使用kerassurgeon对xception模型进行通道剪枝

Question

所以我是神经网络的新手，我手头有修剪神经网络模型的任务。为此，我正在修剪 xception 模型。

我正在为此使用 kerassugeon 并执行频道修剪。以下是相关的导入：

from kerassurgeon import identify
from kerassurgeon.operations import delete_channels, delete_layer

这是相同的代码：

#load pre trained Xception model
model=tf.keras.applications.xception.Xception(weights='imagenet',include_top=True)

weight = model.layers[1].get_weights()[0]
weight_dict = {}
num_filters = len(weight[0,0,0,:])


for j in range(num_filters) :
  w_s = np.sum(abs(weight[:,:,:,j]))
  filt = 'filt_{}'.format(j)
  weight_dict[filt]=w_s


weights_dict_sort = sorted(weight_dict.items(), key = lambda kv:kv[1])
print('L1 norm conv layer {}\n'.format(1), weights_dict_sort)

weights_value = []
for elem in weights_dict_sort :
  weights_value.append(elem[1])

xc = range(num_filters)

layer_0 = model.layers[1]
model_new = delete_channels(model=model, layer=layer_0, channels=[1,7,2,8,16])

model_new.compile(loss = 'binary_crossentropy', optimizer = 'rmsprop', metrics = ['accuracy'])

这给我带来了以下错误：

TypeError                                 Traceback (most recent call last)

<ipython-input-42-8ec328142934> in <module>()
     23 
     24 layer_0 = model.layers[1]
---> 25 model_new = delete_channels(model=model, layer=layer_0, channels=[1,7,2,8,16])
     26 
     27 model_new.compile(loss = 'binary_crossentropy', optimizer = 'rmsprop', metrics = ['accuracy'])

4 frames

/usr/local/lib/python3.7/dist-packages/kerassurgeon/operations.py in delete_channels(model, layer, channels, node_indices, copy)
    103     surgeon = Surgeon(model, copy)
    104     surgeon.add_job('delete_channels', layer, node_indices=node_indices, channels=channels)
--> 105     return surgeon.operate()

/usr/local/lib/python3.7/dist-packages/kerassurgeon/surgeon.py in operate(self)
    157             # Perform surgery at this node
    158             kwargs = self._kwargs_map[node]
--> 159             self._mod_func_map[node](node, outputs, output_masks, **kwargs)
    160 
    161         # Finish rebuilding model

/usr/local/lib/python3.7/dist-packages/kerassurgeon/surgeon.py in _delete_channels(self, node, inputs, input_masks, channels, layer_name)
    317         """
    318         old_layer = node.outbound_layer
--> 319         old_layer_output = utils.single_element(node.output_tensors)
    320         # Create a mask to propagate the deleted channels to downstream layers
    321         new_delete_mask = self._make_delete_mask(old_layer, channels)

/usr/local/lib/python3.7/dist-packages/kerassurgeon/utils.py in single_element(x)
    141         return x
    142 
--> 143     if len(x) == 1:
    144         x = x[0]
    145     return x

/usr/local/lib/python3.7/dist-packages/tensorflow/python/keras/engine/keras_tensor.py in __len__(self)
    238 
    239   def __len__(self):
--> 240     raise TypeError('Keras symbolic inputs/outputs do not '
    241                     'implement `__len__`. You may be '
    242                     'trying to pass Keras symbolic inputs/outputs '

TypeError: Keras symbolic inputs/outputs do not implement `__len__`. You may be trying to pass Keras symbolic inputs/outputs to a TF API that does not register dispatching, preventing Keras from automatically converting the API call to a lambda layer in the Functional Model. This error will also get raised if you try asserting a symbolic input/output directly.

我正在为此开发 google colab：https://colab.research.google.com/drive/1n5uv3d0dBv3uct6pzaB8bs6q0eMO2sID?usp=sharing

非常感谢任何帮助调试这个问题，如果需要任何其他信息来澄清这个问题，请告诉我。

Answer 1

所以我想通了。

这与tensorflows版本有关。据我了解，kerassurgeon 是在早期版本的 tf 上构建的，而新版本的 tf 有一些重载的 len 方法，这导致了一些问题。

因此您可以按照以下步骤之一进行操作：

降级到 TF 1.14（推荐）
删除 TensorFlow 源码中的 len 重载（不推荐）

第一步我是这样操作的：

确保您在重新运行代码以清除已安装的 tf 依赖项之前重置了您的 colab 环境。

希望这对某人有所帮助。

使用kerassurgeon对xception模型进行通道剪枝

Channel pruning of xception model using kerassurgeon

python

deep-learning

conv-neural-network

keras

tensorflow