文件主要分为Train和Model两部分，可以分开两个文件，我在一起。
第三方工具：
tensorflow_gpu == 2.6.0
pandas == 1.3.2
numpy == 1.19.5
附件：Fer2013数据集
模型文件
有两个模型很简单CNN层，两个FC层
# ----------------------模型---------------------------------------- class CNN(tf.keras.Model): def __init__(self, num_class, keep_prob): super(CNN, self).__init__() self.conv1 = tf.keras.layers.Conv2D(filters=32, kernel_size=5, strides=1, use_bias=True, padding='same') self.conv1_act = tf.keras.layers.Activation('relu') self.conv1_pool = tf.keras.layers.MaxPool2D(pool_size=2, strides=2, padding='same') self.conv2 = tf.keras.layers.Conv2D(64, 5, strides=2, use_bias=True, padding='same') self.conv2_act = tf.keras.layers.Activation('relu') self.conv2_pool = tf.keras.layers.MaxPool2D(pool_size=2, strides=2, padding='same') self.flat = tf.keras.layers.Flatten() self.dense1 = tf.keras.layers.Dense(1024, activation='relu', use_bias=True) self.dense1_act = tf.keras.layers.Activation('relu') self.drop = tf.keras.layers.Dropout(rate=keep_prob) self.dense2 = tf.keras.layers.Dense(num_class, use_bias=True) def call(self, inputs): x = self.conv1(inputs) x = self.conv1_act(x) x = self.conv1_pool(x) x = self.conv2(x) x = self.conv2_act(x) x = self.conv2_pool(x) x = self.flat(x) x = self.dense1(x) x = self.dense1_act(x) x = self.drop(x) x = self.dense2(x) return x训练文件
由于fer2013文件是csv文件、思路：
1.使用pandas读取csv文件
2.分为特征和标签
3.将pixel转换为所需类型
4.转换成tensorflow的Dateset
5.提取部分数据集
# ---------------------------参数--------------------- model_path = r'./checkpoint/emotion_analysis.ckpt' # path where to save the trained model penalty_parameter = 0.02 # the SVM C penalty parameter log_path = r'/logs/' # path where to save the TensorBoard logs num_classes = 7 dropout_rate = 0.5 # dropout batch_size = 128 epoch = 100 lr = 3e-4 weight_decay = 1e-4 data_path = r'../Datasets/archive/fer2013.csv' AUTOTUNE = tf.data.experimental.AUTOTUNE np.set_printoptions(precision=3, suppress=True) # 读取csv文件 df = pd.read_csv(filepath_or_buffer=data_path, usecols=["emotion", "pixels"], dtype={"pixels": str}) fer_pixels = df.copy() # 分为特征和标签 fer_label = fer_pixels.pop('emotion') fer_pixels = np.asarray(fer_pixels) # 将特征转换为模型所需的类型 fer_train = [] for i in range(len(fer_label)): pixels_new = np.asarray([float(p) for p in fer_pixels[i][0].split()]).reshape([48,48,1]) fer_train.append(pixels_new) fer_train = np.asarray(fer_train) fer_label = np.asarray(fer_label) # 转换为tf.Dateset类型 dataset = tf.data.Dataset.from_tensor_slices((fer_train, fer_label)) # 数据集验证集测试集的拆分 train_dataset = dataset.take(1000) # 为了测试，这里只使用了1000张图片。 test_dataset = dataset.skip(32297) # 打乱 train_dataset = (train_dataset.cache().shuffle(5 * batch_size).batch(batch_size).prefetch(AUTOTUNE)) # 训练 strategy = tf.distribute.MirroredStrategy() with strategy.scope(): model = CNN(num_class=num_classes, keep_prob=dropout_rate) model.compile(loss=tf.keras.losses.Hinge(), optimizer=tf.keras.optimizers.Adam(learning_rate=1e-3), metrics=['accuracy']) # 断点续训 if os.path.exists(model_path '/saved_model.pb'): print('-加载模型-') model = tf.keras.models.load_model(model_path)#加载model cp_callback = tf.keras.callbacks.ModelCheckpoint(filepath=model_path, save_weights_only=False, monitor='val_accuracy', model='max', save_best_only=True) # 训练 history = model.fit(x=train_dataset, epochs=epoch,callbacks=[cp_callback]) model.summary()