tensorflow训练时的一些注意事项

1，使用batch norm层后，计算损失时，注意添加相应操作

def conv_bn_relu(inputs, num_outputs, phase, kernel_size, stride=1, padding='SAME', scope=None,weight_decay=0):
    with tf.variable_scope(scope):
        bottom = layers.conv2d(inputs=inputs, num_outputs=num_outputs, kernel_size=kernel_size, activation_fn=None, 
                               weights_regularizer=layers.l2_regularizer(weight_decay),reuse=tf.get_variable_scope().reuse,
                               stride=stride, padding=padding, normalizer_fn=None, scope='conv')
        bottom = layers.batch_norm(bottom, center=True, scale=True, is_training=phase, scope='bn')
        return tf.nn.relu(bottom, 'relu')

最后计算损失时：

update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
    train_step = tf.train.GradientDescentOptimizer(0.01).minimize(total_loss)

其中，tf.control_dependencies(update_ops)表示with段中的操作是在update_ops操作执行之后 再执行的

2，稀疏标签转为dense_label (one_hot标签)

# 以下是把稀疏标签ys转为向量形式shape [batch_size, NUM_CLASSES].注意，需要把上面ys的int64替换为int32
    batch_size = train_batch_size #假设是100
    sparse_labels = tf.reshape(ys, [batch_size, 1]) 
    print(np.shape(sparse_labels)) #[100*1]
    indices = tf.reshape(tf.range(batch_size), [batch_size, 1]) 
    print(np.shape(indices)) #[100*1]
    concated = tf.concat( [indices, sparse_labels], 1 )
    print(np.shape(concated)) #[100*2]
    dense_labels = tf.sparse_to_dense(concated,
                                [batch_size, NUM_CLASSES],
                                1.0, 0.0) #[100*7]
    # 转换完成
    #对应的计算loss：
    loss = tf.nn.softmax_cross_entropy_with_logits( logits=logits,labels= dense_labels )
    loss = tf.reduce_mean(loss) #avg loss
    #计算accuracy：
    correct_prediction = tf.equal( tf.argmax( prediction,1 ), tf.argmax( dense_labels,1 ) ) 
    accuracy = tf.reduce_mean( tf.cast(correct_prediction,tf.float32) )

3，待续