1. 简单介绍
本文的应用场景多输入问题，采用的数据集为有，'black_jeans','black_shoes','blue_dress','blue_jeans','blue_shirt','red_dress','red_shirt'七个类别，我们将根据颜色和衣服类型进行两类输出。
2. 加载相关数据包
import tensorflow as tf print('Tensorflow version: {}'.format(tf.__version__)) from tensorflow import keras import matplotlib.pyplot as plt %matplotlib inline import numpy as np import pathlib import os import random
Tensorflow version: 2.0.0 • 1
2.1 图片的路径的配置
#配置数据集路径 path_root = os.path.realpath(".") data_dir = pathlib.Path(path_root) #数量构成 image_count = len(list(data_dir.glob('dataset/*/*.jpg'))) label_names = sorted(item.name for item in data_dir.glob("dataset/*")) #图片路径 all_image_path = list(list(data_dir.glob('dataset/*/*.jpg'))) #将路径打乱 all_image_path = [str(path) for path in all_image_path] random.shuffle(all_image_path) color_label_names = set(name.split('_')[0] for name in label_names) item_label_names = set(name.split('_')[1] for name in label_names) color_lable_to_index = dict((name,index) for index,name in enumerate(color_label_names)) item_lable_to_index = dict((name,index) for index,name in enumerate(item_label_names)) color_label = [color_lable_to_index[pathlib.Path(path).parent.name.split("_")[0]] for path in all_image_path] item_label = [item_lable_to_index[pathlib.Path(path).parent.name.split("_")[1]] for path in all_image_path]
2.2 读取图片
def preprocess_image(image): image = tf.image.decode_jpeg(image, channels=3) image = tf.image.resize(image, [224, 224]) image = tf.cast(image, tf.float32) image /= 255.0 # normalize to [0,1] range image = 2*image-1 return image #加载图片 def load_and_preprocess_image(path): image = tf.io.read_file(path) return preprocess_image(image) #举例 image_path = all_image_path[11] color = color_label[11] item = item_label[11] plt.imshow((load_and_preprocess_image(image_path)+1)/2) plt.grid(False) ##plt.xlabel(caption_image(image_path)) plt.title(str(list(color_lable_to_index.keys())[color])+"_"+str(list(item_lable_to_index.keys())[item]).title()) plt.axis("off") plt.show()
3. 图片预处理
在这一部分我们采用from_tensor_slices的方法对图片数据集进行构建，对比tf1.x版本采用队列形式读取数据，这一种方法比较简单切易于理解。并构建（图片，标签）对数据集。
#%%构建一个tf.data.Dataset #一个图片数据集构建 tf.data.Dataset 最简单的方法就是使用 from_tensor_slices 方法。 #将字符串数组切片，得到一个字符串数据集： path_ds = tf.data.Dataset.from_tensor_slices(all_image_path) print(path_ds) #现在创建一个新的数据集，通过在路径数据集上映射 preprocess_image 来动态加载和格式化图片。 AUTOTUNE = tf.data.experimental.AUTOTUNE image_ds = path_ds.map(load_and_preprocess_image,num_parallel_calls=AUTOTUNE) lable_ds = tf.data.Dataset.from_tensor_slices((color_label, item_label)) for color,item in lable_ds.take(5): print(str(list(color_lable_to_index.keys())[color])+"_"+str(list(item_lable_to_index.keys())[item])) #%%构建一个（图片，标签）对数据集 #因为这些数据集顺序相同，可以将他们打包起来 image_label_ds = tf.data.Dataset.zip((image_ds,lable_ds)) print(image_label_ds)
<TensorSliceDataset shapes: (), types: tf.string> blue_dress blue_dress blue_shirt red_dress blue_shirt <ZipDataset shapes: ((224, 224, 3), ((), ())), types: (tf.float32, (tf.int32, tf.int32))>
4. 训练阶段
4.1 设置验证集与数据集
#%%设置训练数据和验证集数据的大小 test_count = int(image_count*0.2) train_count = image_count - test_count print(test_count,train_count) #跳过test_count个 train_dataset = image_label_ds.skip(test_count) test_dataset = image_label_ds.take(test_count) batch_size = 16 # 设置一个和数据集大小一致的 shuffle buffer size（随机缓冲区大小）以保证数据被充分打乱。 train_ds = train_dataset.shuffle(buffer_size=train_count).repeat().batch(batch_size) test_ds = test_dataset.batch(batch_size)
419 1677 • 1
4.2 构建模型并训练
mobile_net = tf.keras.applications.MobileNetV2(input_shape=(224, 224, 3), include_top=False, weights='imagenet') mobile_net.trianable = False inputs = tf.keras.Input(shape=(224, 224, 3)) x = mobile_net(inputs) x = tf.keras.layers.GlobalAveragePooling2D()(x) x1 = tf.keras.layers.Dense(1024, activation='relu')(x) out_color = tf.keras.layers.Dense(len(color_label_names), activation='softmax', name='out_color')(x1) x2 = tf.keras.layers.Dense(1024, activation='relu')(x) out_item = tf.keras.layers.Dense(len(item_label_names), activation='softmax', name='out_item')(x2) model = tf.keras.Model(inputs=inputs, outputs=[out_color, out_item]) model.summary()
Model: "model" __________________________________________________________________________________________________ Layer (type) Output Shape Param # Connected to ================================================================================================== input_2 (InputLayer) [(None, 224, 224, 3) 0 __________________________________________________________________________________________________ mobilenetv2_1.00_224 (Model) (None, 7, 7, 1280) 2257984 input_2[0][0] __________________________________________________________________________________________________ global_average_pooling2d (Globa (None, 1280) 0 mobilenetv2_1.00_224[1][0] __________________________________________________________________________________________________ dense (Dense) (None, 1024) 1311744 global_average_pooling2d[0][0] __________________________________________________________________________________________________ dense_1 (Dense) (None, 1024) 1311744 global_average_pooling2d[0][0] __________________________________________________________________________________________________ out_color (Dense) (None, 3) 3075 dense[0][0] __________________________________________________________________________________________________ out_item (Dense) (None, 4) 4100 dense_1[0][0] ================================================================================================== Total params: 4,888,647 Trainable params: 4,854,535 Non-trainable params: 34,112 __________________________________________________________________________________________________
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), loss={'out_color':'sparse_categorical_crossentropy', 'out_item':'sparse_categorical_crossentropy'}, metrics=['acc'] ) steps_per_eooch = train_count//batch_size validation_steps = test_count//batch_size history = model.fit(train_ds,epochs=3,steps_per_epoch=steps_per_eooch,validation_data=test_ds,validation_steps=validation_steps)
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), loss={'out_color':'sparse_categorical_crossentropy', 'out_item':'sparse_categorical_crossentropy'}, metrics=['acc'] ) steps_per_eooch = train_count//batch_size validation_steps = test_count//batch_size history = model.fit(train_ds,epochs=3,steps_per_epoch=steps_per_eooch,validation_data=test_ds,validation_steps=validation_steps)
Train for 104 steps, validate for 26 steps Epoch 1/3 104/104 [==============================] - 103s 991ms/step - loss: 0.4240 - out_color_loss: 0.2223 - out_item_loss: 0.2017 - out_color_acc: 0.9123 - out_item_acc: 0.9399 - val_loss: 0.1177 - val_out_color_loss: 0.0892 - val_out_item_loss: 0.0284 - val_out_color_acc: 0.9688 - val_out_item_acc: 0.9880 Epoch 2/3 104/104 [==============================] - 86s 825ms/step - loss: 0.0716 - out_color_loss: 0.0404 - out_item_loss: 0.0312 - out_color_acc: 0.9838 - out_item_acc: 0.9916 - val_loss: 0.0842 - val_out_color_loss: 0.0412 - val_out_item_loss: 0.0430 - val_out_color_acc: 0.9856 - val_out_item_acc: 0.9880 Epoch 3/3 104/104 [==============================] - 87s 833ms/step - loss: 0.0416 - out_color_loss: 0.0170 - out_item_loss: 0.0246 - out_color_acc: 0.9952 - out_item_acc: 0.9922 - val_loss: 0.0546 - val_out_color_loss: 0.0309 - val_out_item_loss: 0.0238 - val_out_color_acc: 0.9904 - val_out_item_acc: 0.9904
5. 模型评估
model.evaluate(test_ds)
27/27 [==============================] - 9s 347ms/step - loss: 0.0526 - out_color_loss: 0.0297 - out_item_loss: 0.0229 - out_color_acc: 0.9905 - out_item_acc: 0.9905 [0.052628928653171494, 0.02974791, 0.022881018, 0.9904535, 0.9904535]
我们在网上找一个图片，检验其准确性
str(list(data_dir.glob('dataset/*.jpg'))[0]) my_image = load_and_preprocess_image(str(list(data_dir.glob('dataset/*.jpg'))[0])) my_image = tf.expand_dims(my_image, 0) pred = model.predict(my_image) plt.grid(False) plt.imshow((tf.squeeze(my_image,axis=0)+1)/2) plt.title(str(list(color_lable_to_index.keys())[np.argmax(pred[0])])+"_"+str(list(item_lable_to_index.keys())[np.argmax(pred[1])]).title()) plt.axis("off") plt.show()
