我想将一些由另一个网络训练的权重传输到 TensorFlow,这些权重存储在像这样的单个向量中:
[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 reshape 为两个 3 x 3 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)
。然而,在 TensorFlow 中,过滤器的形状是 (3,3,2,1)
:
tf_weights = tf.Variable(tf.random_normal([3,3,2,1]))
将 tf_weights reshape 为预期的形状后,权重变得一团糟,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[number,channel,size,size]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,通过像这样使用 tensorflow 的 conv2d:
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feature map是错误的。
最佳答案
不要使用np.reshape
。它可能 mess up the order of your values .
使用np.rollaxis
相反:
>>> 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,两个 rollaxis
操作会导致形状更改为 (1, 2, 31, 32) -> (1, 31, 32, 2) -> (31, 32, 2 , 1).您的最终数组如下所示:
>>> b
array([[[[ 1],
[10]],
[[ 2],
[11]],
[[ 3],
[12]]],
[[[ 4],
[13]],
[[ 5],
[14]],
[[ 6],
[15]]],
[[[ 7],
[16]],
[[ 8],
[17]],
[[ 9],
[18]]]])
关于python - 如何将矢量 reshape 为 TensorFlow 的过滤器?,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/34725157/