有三種計算圖的構建方式:靜態計算圖,動態計算圖,以及Autograph。
TensorFlow 2.0主要使用的是動態計算圖和Autograph。
動態計算圖易於調試,編碼效率較高,但執行效率偏低。
靜態計算圖執行效率很高,但較難調試。
而Autograph機制可以將動態圖轉換成靜態計算圖,兼收執行效率和編碼效率之利。
當然Autograph機制能夠轉換的代碼並不是沒有任何約束的,有一些編碼規范需要遵循,否則可能會轉換失敗或者不符合預期。
前面我們介紹了Autograph的編碼規范和Autograph轉換成靜態圖的原理。
本篇我們介紹使用tf.Module來更好地構建Autograph。
一,Autograph和tf.Module概述
前面在介紹Autograph的編碼規范時提到構建Autograph時應該避免在@tf.function修飾的函數內部定義tf.Variable.
但是如果在函數外部定義tf.Variable的話,又會顯得這個函數有外部變量依賴,封裝不夠完美。
一種簡單的思路是定義一個類,並將相關的tf.Variable創建放在類的初始化方法中。而將函數的邏輯放在其他方法中。
這樣一頓猛如虎的操作之后,我們會覺得一切都如同人法地地法天天法道道法自然般的自然。
驚喜的是,TensorFlow提供了一個基類tf.Module,通過繼承它構建子類,我們不僅可以獲得以上的自然而然,而且可以非常方便地管理變量,還可以非常方便地管理它引用的其它Module,最重要的是,我們能夠利用tf.saved_model保存模型並實現跨平台部署使用。
實際上,tf.keras.models.Model,tf.keras.layers.Layer 都是繼承自tf.Module的,提供了方便的變量管理和所引用的子模塊管理的功能。
因此,利用tf.Module提供的封裝,再結合TensoFlow豐富的低階API,實際上我們能夠基於TensorFlow開發任意機器學習模型(而非僅僅是神經網絡模型),並實現跨平台部署使用。
二,應用tf.Module封裝Autograph
定義一個簡單的function。
import tensorflow as tf x = tf.Variable(1.0,dtype=tf.float32) # 在tf.function中用input_signature限定輸入張量的簽名類型:shape和dtype @tf.function(input_signature=[tf.TensorSpec(shape = [], dtype = tf.float32)]) def add_print(a): x.assign_add(a) tf.print(x) return(x) add_print(tf.constant(3.0)) # add_print(tf.constant(3)) #輸入不符合張量簽名的參數將報錯
4
下面利用tf.Module的子類化將其封裝一下。
class DemoModule(tf.Module): def __init__(self,init_value = tf.constant(0.0),name=None): super(DemoModule, self).__init__(name=name) with self.name_scope: #相當於with tf.name_scope("demo_module") self.x = tf.Variable(init_value,dtype = tf.float32,trainable=True) @tf.function(input_signature=[tf.TensorSpec(shape = [], dtype = tf.float32)]) def addprint(self,a): with self.name_scope: self.x.assign_add(a) tf.print(self.x) return(self.x) # 執行 demo = DemoModule(init_value = tf.constant(1.0)) result = demo.addprint(tf.constant(5.0))
6
# 查看模塊中的全部變量和全部可訓練變量 print(demo.variables) print(demo.trainable_variables)
(<tf.Variable 'demo_module/Variable:0' shape=() dtype=float32, numpy=6.0>,)
(<tf.Variable 'demo_module/Variable:0' shape=() dtype=float32, numpy=6.0>,)
# 查看模塊中的全部子模塊 demo.submodules
# 使用tf.saved_model 保存模型,並指定需要跨平台部署的方法 tf.saved_model.save(demo,"./data/demo/1",signatures = {"serving_default":demo.addprint})
WARNING:tensorflow:From /usr/local/lib/python3.6/dist-packages/tensorflow/python/ops/resource_variable_ops.py:1817: calling BaseResourceVariable.__init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version. Instructions for updating: If using Keras pass *_constraint arguments to layers. INFO:tensorflow:Assets written to: ./data/demo/1/assets
# 加載模型 demo2 = tf.saved_model.load("./data/demo/1") demo2.addprint(tf.constant(5.0))
11
<tf.Tensor: shape=(), dtype=float32, numpy=11.0>
# 查看模型文件相關信息,紅框標出來的輸出信息在模型部署和跨平台使用時有可能會用到 !saved_model_cli show --dir ./data/demo/1 --all
MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs: signature_def['__saved_model_init_op']: The given SavedModel SignatureDef contains the following input(s): The given SavedModel SignatureDef contains the following output(s): outputs['__saved_model_init_op'] tensor_info: dtype: DT_INVALID shape: unknown_rank name: NoOp Method name is: signature_def['serving_default']: The given SavedModel SignatureDef contains the following input(s): inputs['a'] tensor_info: dtype: DT_FLOAT shape: () name: serving_default_a:0 The given SavedModel SignatureDef contains the following output(s): outputs['output_0'] tensor_info: dtype: DT_FLOAT shape: () name: StatefulPartitionedCall:0 Method name is: tensorflow/serving/predict WARNING: Logging before flag parsing goes to stderr. W0411 02:47:30.452981 139671888869248 deprecation.py:506] From /usr/local/lib/python2.7/dist-packages/tensorflow_core/python/ops/resource_variable_ops.py:1786: calling __init__ (from tensorflow.python.ops.resource_variable_ops) with constraint is deprecated and will be removed in a future version. Instructions for updating: If using Keras pass *_constraint arguments to layers. Defined Functions: Function Name: 'addprint' Option #1 Callable with: Argument #1 a: TensorSpec(shape=(), dtype=tf.float32, name=u'a')
在tensorboard中查看計算圖,模塊會被添加模塊名demo_module,方便層次化呈現計算圖結構。
import datetime # 創建日志 stamp = datetime.datetime.now().strftime("%Y%m%d-%H%M%S") logdir = './data/demomodule/%s' % stamp writer = tf.summary.create_file_writer(logdir) # 開啟autograph跟蹤 tf.summary.trace_on(graph=True, profiler=True) # 執行autograph demo = DemoModule(init_value = tf.constant(0.0)) result = demo.addprint(tf.constant(5.0)) # 將計算圖信息寫入日志 with writer.as_default(): tf.summary.trace_export( name="demomodule", step=0, profiler_outdir=logdir) # 啟動 tensorboard在jupyter中的魔法命令 %reload_ext tensorboard from tensorboard import notebook notebook.list() notebook.start("--logdir ./data/demomodule/")
除了利用tf.Module的子類化實現封裝,我們也可以通過給tf.Module添加屬性的方法進行封裝。
mymodule = tf.Module() mymodule.x = tf.Variable(0.0) @tf.function(input_signature=[tf.TensorSpec(shape = [], dtype = tf.float32)]) def addprint(a): mymodule.x.assign_add(a) tf.print(mymodule.x) return (mymodule.x) mymodule.addprint = addprint mymodule.addprint(tf.constant(1.0)).numpy() print(mymodule.variables) # 使用tf.saved_model 保存模型 tf.saved_model.save(mymodule,"./data/mymodule", signatures = {"serving_default":mymodule.addprint}) # 加載模型 mymodule2 = tf.saved_model.load("./data/mymodule") mymodule2.addprint(tf.constant(5.0))
1 (<tf.Variable 'Variable:0' shape=() dtype=float32, numpy=1.0>,) INFO:tensorflow:Assets written to: ./data/mymodule/assets 6
<tf.Tensor: shape=(), dtype=float32, numpy=6.0>
三,tf.Module和tf.keras.Model,tf.keras.layers.Layer
tf.keras中的模型和層都是繼承tf.Module實現的,也具有變量管理和子模塊管理功能。
import tensorflow as tf from tensorflow.keras import models,layers,losses,metrics print(issubclass(tf.keras.Model,tf.Module)) print(issubclass(tf.keras.layers.Layer,tf.Module)) print(issubclass(tf.keras.Model,tf.keras.layers.Layer))
True
True
True
tf.keras.backend.clear_session() model = models.Sequential() model.add(layers.Dense(4,input_shape = (10,))) model.add(layers.Dense(2)) model.add(layers.Dense(1)) model.summary()
model.variables
[<tf.Variable 'dense/kernel:0' shape=(10, 4) dtype=float32, numpy= array([[-0.24266458, -0.45152673, -0.5430875 , -0.35098866], [ 0.36108053, -0.32325 , 0.3329792 , 0.33279514], [ 0.2944306 , -0.5975202 , -0.06157887, 0.25049144], [ 0.5707406 , 0.6214677 , -0.32870707, -0.12539297], [ 0.41170907, -0.5257766 , 0.12482923, -0.11132008], [-0.41743976, -0.3998926 , -0.46740663, 0.6105366 ], [ 0.54347396, 0.5108323 , 0.4747305 , -0.404514 ], [ 0.4390788 , -0.1988923 , 0.40562296, 0.57931125], [-0.2694599 , -0.4149857 , 0.07898462, -0.05845898], [-0.02557009, -0.440827 , -0.26627067, -0.0769726 ]], dtype=float32)>, <tf.Variable 'dense/bias:0' shape=(4,) dtype=float32, numpy=array([0., 0., 0., 0.], dtype=float32)>, <tf.Variable 'dense_1/kernel:0' shape=(4, 2) dtype=float32, numpy= array([[ 0.17386723, 0.9130187 ], [-0.88832307, -0.20379901], [ 0.9303725 , -0.4667368 ], [-0.8743646 , -0.31934786]], dtype=float32)>, <tf.Variable 'dense_1/bias:0' shape=(2,) dtype=float32, numpy=array([0., 0.], dtype=float32)>, <tf.Variable 'dense_2/kernel:0' shape=(2, 1) dtype=float32, numpy= array([[ 0.4870274 ], [-0.71679246]], dtype=float32)>, <tf.Variable 'dense_2/bias:0' shape=(1,) dtype=float32, numpy=array([0.], dtype=float32)>]
model.layers[0].trainable = False #凍結第0層的變量,使其不可訓練 model.trainable_variables
[<tf.Variable 'dense_1/kernel:0' shape=(4, 2) dtype=float32, numpy= array([[ 0.17386723, 0.9130187 ], [-0.88832307, -0.20379901], [ 0.9303725 , -0.4667368 ], [-0.8743646 , -0.31934786]], dtype=float32)>, <tf.Variable 'dense_1/bias:0' shape=(2,) dtype=float32, numpy=array([0., 0.], dtype=float32)>, <tf.Variable 'dense_2/kernel:0' shape=(2, 1) dtype=float32, numpy= array([[ 0.4870274 ], [-0.71679246]], dtype=float32)>, <tf.Variable 'dense_2/bias:0' shape=(1,) dtype=float32, numpy=array([0.], dtype=float32)>]
model.submodules
(<tensorflow.python.keras.engine.input_layer.InputLayer at 0x7fac6c6c2278>, <tensorflow.python.keras.layers.core.Dense at 0x7fac6e3e1908>, <tensorflow.python.keras.layers.core.Dense at 0x7fac6c6c2438>, <tensorflow.python.keras.layers.core.Dense at 0x7fac6c6c2470>)
model.layers
[<tensorflow.python.keras.layers.core.Dense at 0x7fac6e3e1908>, <tensorflow.python.keras.layers.core.Dense at 0x7fac6c6c2438>, <tensorflow.python.keras.layers.core.Dense at 0x7fac6c6c2470>]
print(model.name) print(model.name_scope())
sequential
sequential
參考:
開源電子書地址:https://lyhue1991.github.io/eat_tensorflow2_in_30_days/
GitHub 項目地址:https://github.com/lyhue1991/eat_tensorflow2_in_30_days