在keras中重用函数的权重

Question

我有一个函数叫做 block

from tensorflow.keras import layers
from tensorflow.keras.initializers import glorot_uniform

def block(X, classes):
    X = layers.Conv2D(filters = 3, kernel_size = (1, 1), strides = (1,1), name = 'a', kernel_initializer = glorot_uniform())(X)
    X = layers.Activation('relu')(X)

    X = layers.Flatten()(X)
    X = layers.Dense(classes, activation='linear', name='fc', kernel_initializer = glorot_uniform())(X)
    return X

block是一个神经网络模块。我想为多个输入重用 block 的权重。例如，假设我有两个输入：input1、input2。我将如何通过 block 传递它们，以便权重再次用于 input2。下面的代码不起作用，因为它为 B.

初始化了一组新的权重

input1 = layers.Input((64, 64, 3))
input2 = layers.Input((64, 64, 3))

A = block(input1, 10)
B = block(input2, 10)

print(A)
print(B)

Answer 1

如果您使 block 成为继承自 tf.keras.Model 的 class 或将其更改为 returns 成为 Sequential 模型，这可能会有所帮助与所有你想要的层。目前，您似乎没有在任何地方节省体重。

例如：

A = tf.keras.Sequential([
    # ... your layers
])

B = tf.keras.Sequential([
    # ... your layers
])


# Or use:

def block(params):
    model = tf.keras.Sequential([
        # ... your layers with params
    ])
    return model

A = block(params)
B = block(params)

# Or use

class block(tf.keras.Model):
  def __init__(self, params):
    super(block, self).__init__()
    self.layer1 = # ... first layer
    # ... your layers

  def call(self, inputs):
    X = self.layer1(inputs)
    # ... the rest of your layers
    return X

然后您应该能够像这样获得图层权重 A.layers[layer_number].get_weights() 并使用 B.layers[layer_number].set_weights([np_weight_arr]) 设置权重。

此外，我知道一些单独的图层接受 weights 参数，但我必须仔细检查它是否适用于您提到的所有图层。我知道你可以做类似 layers.Embedding(vocab_size+1, emb_dim, weights=[embedding_matrix], input_length=maxlength) 的事情，其中 embedding_matrix 是一个 numpy 数组。

最后，如果你可以重复使用 A，那将是最简单的解决方案。

Answer 2

重用的一种选择是使用共享模型。我们可以从 block() 中创建一个单独的模型，我们可以随时重复使用它。

import tensorflow as tf

from tensorflow.keras import layers
from tensorflow.keras import models
from tensorflow.keras.initializers import glorot_uniform

# this is my block
def block(X, classes):
    X = layers.Conv2D(filters = 3, kernel_size = (1, 1), strides = (1,1), kernel_initializer = glorot_uniform())(X)
    X = layers.Activation('relu')(X)

    X = layers.Flatten()(X)
    X = layers.Dense(classes, activation='linear', kernel_initializer = glorot_uniform())(X)
    return X


# block model

ip_base = layers.Input((64, 64, 3)) # decide on it
op = block(ip_base, 3)
block_model = models.Model(ip_base, op)

block_model_weights = block_model.get_weights()  # we can save this as pickle and later load for another model
print(block_model_weights)

# now we can save/re-use the weight of this model anytime we want
# secondary

input1 = layers.Input((64, 64, 3))
input2 = layers.Input((64, 64, 3))

A = block_model(input1)
B = block_model(input2)

model = models.Model([input1, input2], [A, B])

tf.keras.utils.plot_model(
    model)

输出：

[array([[[[ 0.07414746, -0.41065478, -0.5521915 ],
         [-0.6518545 ,  0.09209108,  0.9978211 ],
         [ 0.8638036 ,  0.5552795 ,  0.2620473 ]]]], dtype=float32), array([0., 0., 0.], dtype=float32), array([[-0.02172941,  0.00599053,  0.00549869],
       [-0.01580501, -0.01591844,  0.0140271 ],
       [ 0.0102962 ,  0.01141792, -0.00722483],
       ...,
       [ 0.00119783,  0.0168376 , -0.00014478],
       [ 0.00626496, -0.0100358 ,  0.01791514],
       [-0.00650628, -0.00537593,  0.00106668]], dtype=float32), array([0., 0., 0.], dtype=float32)]

如您所见，如果我们从基础创建一个 Model 对象，然后通过它传递任何输入，权重将被共享。

如果我们想要加载预训练的 block 模型，然后使用其先前的权重并传递我们的输入之一，我们可以简单地执行此操作。

# save the weights of the block model

weights = np.array(block_model.get_weights())

np.save('block_weights.npy', weights, allow_pickle = True)

# now we create another new block model
# block model

ip_base = layers.Input((64, 64, 3)) # decide on it
op = block(ip_base, 3)
new_block_model = models.Model(ip_base, op)

# currently it's randomly initialized but we can take the previous weights and load that in the current model

# load weight

old_weights = np.load('block_weights.npy', allow_pickle = True)

new_block_model.set_weights( old_weights )

# compare all of the weights

print(weights)
print('---------------')
print(old_weights)
print('---------------')
print(new_block_model.get_weights())

输出：

[array([[[[ 0.38135457, -0.28602505,  0.6248524 ],
         [-0.10373783,  0.20868587, -0.0295043 ],
         [ 0.073596  , -0.85106397,  0.86780167]]]], dtype=float32)
 array([0., 0., 0.], dtype=float32)
 array([[-0.00512073,  0.00298601,  0.015619  ],
       [ 0.01500274,  0.01736909, -0.0106204 ],
       [ 0.00690563,  0.00548493,  0.00449893],
       ...,
       [-0.00202469, -0.00226198,  0.0212799 ],
       [-0.01203138, -0.00065516, -0.01211848],
       [-0.01238732, -0.00368575,  0.02146613]], dtype=float32)
 array([0., 0., 0.], dtype=float32)]
---------------
[array([[[[ 0.38135457, -0.28602505,  0.6248524 ],
         [-0.10373783,  0.20868587, -0.0295043 ],
         [ 0.073596  , -0.85106397,  0.86780167]]]], dtype=float32)
 array([0., 0., 0.], dtype=float32)
 array([[-0.00512073,  0.00298601,  0.015619  ],
       [ 0.01500274,  0.01736909, -0.0106204 ],
       [ 0.00690563,  0.00548493,  0.00449893],
       ...,
       [-0.00202469, -0.00226198,  0.0212799 ],
       [-0.01203138, -0.00065516, -0.01211848],
       [-0.01238732, -0.00368575,  0.02146613]], dtype=float32)
 array([0., 0., 0.], dtype=float32)]
---------------
[array([[[[ 0.38135457, -0.28602505,  0.6248524 ],
         [-0.10373783,  0.20868587, -0.0295043 ],
         [ 0.073596  , -0.85106397,  0.86780167]]]], dtype=float32), array([0., 0., 0.], dtype=float32), array([[-0.00512073,  0.00298601,  0.015619  ],
       [ 0.01500274,  0.01736909, -0.0106204 ],
       [ 0.00690563,  0.00548493,  0.00449893],
       ...,
       [-0.00202469, -0.00226198,  0.0212799 ],
       [-0.01203138, -0.00065516, -0.01211848],
       [-0.01238732, -0.00368575,  0.02146613]], dtype=float32), array([0., 0., 0.], dtype=float32)]

在keras中重用函数的权重

Reuse weights of a function in keras

keras

tensorflow2.0