使用Tensorflow Slim复制Tensorflow Hub模块输出

问题描述

我试图从基于Tensorflow Slim Tensorflow Hub模块的输出中重现使用Tensorflow Slim模块进行nofollow noreferrer>检查点.但是，我似乎无法获得预期的输出.例如，让我们加载所需的库，创建示例输入，并使用占位符填充数据:

I am trying to reproduce the output from a Tensorflow Hub module that is based on a Tensorflow Slim checkpoint, using the Tensorflow Slim modules. However, I can't seem to get the expected output. For example, let us load the required libraries, create a sample input and the placeholder to feed the data:

import tensorflow_hub as hub
from tensorflow.contrib.slim import nets

images = np.random.rand(1,224,224,3).astype(np.float32)
inputs = tf.placeholder(shape=[None, 224, 224, 3], dtype=tf.float32)

加载TF Hub模块:

resnet_hub = hub.Module("https://tfhub.dev/google/imagenet/resnet_v2_152/feature_vector/3")
features_hub = resnet_hub(inputs, signature="image_feature_vector", as_dict=True)["resnet_v2_152/block4"]

现在，让我们对TF Slim做同样的事情，并创建一个将加载检查点的加载器:

Now, let's do the same with TF Slim and create a loader that will load the checkpoint:

with slim.arg_scope(nets.resnet_utils.resnet_arg_scope()):
    _, end_points = nets.resnet_v2.resnet_v2_152(image, is_training=False)
    features_slim = end_points["resnet_v2_152/block4"]
loader = tf.train.Saver(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope="resnet_v2_152"))

现在，一旦一切就绪，我们就可以测试输出是否相同:

Now, once we have everything in place we can test whether the outputs are the same:

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    loader.restore(sess, "resnet_v2_152_2017_04_14/resnet_v2_152.ckpt")
    slim_output = sess.run(features_slim, feed_dict={inputs: images})
    hub_output = sess.run(features_hub, feed_dict={inputs: images})
    np.testing.assert_array_equal(slim_output, hub_output)

但是，断言失败，因为两个输出不相同.我认为这是因为TF Hub对TF Slim实现缺少的输入使用了内部预处理.

However, the assertion fails because the two outputs are not the same. I assume that this is because TF Hub uses an internal preprocessing of the inputs that the TF Slim implementation lacks.

让我知道您的想法！

推荐答案

这些Hub模块将其输入从标准范围[0,1]扩展到相应的苗条检查点对其进行培训的预处理所期望的值(通常为[- 1，+ 1](用于初始样式"预处理).将它们传递给相同的输入可以解释很大的不同.甚至在进行线性重新缩放以解决该问题之后，达到复合数值误差的差异也不会令我感到惊讶(考虑到TF内部的许多自由度)，但是主要差异可能表明存在错误.

Those Hub modules scale their inputs from the canonical range [0,1] to what the respective slim checkpoint expects from the preprocessing it was trained with (typically [-1,+1] for "Inception-style" preprocessing). Passing them the same inputs can explain a large difference. Even after linear rescaling to fix that, a difference up to compounded numerical error wouldn't surprise me (given the many degrees of freedom inside TF), but major discrepancies might indicate a bug.

这篇关于使用Tensorflow Slim复制Tensorflow Hub模块输出的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持！