_score = roc_auc_score(test_mask.flatten(), test_pred[0, :, :, 1].flatten())
f1 = f1_score(test_mask.flatten(), test_prediction.flatten(), average='binary', zero_division=1)
f1_micro = f1_score(test_mask.flatten(), test_prediction.flatten(), average='micro', zero_division=1)
f1_macro = f1_score(test_mask.flatten(), test_prediction.flatten(), average='macro', zero_division=1)
precision = precision_score(test_mask.flatten(), test_prediction.flatten(), average='macro', zero_division=0)
recall = recall_score(test_mask.flatten(), test_prediction.flatten(), average='macro', zero_division=0)
pa = pixel_accuracy(test_mask, test_prediction)
pa_dust = pixel_accuracy_for_class(test_mask, test_prediction, 1)
mcc = matthews_corrcoef(test_mask.flatten(), test_prediction.flatten())
metrics_dict['文件名'].append(test_names[i])
metrics_dict['IoU 分数'].append(iou)
metrics_dict['AUC 分数'].append(auc_score)
metrics_dict['F1 分数'].append(f1)
metrics_dict['F1 微平均'].append(f1_micro)
metrics_dict['F1 宏平均'].append(f1_macro)
metrics_dict['精确率 (宏平均)'].append(precision)
metrics_dict['召回率 (宏平均)'].append(recall)
metrics_dict['像素精度 (整体)'].append(pa)
metrics_dict['像素精度 (尘暴)'].append(pa_dust)
metrics_dict['MCC'].append(mcc)
# 可视化
plt.figure(figsize=(15, 5))
plt.subplot(1, 3, 1)
plt.title('图像')
plt.imshow(test_img)
plt.axis('off')
plt.subplot(1, 3, 2)
plt.title('图像 + 掩码')
plt.imshow(test_img)
plt.imshow(test_mask, cmap='jet', alpha=0.5)
plt.axis('off')
plt.subplot(1, 3, 3)
plt.title('预测 + 掩码 + IoU')
plt.imshow(test_prediction, cmap='gray')
plt.imshow(test_mask, cmap='jet', alpha=0.5)
plt.text(0.5, -0.1, f'IoU: {iou:.3f}', ha='center', va='center', transform=plt.gca().transAxes, fontsize=10)
plt.axis('off')
plt.savefig(os.path.join(result_directory, f'test_{test_names[i]}'), bbox_inches='tight')
plt.close()
# 保存测试指标
pd.DataFrame(metrics_dict).to_csv(f'{checkpoint_directory}/test_MY30_31_inceptionv3_metrics_{timestamp}.csv', index=False)
现在是可以运行，但是运行不是用gpu运行，是使用cpu运行
最新发布
本文解决了TensorFlow运行时找不到cupti64_80.dll的问题，详细步骤包括检查CUDA文件夹，确认cupti64_80.dll存在，调整环境变量，将该DLL复制到CUDA8.0/bin目录。
扫一扫
30万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
