python - TF2.0 中的 saved_model.prune()

Question

我正在尝试修剪SavedModel使用 tf.keras 生成的节点。剪枝脚本如下：

svmod = tf.saved_model.load(fn) #version 1
#svmod = tfk.experimental.load_from_saved_model(fn) #version 2
feeds = ['foo:0']
fetches = ['bar:0']
svmod2 = svmod.prune(feeds=feeds, fetches=fetches)
tf.saved_model.save(svmod2, '/tmp/saved_model/') #version 1
#tfk.experimental.export_saved_model(svmod2, '/tmp/saved_model/') #version 2

如果我使用版本 #1 修剪有效，但ValueError: Expected a Trackable object for export在保存时会给出。在版本 2 中，没有 prune() 方法。

如何修剪 TF2.0 Keras SavedModel？

标签： pythontensorflowtf.keras

Answer 1

看起来您在版本 1 中修剪模型的方式很好；根据您的错误消息，生成的修剪模型无法保存，因为它不是“可跟踪的”，这是保存模型的必要条件tf.saved_model.save。制作可跟踪对象的一种方法是从tf.Module类继承，如使用 SavedModel 格式和具体函数的指南中所述。下面是尝试保存tf.function对象（由于对象不可跟踪而失败）、继承自tf.module并保存结果对象的示例：

（使用 Python 3.7.6 版、TensorFlow 2.1.0 版和 NumPy 1.18.1 版）

import tensorflow as tf, numpy as np

# Define a random TensorFlow function and generate a reference output
conv_filter = tf.random.normal([1, 2, 4, 2], seed=1254)
@tf.function
def conv_model(x):
    return tf.nn.conv2d(x, conv_filter, 1, "SAME")

input_tensor = tf.ones([1, 2, 3, 4])
output_tensor = conv_model(input_tensor)
print("Original model outputs:", output_tensor, sep="\n")

# Try saving the model: it won't work because a tf.function is not trackable
export_dir = "./tmp/"
try: tf.saved_model.save(conv_model, export_dir)
except ValueError: print(
    "Can't save {} object because it's not trackable".format(type(conv_model)))

# Now define a trackable object by inheriting from the tf.Module class
class MyModule(tf.Module):
    @tf.function
    def __call__(self, x): return conv_model(x)

# Instantiate the trackable object, and call once to trace-compile a graph
module_func = MyModule()
module_func(input_tensor)
tf.saved_model.save(module_func, export_dir)

# Restore the model and verify that the outputs are consistent
restored_model = tf.saved_model.load(export_dir)
restored_output_tensor = restored_model(input_tensor)
print("Restored model outputs:", restored_output_tensor, sep="\n")
if np.array_equal(output_tensor.numpy(), restored_output_tensor.numpy()):
    print("Outputs are consistent :)")
else: print("Outputs are NOT consistent :(")

控制台输出：

Original model outputs:
tf.Tensor(
[[[[-2.3629642   1.2904963 ]
   [-2.3629642   1.2904963 ]
   [-0.02110204  1.3400152 ]]

  [[-2.3629642   1.2904963 ]
   [-2.3629642   1.2904963 ]
   [-0.02110204  1.3400152 ]]]], shape=(1, 2, 3, 2), dtype=float32)
Can't save <class 'tensorflow.python.eager.def_function.Function'> object
because it's not trackable
Restored model outputs:
tf.Tensor(
[[[[-2.3629642   1.2904963 ]
   [-2.3629642   1.2904963 ]
   [-0.02110204  1.3400152 ]]

  [[-2.3629642   1.2904963 ]
   [-2.3629642   1.2904963 ]
   [-0.02110204  1.3400152 ]]]], shape=(1, 2, 3, 2), dtype=float32)
Outputs are consistent :)

因此，您应该尝试如下修改您的代码：

svmod = tf.saved_model.load(fn) #version 1
svmod2 = svmod.prune(feeds=['foo:0'], fetches=['bar:0'])

class Exportable(tf.Module):
    @tf.function
    def __call__(self, model_inputs): return svmod2(model_inputs)

svmod2_export = Exportable()
svmod2_export(typical_input)    # call once with typical input to trace-compile
tf.saved_model.save(svmod2_export, '/tmp/saved_model/')

如果您不想从 继承tf.Module，您也可以只实例化一个tf.Module对象并tf.function通过替换该部分代码来添加方法/可调用属性，如下所示：

to_export = tf.Module()
to_export.call = tf.function(conv_model)
to_export.call(input_tensor)
tf.saved_model.save(to_export, export_dir)

restored_module = tf.saved_model.load(export_dir)
restored_func = restored_module.call