如何将矢量重塑为TensorFlow的过滤器？-Java 学习之路

我想将一些由另一个网络训练的权重转移到TensorFlow，权重存储在一个向量中，如下所示：

[1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18]

通过使用numpy，我可以将它重塑为两个3乘3的过滤器，如下所示：

1 2 3     9  10 11
3 4 5     12 13 14
6 7 8     15 16 17

因此，我的过滤器的形状是 (1,2,3,3) . 但是，在TensorFlow中，过滤器的形状为 (3,3,2,1) ：

tf_weights = tf.Variable(tf.random_normal([3,3,2,1]))

在将tf_weights重塑为预期形状后，重量变得混乱，我无法获得预期的卷积结果 .

具体来说，当图像或滤镜的形状是[数字，通道，大小，大小]时，我写了一个卷积函数，它给出了正确的答案，但它太慢了：

def convol(images,weights,biases,stride):
    """
    Args:
      images:input images or features, 4-D tensor
      weights:weights, 4-D tensor
      biases:biases, 1-D tensor
      stride:stride, a float number
    Returns:
      conv_feature: convolved feature map
    """
    image_num = images.shape[0] #the number of input images or feature maps
    channel = images.shape[1] #channels of an image,images's shape should be like [n,c,h,w]
    weight_num = weights.shape[0] #number of weights, weights' shape should be like [n,c,size,size]
    ksize = weights.shape[2]
    h = images.shape[2]
    w = images.shape[3]
    out_h = (h+np.floor(ksize/2)*2-ksize)/2+1
    out_w = out_h

    conv_features = np.zeros([image_num,weight_num,out_h,out_w])
    for i in range(image_num):
        image = images[i,...,...,...]
        for j in range(weight_num):
            sum_convol_feature = np.zeros([out_h,out_w])
            for c in range(channel):
                #extract a single channel image
                channel_image = image[c,...,...]
                #pad the image
                padded_image = im_pad(channel_image,ksize/2)
                #transform this image to a vector
                im_col = im2col(padded_image,ksize,stride)

                weight = weights[j,c,...,...]
                weight_col = np.reshape(weight,[-1])
                mul = np.dot(im_col,weight_col)
                convol_feature = np.reshape(mul,[out_h,out_w])
                sum_convol_feature = sum_convol_feature + convol_feature
            conv_features[i,j,...,...] = sum_convol_feature + biases[j]
    return conv_features

相反，通过使用tensorflow的conv2d，如下所示：

img = np.zeros([1,3,224,224])
img = img - 1
img = np.rollaxis(img, 1, 4)

weight_array = googleNet.layers[1].weights
weight_array = np.reshape(weight_array,[64,3,7,7])

biases_array = googleNet.layers[1].biases

tf_weight = tf.Variable(weight_array)

tf_img = tf.Variable(img)
tf_img = tf.cast(tf_img,tf.float32)

tf_biases = tf.Variable(biases_array)

conv_feature = tf.nn.bias_add(tf.nn.conv2d(tf_img,tf_weight,strides=[1,2,2,1],padding='SAME'),tf_biases)
sess = tf.Session()
sess.run(tf.initialize_all_variables())
feautre = sess.run(conv_feature)

我得到的功能图是错误的 .

2 回答

不要使用 np.reshape . 它可能mess up the order of your values .

请改用np.rollaxis：

>>> a = np.array([1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18])
>>> a = a.reshape((1,2,3,3))
>>> a
array([[[[ 1,  2,  3],
         [ 4,  5,  6],
         [ 7,  8,  9]],

        [[10, 11, 12],
         [13, 14, 15],
         [16, 17, 18]]]])
>>> b = np.rollaxis(a, 1, 4)
>>> b.shape
(1, 3, 3, 2)
>>> b = np.rollaxis(b, 0, 4)
>>> b.shape
(3, 3, 2, 1)

请注意，尺寸为3的两个轴的顺序没有改变 . 如果我要标记它们，那么两个 rollaxis 操作会导致形状变为（1,2,31,32） - >（1,31,32,2） - >（31,32,2,1） . 你的最终数组如下：

>>> b
array([[[[ 1],
         [10]],

        [[ 2],
         [11]],

        [[ 3],
         [12]]],


       [[[ 4],
         [13]],

        [[ 5],
         [14]],

        [[ 6],
         [15]]],


       [[[ 7],
         [16]],

        [[ 8],
         [17]],

        [[ 9],
         [18]]]])

回复于 2024-04-30T15:29:30+08:00

样本张量操作

我不知道这是否有帮助 . 考虑Reshape，Gather，Dynamic_partition和Split操作，并根据您的需要进行调整 . 下面是这些操作的说明，可以适应您的情况 . 我从我的git repo中复制了这个 . 我相信如果你在ipython中运行这些例子，你可以弄清楚你真正想要的是什么，并获得更好的洞察力 .

重塑，收集，动态分区和拆分

收集操作（tf.gather（））

生成数组并测试收集操作 . 请注意这种快速原型制作方法：

我们在Numpy中生成一个数组并测试其上的张量流的操作 .

使用：根据指数从params收集切片 .

indices必须是任何维度的整数张量（通常为0-D或1-D） . 最好通过一个例子来说明：

array = np.array([[1,2,3],[4,9,6],[2,3,4],[7,8,0]])

array.shape


(4, 3)

In [27]:

gather_output0  = tf.gather(array,1)
gather_output01  = tf.gather(array,2)
gather_output02  = tf.gather(array,3)

gather_output11  = tf.gather(array,[1,2])
gather_output12  = tf.gather(array,[1,3])
gather_output13  = tf.gather(array,[3,2])




gather_output  = tf.gather(array,[1,0,2])
gather_output1  = tf.gather(array,[1,1,2])
gather_output2  = tf.gather(array,[1,2,1])

In [28]:

with tf.Session() as sess:
    print (gather_output0.eval());print("\n")
    print (gather_output01.eval());print("\n")
    print (gather_output02.eval());print("\n")  
    print (gather_output11.eval());print("\n")
    print (gather_output12.eval());print("\n")
    print (gather_output13.eval());print("\n")

    print (gather_output.eval());print("\n")
    print (gather_output1.eval());print("\n")
    print (gather_output2.eval());print("\n")
    #print (gather_output2.eval());print("\n")

[4 9 6]


[2 3 4]


[7 8 0]


[[4 9 6]
 [2 3 4]]


[[4 9 6]
 [7 8 0]]


[[7 8 0]
 [2 3 4]]


[[4 9 6]
 [1 2 3]
 [2 3 4]]


[[4 9 6]
 [4 9 6]
 [2 3 4]]


[[4 9 6]
 [2 3 4]
 [4 9 6]]

看一下这个简单的例子：

初始化简单数组
测试收集操作

在[11]中：

array_simple = np.array([1,2,3])

In [15]:

print "shape of simple array is: ", array_simple.shape

shape of simple array is:  (3,)

In [57]:

gather1  = tf.gather(array1,[0])
gather01 = tf.gather(array1,[1])
gather02 = tf.gather(array1,[2])

gather2 = tf.gather(array1,[1,2])
gather3 = tf.gather(array1,[0,1])

with tf.Session() as sess:
    print (gather1.eval());print("\n")
    print (gather01.eval());print("\n")
    print (gather02.eval());print("\n")
    print (gather2.eval());print("\n")
    print (gather3.eval());print("\n")

[1]


[2]


[3]


[2 3]


[1 2]


tf.reshape( )

Note:

*  Use the same array that was initiated
*  Do reshape using tf.reshape( )

In [64]:

array.shape # Confirm array shape

Out[64]:

(4, 3)

In [74]:

print ("This is the array\n" ,array) # see the output and compare with the initial array,

This is the array
 [[1 2 3]
 [4 9 6]
 [2 3 4]
 [7 8 0]]

In [84]:

reshape_ops= tf.reshape(array,[-1,4]) # Note the parameters in reshpe
reshape_ops1= tf.reshape(array,[-1,3]) # Note the parameters in reshpe
reshape_ops2= tf.reshape(array,[-1,6]) # Note the parameters in reshpe

reshape_ops_back1= tf.reshape(array,[6,-1]) # Note the parameters in reshpe
reshape_ops_back2= tf.reshape(array,[3,-1]) # Note the parameters in reshpe
reshape_ops_back3= tf.reshape(array,[4,-1]) # Note the parameters in reshpe

In [86]:

with tf.Session() as sess:
    print(reshape_ops.eval());print("\n")
    print(reshape_ops1.eval());print("\n")
    print(reshape_ops2.eval());print("\n")
    print ("Output when we reverse the parameters:");print("\n")
    print(reshape_ops_back1.eval());print("\n")
    print(reshape_ops_back2.eval());print("\n")
    print(reshape_ops_back3.eval());print("\n")

[[1 2 3 4]
 [9 6 2 3]
 [4 7 8 0]]


[[1 2 3]
 [4 9 6]
 [2 3 4]
 [7 8 0]]


[[1 2 3 4 9 6]
 [2 3 4 7 8 0]]


Output when we reverse the parameters:


[[1 2]
 [3 4]
 [9 6]
 [2 3]
 [4 7]
 [8 0]]


[[1 2 3 4]
 [9 6 2 3]
 [4 7 8 0]]


[[1 2 3]
 [4 9 6]
 [2 3 4]
 [7 8 0]]

注意：输入大小和输出大小必须相同 . ---否则会出错 . 检查这个的简单方法是通过简单的乘法确保输入可以分成重塑参数 .

Dynamic_cell_partitions

This is declared as :

tf.dynamic_partition (array, partitions, num_partitions, name=None)

Note:

* we decalare number_partitions --- number of partitions
* Use our array initialised earlier
* We declare the partition as [0 1 0 1] . This signifies the partitions we want 0's fall to one partition and 1 the other partitions given that we have two num_partitions=2.

* The output is a list

In [96]:

    print ("This is the array\n" ,array) # This is output array

    This is the array
     [[1 2 3]
     [4 9 6]
     [2 3 4]
     [7 8 0]]

    We show how to make two and three partitions below
    In [123]:

    num_partitions = 2
    num_partitions1 = 3

    partitions = [0, 0, 1, 1]
    partitions1 = [0 ,1 ,1, 2 ]

    In [119]:

    dynamic_ops =tf.dynamic_partition(array, partitions, num_partitions, name=None) # 2 partitions
    dynamic_ops1 =tf.dynamic_partition(array, partitions1, num_partitions1, name=None) # 3 partitions

    In [125]:

    with tf.Session() as sess:
        run = sess.run(dynamic_ops)
        run1 = sess.run(dynamic_ops1)
        print("Output for 2 partitions: ")
        print (run[0]);print("\n")
        print(run[1]) ;print("\n")# Compare result with initial array. Out is list
        print("Output for three partitions: ")

        print (run1[0]);print("\n")
        print (run1[1]);print("\n")
        print (run1[2]);print("\n")

    Output for 2 partitions: 
    [[1 2 3]
     [4 9 6]]


    [[2 3 4]
     [7 8 0]]


    Output for three partitions: 
    [[1 2 3]]


    [[4 9 6]
     [2 3 4]]


    [[7 8 0]]

tf.split（）

确保使用最新的tensorflow版本 . 否则在旧版本中，此实现将给出错误

这在文档中指定如下：

tf.split（value，num_or_size_splits，axis = 0，num = None，name ='split'） .

它将张量分割为子张量 . 最好通过一个例子来说明：

* we define (5,30) aray in numpy
* we split the array along axis 1
* We  specify the number of splits as 1-Dimen Tensor along axis 1. So we have 3 splits.

Specify an array

    Create a (5 by 30) numpy array. The syntax using numpy is shown below
    In [2]:

    ArrayBeforeSplitting = np.arange(150).reshape(5,30) 
    print ("Array shape without split operation is : " ,ArrayBeforeSplitting.shape)

    ('Array shape without split operation is : ', (5, 30))

    specify number of splits
    In [3]:

    split_1D = tf.Variable([8,13,9])
    print("specify number of partions using 1-Dimen Variable:" , tf.shape(split_1D))

    ('specify number of partions using 1-Dimen Variable:', <tf.Tensor 'Shape:0' shape=(1,) dtype=int32>)

    Use tf.split

    Make 3 splits aong y axis so that we have (5,8) ,(5,13),(5,9) splits. The axis 1 add up to give 30-- we can see axis 1 has 30 elements so the partition along that axis should add up to 30 otherwise it gives error.
    In [6]:

    split1,split2,split3 = tf.split(ArrayBeforeSplitting,split_1D,1)
    # we have 3 splits along axis 1 specified spcifically
    # by the split_1D . That is split axis 1D (with 30 elements) into partions with 8 ,13, and 9 elements while the x axis
    #remains constant

    In [7]:

    #INitialise global variables. because split_ID is a variable and needs to be initialised before being
    #used in a computational graph
    init_op = tf.global_variables_initializer()

    In [16]:

    with tf.Session() as sess:
        sess.run(init_op) # run variable initialisation.
        result=split1.eval();print("\n")
        print(result)
        print("the shape of the first split operation is : ",result.shape)
        result2=split2.eval();print("\n")
        print(result2)
        print("the shape of the second split operation is : ",result2.shape)

        result3=split3.eval();print("\n")
        print(result3)
        print("the shape of the third split operation is : ",result3.shape)


    [[  0   1   2   3   4   5   6   7]
     [ 30  31  32  33  34  35  36  37]
     [ 60  61  62  63  64  65  66  67]
     [ 90  91  92  93  94  95  96  97]
     [120 121 122 123 124 125 126 127]]
    ('the shape of the first split operation is : ', (5, 8))


    [[  8   9  10  11  12  13  14  15  16  17  18  19  20]
     [ 38  39  40  41  42  43  44  45  46  47  48  49  50]
     [ 68  69  70  71  72  73  74  75  76  77  78  79  80]
     [ 98  99 100 101 102 103 104 105 106 107 108 109 110]
     [128 129 130 131 132 133 134 135 136 137 138 139 140]]
    ('the shape of the second split operation is : ', (5, 13))

希望这可以帮助！