我已经将theano模型中的权重和偏差读作numpy数组:
w1 = np.zeros((3,3,3,100))
w1 = tf.convert_to_tensor(w1,tf.float32)
w = tf.get_variable('get_variable', dtype=tf.float32, initializer=w1)
net['left'] = tf.contrib.layers.convolution2d(
inputs=input_l,
num_outputs=100,
kernel_size=kernel_size,
weights_initializer=w,
padding='SAME',
data_format='NCHW',
)
当我收到以下错误时,我必须更改 w ?
ValueError: If initializer is a constant, do not specify shape.
编辑:
这对我有用:
layer1 = tf.nn.conv2d(t_in_left, w, strides=[1,1,1,1], padding='SAME')