我正在尝试使用TensorFlow创建一个CNN来对Google关于CNN的教程中的图像进行分类。我创建了一个函数来加载图像数据集，并有一个函数来提取用于训练的批次。但即使我移动到下一批次，网络仍然总是在同一个批次上进行训练。数据集有10000张图像。我认为模型没有使用下一批次的原因是我在不到10次迭代中就达到了100%的准确率。这里是代码:
#Training Parameterslearning_rate = 0.001batch_size = 128epochs = 10MODE = 'TRAIN'# Function that loads the entire dataset of images (X) with their respective labels (Y). X and Y are two np.arraylen_X, X, Y = get_images( files_path=dataset_path, img_size_h=1000, img_size_w=48, mode='TRAIN', randomize=True)# Function that load the batch from X and YX_batch, Y_X_batch = next_batch( total=len_X, images=X, labels=Y, batch_size=batch_size, index=0)logits = cnn_model_fn(X_batch, MODE)prediction = tf.nn.softmax(logits)loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=Y_X_batch))optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)train_op = optimizer.minimize(loss)correct_predict = tf.equal(tf.argmax(prediction, 1), tf.argmax(Y_X_batch, 1))accuracy = tf.reduce_mean(tf.cast(correct_predict, tf.float32))init = tf.global_variables_initializer()best_acc=0with tf.Session() as sess: sess.run(init) saver = tf.train.Saver() if MODE == 'TRAIN': print("TRAINING MODE") for step in range(1,epochs+1): for i in range(0, int(len_X/batch_size)+1): if i > 0: X_batch, Y_X_batch = next_batch( total=len_X, images=X, labels=Y, batch_size=batch_size, index=i ) sess.run(train_op) los, acc= sess.run([loss, accuracy]) if acc >= best_acc: best_acc = acc writer = tf.summary.FileWriter(TensorBoard_path, sess.graph) elif MODE=='TEST': # TEST MODE #sess.close()这是神经网络模型的结构:
def cnn_model_fn(X, MODE): # INPUT LAYER input_layer = tf.reshape(X, [-1, 1000, 48, 1]) # CONVOLUTIONAL LAYER #1 conv1 = tf.layers.conv2d( inputs=input_layer, filters=4, kernel_size=[10, 10], strides=(2, 2), padding="valid", ) conv1_relu = tf.nn.relu(conv1) # POOLING LAYER #1 pool1 = tf.layers.max_pooling2d( inputs=conv1_relu, pool_size=[2, 2], strides=2 ) # CONVOLUTIONAL LAYER #2 conv2 = tf.layers.conv2d( inputs=pool1, filters=64, kernel_size=[5, 5], padding="same", ) conv2_relu = tf.nn.relu(conv2) # POOLING LAYER #2 pool2 = tf.layers.max_pooling2d( inputs=conv2_relu, pool_size=[2, 2], strides=2 ) x = tf.TensorShape.as_list(pool2.shape) pool2_flat = tf.reshape(pool2, [-1, x[1] * x[2] * x[3]]) # DENSE LAYER dense = tf.layers.dense( inputs=pool2_flat, units=1024, ) dense_relu = tf.nn.relu(dense) # AGGIUNGO L'OPERAZIONE DI DROPOUT dropout = tf.layers.dropout( inputs=dense_relu, rate=0.4, training=MODE == tf.estimator.ModeKeys.TRAIN ) # LOGIT LAYER logits = tf.layers.dense( inputs=dropout, units=2 ) return logits非常感谢您的时间。
回答：
您没有使用TensorFlow的占位符。请确保定义并在损失函数中使用X_batch_placeholder和Y_batch_placeholder，如下所示:
Y_batch_placeholder = tf.placeholder(tf.float32 ,[None, 2] )X_batch_placeholder = tf.placeholder(tf.float32 ,[None, 1000, 48])logits = cnn_model_fn(X_batch_placeholder, MODE)prediction = tf.nn.softmax(logits)loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=Y_batch_placeholder))optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)train_op = optimizer.minimize(loss)correct_predict = tf.equal(tf.argmax(prediction, 1), tf.argmax(Y_batch_placeholder, 1))accuracy = tf.reduce_mean(tf.cast(correct_predict, tf.float32))然后使用feed_dict运行您的会话。有些TensorFlow变量是占位符，您需要在运行会话时使用feed_dict来为它们提供相关信息:
_, los, acc = sess.run([train_op, loss, accuracy], feed_dict={X_batch_placeholder: X_batch, Y_batch_placeholder: Y_batch})请参考这里: Tensorflow: When should I use or not use `feed_dict`? 和这里: https://hanxiao.github.io/2017/07/07/Get-10x-Speedup-in-Tensorflow-Multi-Task-Learning-using-Python-Multiprocessing/
