import tensorflow as tf def read_img(path): return tf.image.decode_image(tf.io.read_file(path)) def psnr(tf_img1, tf_img2): return tf.image.psnr(tf_img1, tf_img2, max_val=255) def _main(): t1 = read_img('D:/SteganoGAN-master/research/ruanjian.png') t2 = read_img('D:/SteganoGAN-master/research/outtt.png') with tf.Session() as sess: sess.run(tf.global_variables_initializer()) y = sess.run(psnr(t1, t2)) print(y) if name == 'main': _main()

时间: 2023-07-16 18:12:26 浏览: 84

这段代码使用TensorFlow计算两张图片之间的PSNR值。具体来说，它使用了TensorFlow中的tf.image.decode_image()函数读取图片，并将其作为参数传递给tf.image.psnr()函数，该函数计算两张图片的PSNR值。最后，使用TensorFlow会话来执行计算并打印结果。需要注意的是，这段代码需要运行在已经安装了TensorFlow的环境中。

import tensorflow as tf from tensorflow import keras from keras import layers import xml.etree.ElementTree as ET import pathlib from pathlib import Path file_path = Path('C:/1') def net_init(): model = keras.Sequential([layers.Input(shape=(1200, 1600, 3))]) model.add(layers.Conv2D(filters=3, activation="relu", kernel_size=(3, 3), padding="same", strides=2)) model.add(layers.MaxPool2D(pool_size=(2, 2))) model.add(layers.Conv2D(filters=3, activation="relu", kernel_size=(3, 3), padding="same", strides=2)) model.add(layers.MaxPool2D(pool_size=(2, 2))) model.add(layers.Conv2D(filters=1, activation="relu", kernel_size=(3, 3), padding="same", strides=2)) model.add(layers.MaxPool2D(pool_size=(2, 2))) model.add(layers.Dense(48, activation='relu')) model.add(layers.Dense(2, activation='softmax')) return model def load_xml(folder_path: Path) -> list: feature_list = [] file_list = [] label_list = [] for file_name in folder_path.glob('.xml'): xml_tree = ET.parse(file_name) root = xml_tree.getroot() feature = ( int(root.find('object/bndbox/xmin').text), int(root.find('object/bndbox/ymin').text), int(root.find('object/bndbox/xmax').text), int(root.find('object/bndbox/ymax').text) ) feature_list.append(feature) file_list.append(file_name) label_list.append(root.find('object/name').text) return feature_list, file_list, label_list def load_img(folder_path : Path, xml_list : list): img_list = [] print(xml_list) for img_name in folder_path.glob('.jpg'): print(img_name) xml_name = img_name.with_suffix('.xml') print(xml_name) if xml_name in xml_list: print("yes") img = tf.io.read_file(img_name.as_posix()) img = tf.image.decode_image(img, channels=3) img = tf.image.per_image_standardization(img) img_list.append(img) return img_list def main(): feature_list, file_list, label_list = load_xml(file_path) img_list = load_img(file_path, file_list) model = net_init() model.compile(optimizer='adam', loss=tf.keras.losses.mse, metrics=['accuracy']) model.fit(img_list, feature_list, epochs=1) main()这段程序有什么问题

这段程序可能会有以下问题： 1. 程序中没有明确指定输入的图像尺寸，而模型中的第一层需要接受一个 (1200, 1600, 3) 的输入，可能会导致输入图像的尺寸不符合要求而导致程序出错。 2. 程序中使用了标准化函数 tf.image.per_image_standardization() 对输入图像进行标准化处理，但这个函数会将图像像素值减去均值再除以标准差，这有可能会导致图像像素值超出了合理范围，需要使用其他的标准化方法。 3. 程序中编译模型时，损失函数使用的是均方误差（MSE），但这个损失函数通常用于回归问题，对于分类问题来说，应该使用交叉熵损失函数。 4. 程序中训练模型时，只训练了一个 epoch，这可能会导致模型未能充分学习特征，需要增加训练 epochs 的数量。

import tensorflow as tf import os import matplotlib.pyplot as plt from time import time # 定义标签字典（根据实际类别修改） label_dict = { 'electrodrill': 0, 'headphones': 1, 'keyboard': 2, 'mobile_phone': 3, 'monitor': 4, 'mouse': 5, 'multimeter': 6, 'number': 7, 'oscillograph': 8, 'pliers': 9, 'printer': 10, 'screwdriver': 11, 'soldering_iron': 12, 'speaker': 13, 'tape_measure': 14, 'wrench': 15 } def data_load(data_dir, test_data_dir, img_height, img_width, batch_size): def process_image(image_path): image = tf.io.read_file(image_path) image = tf.image.decode_jpeg(image, channels=3) image = tf.image.resize(image, [img_height, img_width]) image = (image / 127.5) - 1.0 # MobileNet标准化 return image def parse_path(path): # 转换为numpy字符串处理路径 path_str = path.numpy().decode('utf-8') # 获取类别文件夹名 class_name = tf.strings.split(path_str, os.path.sep)[-2].numpy().decode('utf-8') # 从预定义的字典获取标签索引 label_idx = label_dict.get(class_name, -1) # -1表示未知类别 if label_idx == -1: raise ValueError(f"未知类别: {class_name}") return process_image(path), label_idx def map_fn(path): # 使用py_function包装Python逻辑 image, label = tf.py_function( func=parse_path, inp=[path], Tout=[tf.float32, tf.int32] ) # 设置明确的Tensor形状 image.set_shape([img_height, img_width, 3]) label.set_shape([]) # 将标签转换为one-hot编码 label = tf.one_hot(label, depth=len(label_dict)) return image, label def load_dataset(directory): # 获取所有图片路径 dataset = tf.data.Dataset.list_files(directory + '//.jpg', shuffle=True) # 应用处理函数 dataset = dataset.map( map_fn, num_parallel_calls=tf.data.experimental.AUTOTUNE ) return dataset # 加载数据集 train_ds = load_dataset(data_dir) val_ds = load_dataset(test_data_dir) # 批处理和预取 train_ds = train_ds.batch(batch_size).prefetch(tf.data.experimental.AUTOTUNE) val_ds = val_ds.batch(batch_size).prefetch(tf.data.experimental.AUTOTUNE) # 验证数据预处理是否正确 for images, labels in train_ds.take(1): # 检查图像标准化是否正确 min_value = tf.reduce_min(images) max_value = tf.reduce_max(images) print(f"图像标准化检查: 最小值 = {min_value.numpy()}, 最大值 = {max_value.numpy()}") assert min_value >= -1 and max_value <= 1, "图像标准化错误，范围应为[-1, 1]" # 检查标签是否为one-hot编码且正确 print("标签示例:", labels[0].numpy()) # 应为one-hot如[0,0,1,...,0] assert tf.reduce_sum(labels[0]).numpy() == 1, "标签应该是one-hot编码，其中只有一个值为1，其余为0" return train_ds, val_ds def model_load(IMG_SHAPE=(224, 224, 3), class_num=16, learning_rate=0.01): # 添加learning_rate参数 base_model = tf.keras.applications.MobileNetV2( input_shape=IMG_SHAPE, include_top=False, weights='imagenet' ) base_model.trainable = False model = tf.keras.Sequential([ base_model, tf.keras.layers.GlobalAveragePooling2D(), tf.keras.layers.Dense(class_num, activation='softmax') ]) # 显式设置学习率的优化器 optimizer = tf.keras.optimizers.Adam(learning_rate=learning_rate) model.compile( optimizer=optimizer, # 使用自定义优化器 loss='categorical_crossentropy', metrics=['accuracy'] ) model.summary() return model def show_loss_acc(history): acc = history.history['accuracy'] val_acc = history.history['val_accuracy'] loss = history.history['loss'] val_loss = history.history['val_loss'] plt.figure(figsize=(8, 8)) plt.subplot(2, 1, 1) plt.plot(acc, label='Training Accuracy') plt.plot(val_acc, label='Validation Accuracy') plt.legend(loc='lower right') plt.ylabel('Accuracy') plt.ylim([min(plt.ylim()), 1]) plt.title('Training and Validation Accuracy') plt.subplot(2, 1, 2) plt.plot(loss, label='Training Loss') plt.plot(val_loss, label='Validation Loss') plt.legend(loc='upper right') plt.ylabel('Cross Entropy') plt.title('Training and Validation Loss') plt.xlabel('epoch') plt.savefig('results/results_mobilenet.png', dpi=100) def train(epochs): begin_time = time() # 创建必要目录 os.makedirs("models", exist_ok=True) os.makedirs("results", exist_ok=True) try: print("加载数据集中...") train_ds, val_ds = data_load( "C:/Users/dll20/Desktop/vegetables_tf2.3-master/new_data/train", "C:/Users/dll20/Desktop/vegetables_tf2.3-master/new_data/val", 224, 224, 16 ) # 验证数据加载 for images, labels in train_ds.take(1): print(f"图像形状: {images.shape}, 标签形状: {labels.shape}") print(f"标签示例: {labels[0].numpy()}") print("类别数量:", len(label_dict)) print("类别映射:", label_dict) model = model_load(class_num=len(label_dict)) print("开始训练...") history = model.fit( train_ds, validation_data=val_ds, epochs=epochs, verbose=1 ) model.save("models/mobilenet_engineer.h5") show_loss_acc(history) except Exception as e: print(f"训练出错: {str(e)}") import traceback traceback.print_exc() finally: print(f"总耗时: {time() - begin_time:.2f}秒") if name == 'main': # 配置TensorFlow tf.config.run_functions_eagerly(False) physical_devices = tf.config.list_physical_devices('GPU') if physical_devices: tf.config.experimental.set_memory_growth(physical_devices[0], True) 图像标准化检查: 最小值 = -1.0, 最大值 = 1.0 标签示例: [0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0.] 图像形状: (16, 224, 224, 3), 标签形状: (16, 16) 标签示例: [0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0.] 类别数量: 16 类别映射: {'electrodrill': 0, 'headphones': 1, 'keyboard': 2, 'mobile_phone': 3, 'monitor': 4, 'mouse': 5, 'multimeter': 6, 'number': 7, 'oscillograph': 8, 'pliers': 9, 'printer': 10, 'screwdriver': 11, 'soldering_iron': 12, 'speaker': 13, 'tape_measure': 14, 'wrench': 15} Model: "sequential" _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= mobilenetv2_1.00_224 (Functi (None, 7, 7, 1280) 2257984 _________________________________________________________________ global_average_pooling2d (Gl (None, 1280) 0 _ dense (Dense) (None, 16) 20496 ================================================================= Total params: 2,278,480 Trainable params: 20,496 Non-trainable params: 2,257,984 _ 开始训练... Epoch 1/100 2025-05-17 20:14:57.943383: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cublas64_10.dll 2025-05-17 20:16:05.881342: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cudnn64_7.dll 2025-05-17 20:19:28.437448: W tensorflow/stream_executor/gpu/redzone_allocator.cc:314] Internal: Invoking GPU asm compilation is supported on Cuda non-Windows platforms only Relying on driver to perform ptx compilation. Modify $PATH to customize ptxas location. This message will be only logged once. 1024/1024 [==============================] - 53s 52ms/step - loss: 9.9016 - accuracy: 0.0606 - val_loss: 9.3069 - val_accuracy: 0.0701 Epoch 2/100 1024/1024 [==============================] - 77s 75ms/step - loss: 10.5672 - accuracy: 0.0642 - val_loss: 10.8782 - val_accuracy: 0.0718 Epoch 3/100 1024/1024 [==============================] - 80s 78ms/step - loss: 10.6035 - accuracy: 0.0639 - val_loss: 10.8998 - val_accuracy: 0.0658 Epoch 4/100 1024/1024 [==============================] - 78s 76ms/step - loss: 10.4597 - accuracy: 0.0658 - val_loss: 9.5053 - val_accuracy: 0.0581 Epoch 5/100 1024/1024 [==============================] - 77s 75ms/step - loss: 10.1673 - accuracy: 0.0596 - val_loss: 12.2643 - val_accuracy: 0.0620 Epoch 6/100 1024/1024 [==============================] - 81s 79ms/step - loss: 10.1886 - accuracy: 0.0628 - val_loss: 9.2048 - val_accuracy: 0.0641 Epoch 7/100 1024/1024 [==============================] - 78s 76ms/step - loss: 10.2992 - accuracy: 0.0630 - val_loss: 10.0681 - val_accuracy: 0.0658 Epoch 8/100 1024/1024 [==============================] - 65s 63ms/step - loss: 10.2812 - accuracy: 0.0665 - val_loss: 12.2382 - val_accuracy: 0.0645 Epoch 9/100 1024/1024 [==============================] - 76s 74ms/step - loss: 11.4436 - accuracy: 0.0637 - val_loss: 9.5845 - val_accuracy: 0.0697 Epoch 10/100 1024/1024 [==============================] - 55s 54ms/step - loss: 10.2822 - accuracy: 0.0664 - val_loss: 9.9871 - val_accuracy: 0.0632 Epoch 11/100 1024/1024 [==============================] - 56s 55ms/step - loss: 10.9518 - accuracy: 0.0647 - val_loss: 12.8353 - val_accuracy: 0.0603 Epoch 12/100 1024/1024 [==============================] - 57s 55ms/step - loss: 10.7480 - accuracy: 0.0646 - val_loss: 10.8068 - val_accuracy: 0.0607 Epoch 13/100 1024/1024 [==============================] - 56s 54ms/step - loss: 10.3040 - accuracy: 0.0618 - val_loss: 11.6900 - val_accuracy: 0.0628 Epoch 14/100 1024/1024 [==============================] - 54s 52ms/step - loss: 10.5912 - accuracy: 0.0630 - val_loss: 14.3563 - val_accuracy: 0.0778 Epoch 15/100 1024/1024 [==============================] - 53s 52ms/step - loss: 10.7772 - accuracy: 0.0635 - val_loss: 11.0138 - val_accuracy: 0.0641 Epoch 16/100 1024/1024 [==============================] - 53s 52ms/step - loss: 10.1329 - accuracy: 0.0651 - val_loss: 11.0438 - val_accuracy: 0.0632 Epoch 17/100 1024/1024 [==============================] - 54s 52ms/step - loss: 10.4157 - accuracy: 0.0617 - val_loss: 11.4240 - val_accuracy: 0.0662 Epoch 18/100 1024/1024 [==============================] - 57s 55ms/step - loss: 10.4042 - accuracy: 0.0635 - val_loss: 11.6729 - val_accuracy: 0.0624 train(epochs=100) 我上述代码运行输出一共16个类正确率一直这么低基本没变化感觉就是没用上这个模型的感觉不是微调的问题我的目的是图像分类出16个种类帮我检查这个代码帮我找找原因

### TensorFlow MobileNetV2 图像分类 16 类准确率低的原因分析针对使用 MobileNetV2 进行 16 类图像分类时准确率较低的情况，可能涉及多个方面的因素。以下是对问题的深入剖析： #### 1. **模型初始化与权重加载** 确保在创建 MobileNetV2 模型实例时正确指定了预训练权重。如果未指定 `weights='imagenet'` 或者自定义了顶层结构而未重新训练，则可能导致模型表现不佳[^2]。例如： ```python base_model = tf.keras.applications.MobileNetV2( weights='imagenet', include_top=False, input_shape=(224, 224, 3) ) ``` #### 2. **数据预处理的一致性** MobileNetV2 对输入图像有特定的要求，包括尺寸和像素值范围。如果不遵循这些要求，会影响模型的表现。应使用 `tf.keras.applications.mobilenet_v2.preprocess_input` 方法对输入数据进行标准化处理[^4]。 ```python preprocessed_image = tf.keras.applications.mobilenet_v2.preprocess_input(raw_image) ``` #### 3. **迁移学习策略** 当目标类别数量较少（如本例中的 16 类）时，通常需要冻结基础模型的层并仅微调顶部新增的全连接层。然而，若完全冻结所有层或过度解冻某些层，均可能导致欠拟合或过拟合现象发生[^3]。可以尝试逐步解冻更多层来进行更深层次的学习。 #### 4. **批量归一化的影响** 由于 Batch Normalization 层的存在，在预测阶段必须保证其统计特性与训练阶段保持一致。这可以通过设置 `training=False` 来实现[^2]。 ```python predictions = model.predict(test_images, training=False) ``` #### 5. **优化器与损失函数的选择** 选择合适的优化算法以及匹配的目标函数至关重要。对于多分类任务，交叉熵损失通常是首选方案之一。同时注意调整初始学习速率以适应具体应用场景下的收敛速度需求[^3]。 #### 6. **数据集质量及分布情况** 最后还需审视所用数据集中是否存在标签噪声、样本不平衡等问题。这些问题都可能是造成最终评测指标偏低的重要原因之一[^1]。 --- ### 提供一段修正后的代码示例下面给出了一段经过改进后可用于解决上述提到的一些常见问题的代码片段： ```python import tensorflow as tf from tensorflow.keras.layers import Dense, GlobalAveragePooling2D from tensorflow.keras.models import Model from tensorflow.keras.optimizers import Adam from sklearn.utils.class_weight import compute_class_weight import numpy as np # 构建基本模型 base_model = tf.keras.applications.MobileNetV2( weights='imagenet', include_top=False, input_shape=(224, 224, 3) ) x = base_model.output x = GlobalAveragePooling2D()(x) x = Dense(1024, activation='relu')(x) # 添加新的隐藏层 predictions = Dense(16, activation='softmax')(x) model = Model(inputs=base_model.input, outputs=predictions) # 冻结基底模型的所有层 for layer in base_model.layers: layer.trainable = False # 编译模型 optimizer = Adam(lr=0.0001) loss_fn = 'categorical_crossentropy' metrics_list = ['accuracy'] model.compile(optimizer=optimizer, loss=loss_fn, metrics=metrics_list) # 训练前准备 class_weights = compute_class_weight(class_weight='balanced', classes=np.unique(y_train), y=y_train.flatten()) weight_dict = dict(zip(np.unique(y_train), class_weights)) history = model.fit(train_generator, epochs=EPOCHS, validation_data=val_generator, class_weight=weight_dict) # 微调过程 for layer in model.layers[:int(len(model.layers)*0.9)]: layer.trainable = True optimizer_fine_tune = Adam(lr=0.00001) model.compile(optimizer=optimizer_fine_tune , loss=loss_fn, metrics=metrics_list) fine_tuned_history = model.fit(train_generator, epochs=FINE_TUNE_EPOCHS, initial_epoch=history.epoch[-1], validation_data=val_generator, class_weight=weight_dict) ```

阅读全文

相关推荐

tensorflow入门:tfrecord 和tf.data.TFRecordDataset的使用

intra.zip_intra_intra matlab_intra.zhengshang.cn_intra.zhenshang

VGG19_with_tensorflow-master111.zip_QZI_mase111_neural network_t

image = tf.io.read_file(path)

tensorflow2图片分类训练集

【YOLOv8深度学习融合】：TensorFlow_Keras中的YOLOv8应用详解

【图像转换专家】：Python ImageFile库的进阶图像处理技巧

labelimg与TensorFlow模型集成实践指南

TensorFlow建模实践：图像分类与识别

【卷积神经网络】：TensorFlow实战演示从零开始

【imgaug与深度学习框架整合】：TensorFlow和PyTorch的最佳实践

【深度学习实战】：用TensorFlow构建高效模型的5大技巧

【MSE的Python实现】：Scikit-learn与TensorFlow的对比实战

tensorflow,cv

标签字段名称调整：将 file_name 和 label 改为 文件名 和 标签，以匹配 JSON 文件中的实际字段名称。按照以上要求修改代码，并将修改后代码列出

tensorflow2.6图像处理

tensorflow 风格迁移 预训练

大家在看

粒子群算法matlab编写代码

PT650D称重显示器在配料秤中的应用

超实用zimo21取字模软件.7z

ipmitool v1.8.18 for windows

matlab开发-高斯系数模型中DoLoanPortfolio的累积分布函数

最新推荐

课程设计-jsp734停车场收费系统ssh-qr-修改.zip

谭浩强C语言电子教案第三版权威教程下载

【性能测试基准】：为RK3588选择合适的NVMe性能测试工具指南

centos 修改密码失败ERROR 1820 (HY000): You must reset your password using ALTER USER statement before executing this statement.

50万吨原油常压塔设计与改造分析

【固态硬盘寿命延长】：RK3588平台NVMe维护技巧大公开

M10050-KB

用友860数据字典详细解读与应用指南

【故障恢复策略】：RK3588与NVMe固态硬盘的容灾方案指南

idea运行keycloak18.0.0

标签字段名称调整：将 file_name 和 label 改为文件名和标签，以匹配 JSON 文件中的实际字段名称。按照以上要求修改代码，并将修改后代码列出

tensorflow 风格迁移预训练