Question

我正在尝试在一组新类上使用预训练的Inception V4模型。我的想法是削减最后几步并重新训练上一层输出的1536值的softmax。

slim = tf.contrib.slim

from inception_v4 import inception_v4 as net
from inception_v4 import inception_v4_arg_scope as net_scope

from tensorflow.python.framework import graph_util
from tensorflow.python.platform import gfile

def main(argv):

    with tf.Session() as sess:
        with slim.arg_scope(net_scope()):
            input_units = tf.placeholder(tf.float32, (None, 299, 299, 3), 'inputs')
            net(input_units, num_classes=1001)
            last_node = tf.get_default_graph().as_graph_def().node[-1].name
            saver = tf.train.Saver()
            saver.restore(sess, argv[1])

            output_graph_def = graph_util.convert_variables_to_constants(
                sess,
                tf.get_default_graph().as_graph_def(),
                [last_node]
            )
            with gfile.FastGFile(argv[2], 'w') as f:
                f.write(output_graph_def.SerializeToString())


if __name__=='__main__':
    main(sys.argv)

这将创建.pb图形文件并冻结预训练值。

之后我重新加载图表进行训练：

def add_final_training_ops(class_count, bottleneck_tensor):
    with tf.name_scope('input'):
        bottleneck_input = tf.placeholder_with_default(
            bottleneck_tensor,
            shape=[None, BOTTLENECK_TENSOR_SIZE],
            name='BottleneckInputPlaceholder'
        )

        ground_truth_input = tf.placeholder(tf.float32,
                                    [None, class_count],
                                    name='GroundTruthInput')

        layer_name = 'final_training_ops'
        with tf.name_scope(layer_name):
            with tf.name_scope('weights'):
               layer_weights = tf.Variable(tf.truncated_normal([BOTTLENECK_TENSOR_SIZE, class_count], stddev=0.001), name='final_weights')
                variable_summaries(layer_weights)
            with tf.name_scope('biases'):
                layer_biases = tf.Variable(tf.zeros([class_count]), name='final_biases')
                variable_summaries(layer_biases)
            with tf.name_scope('Wx_plus_b'):
                logits = tf.matmul(bottleneck_input, layer_weights) + layer_biases
                tf.summary.histogram('pre_activations', logits)

        final_tensor = tf.nn.softmax(logits, name='final_result')
        tf.summary.histogram('activations', final_tensor)

        with tf.name_scope('cross_entropy'):
            cross_entropy = tf.nn.softmax_cross_entropy_with_logits(
                labels=ground_truth_input, logits=logits)
            with tf.name_scope('total'):
                cross_entropy_mean = tf.reduce_mean(cross_entropy)
                tf.summary.scalar('cross_entropy', cross_entropy_mean)

        with tf.name_scope('train'):
            train_step = tf.train.GradientDescentOptimizer(0.01).minimize(
            cross_entropy_mean)

        return (train_step, cross_entropy_mean, bottleneck_input, 
            ground_truth_input,
            final_tensor)

代码是自定义的混合，从转移学习中获取InceptionV3教程。

培训将通过，两个相对不同的课程的准确率将是100％。

当我测试结果图时，我的准确度大约为50％，因此几乎是随机的。

使用Inception v4的转移学习进行重新训练的模型具有不稳定的表现

0 个答案: