Question

我想使用张量流（使用tensorlayer）制作图像分类器，我遇到了几个问题。其中之一就是我无法通过网络中的丢失层获得调试使用的张量值。

网络看起来像这样：

sess = tf.InteractiveSession()

x = tf.placeholder(tf.float32, shape=[None, 100, 100, 3], name='x')
y_ = tf.placeholder(tf.float32, shape=[None, ], name='y_')

w_init = tf.truncated_normal_initializer(stddev=0.01)
b_init = tf.constant_initializer(value=0.0)
inputs = tl.layers.InputLayer(x, name='inputs')
conv1 = tl.layers.Conv2d(inputs, 64, (5, 5), act=tf.nn.relu, padding='SAME', W_init=w_init, b_init=b_init, name='conv1')
    #(?, 96, 96, 64)
pool1 = tl.layers.MaxPool2d(conv1, (2, 2), padding='SAME', name='pool1')
    #(?, 48, 48, 64)
conv2 = tl.layers.Conv2d(pool1, 128, (5, 5), act=tf.nn.relu, padding='SAME', W_init=w_init, b_init=b_init, name='conv2')
    #(?, 44, 44, 64)
pool2 = tl.layers.MaxPool2d(conv2, (2, 2), padding='SAME', name='pool2')
    #(?, 22, 22, 128)
conv3 = tl.layers.Conv2d(pool2, 256, (5, 5), act=tf.nn.relu, padding='SAME', W_init=w_init, b_init=b_init, name='conv3')
    #(?, 18, 18, 256)
pool3 = tl.layers.MaxPool2d(conv3, (2, 2), padding='SAME', name='pool3')
    #(?, 9, 9, 256)
flatten = tl.layers.FlattenLayer(pool3, name='flatten_layer')
    #(?, 20736)
drop1 = tl.layers.DropoutLayer(flatten, keep=0.8, name='drop1')
    #(?, 20736)
relu2 = tl.layers.DenseLayer(drop1, n_units=512, act = tf.nn.relu, name='relu2')
    #(?, 512)
drop3 = tl.layers.DropoutLayer(relu2, keep=0.5, name='drop3')
    #(?, 512)
network = tl.layers.DenseLayer(drop3, n_units=1, act = tf.identity, name='output_layer')
    #(?, 2)

y = network.outputs
correct_prediction = tf.equal(tf.reduce_sum(y, 1), y_)

我想找到correct_prediction是什么，所以我运行会话，并将输入提供给网络。

print(sess.run(correct_prediction, feed_dict={
    x: np.asarray(X_train[0:10],dtype=np.float32), 
    y_: np.asarray(Y_train[0:10],dtype=np.float32), 
    drop1:0.8, 
    drop3:0.5
}))

但在那之后我得到了错误：

InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'Placeholder' with dtype float

通过跟踪到堆栈，我发现导致问题的是丢失层：

......
Caused by op 'Placeholder', defined at:
  File "run_linear_output.py", line 33, in <module>
    drop1 = tl.layers.DropoutLayer(flatten, keep=0.8, name='drop1')
  File "/home/neoclub/wanghong/env/lib/python3.5/site-packages/tensorlayer/layers.py", line 963, in __init__
    set_keep[name] = tf.placeholder(tf.float32)
  File "/home/neoclub/wanghong/env/lib/python3.5/site-packages/tensorflow/python/ops/array_ops.py", line 1507, in placeholder
    name=name)
......

从网络中删除这些图层后，我可以成功获得结果。

......
[False False False False False False False False False False]
......

显然这是错的（编辑：因为我忘了在这里舍入输出），但这是另一个问题。

之后，我注意到错误消息中有set_keep[name] = tf.placeholder(tf.float32)，所以我改变了我的Feed

print(sess.run(correct_prediction, feed_dict={
    x: np.asarray(X_train[0:10],dtype=np.float32), 
    y_: np.asarray(Y_train[0:10],dtype=np.float32), 
    drop1:0.8, 
    drop3:0.5
}))

但是我又犯了一个错误:(

Traceback (most recent call last):
  File "run_linear_output.py", line 79, in <module>
    print(sess.run(correct_prediction, feed_dict={x: np.asarray(X_train[0:10],dtype=np.float32), y_: np.asarray(Y_train[0:10],dtype=np.float32), drop1:0.8, drop3:0.5}))
  File "/home/neoclub/wanghong/env/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 778, in run
    run_metadata_ptr)
  File "/home/neoclub/wanghong/env/lib/python3.5/site-packages/tensorflow/python/client/session.py", line 933, in _run
    + e.args[0])
TypeError: Cannot interpret feed_dict key as Tensor: Can not convert a DropoutLayer into a Tensor.

我在这里完全糊涂了。如何正确使用dropout图层？我应该提供什么来使会话可以运行？

Answer 1

Feed通常通过占位符用于将数据放入图表。在您的Feed dict中，您尝试将drop1输入值0.8但drop1 =您的dropout图层。我想你想要做的就是创建两个占位符，这将是你的辍学率。

类似的东西：

drop_keep_1=tf.placeholder(tf.float32,None,name='drop_keep_1')
drop1=tl.layers.DropoutLayer(flatten, keep=drop_keep_1, name='drop1')

我注意到你没有使用Tensorflow Layers API，而是使用tl.layers（我从未听说过），但我想问题仍然是一样的。

Answer 2

您有两种方式在TensorLayer中使用 DropoutLayer ：

1）使用由TensorLayer创建的内部保留概率占位符，请参阅tutorial_mlp_dropout1.py

2）我们不是使用占位符来控制保持概率，而是构建两个用于训练和测试的图表，请参阅tutorial_mlp_dropout2.py

Answer 3

network = tl.layers.DropoutLayer(network, keep = 0.5, is_fix = True, name = 'drop1')

添加is_fix = True是重点

当存在丢失层时，Tensorflow无法提供（运行会话）

3 个答案: