如何计算张量流中不可训练的权重数？

Question

在 keras 中可以通过

计算不可训练权重的数量

from tensorflow.keras import backend as K
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Conv2D, Dense

#Dummy Keras model
model = tf.keras.Sequential([
        tf.keras.layers.Dense(3, input_dim=3072,activation="relu", name="inputlayer"),
        tf.keras.layers.Dense(512, activation="tanh", name="2ndlayer"),
        tf.keras.layers.Dense(512, activation="tanh", name="3rdlayer"),
        tf.keras.layers.Dense(512, activation="tanh", name="4thlayer"),
        tf.keras.layers.Dense(512, activation="tanh", name="5thlayer"),
        tf.keras.layers.Dense(3, name="lastlayer"),
])

trainable_count = int(np.sum([K.count_params(p) for p in set(model.trainable_weights)]))
non_trainable_count = int(np.sum([K.count_params(p) for p in set(model.non_trainable_weights)]))

print('Total params: {:,}'.format(trainable_count + non_trainable_count))
print('Trainable params: {:,}'.format(trainable_count))
print('Non-trainable params: {:,}'.format(non_trainable_count))

我有一个带有自定义层的自定义 tesorflow 1.15 模型，所以我无法利用 Keras。我知道我可以通过以下方式计算自定义模型中的可训练参数：

#Dummy data 
x = np.zeros((1,16,16,3))

#Dummy tf model 
x_tf = tf.convert_to_tensor(x, np.float32)
z_tf = tf.layers.conv2d(x_tf, filters=32, kernel_size=(3,3))
zz_tf =  tf.layers.conv2d(z_tf, filters=32, kernel_size=(3,3))

trainable_count = np.sum([np.prod(v.shape) for v in tf.trainable_variables()])
print(trainable_count)

我的问题是如何访问我的 tensorflow 模型的不可训练权重？

在 keras 中，它是用 non_trainable_count = int(np.sum([K.count_params(p) for p in set(model.non_trainable_weights)])) 完成的，所以我正在寻找等效的命令而不调用 keras api。

Answer 1

您可以使用以下脚本计算一层的不可训练权重，然后对网络的所有层求和

    class Dense(tf.Module):
        def __init__(self, input_shape, output_shape):
            super(Dense, self).__init__()
            self.w = tf.Variable(tf.random.normal([input_shape, output_shape]), 
                                 trainable= False)
            self.b = tf.Variable(tf.zeros([output_shape]))
        def __call__(self,x):
            y  = tf.matmul(x,self.w) + self.b
            return tf.nn.relu(y)
        def count_non_trainable_weights(self):
            return(len(self.variables) - len(self.trainable_variables))

如何计算张量流中不可训练的权重数？

How to count the number of non trainable weights in tensorflow?

python

memory

keras

tensorflow

tf.keras