tensorflow - 使用带有 SavedModel 的 GradientTape 对输入进行梯度下降

Question

我正在尝试对使用 SavedModel 加载的模型进行对抗性攻击API。我想针对给定目标的模型损失对输入进行梯度下降。代码有点长，但它是说明问题的最低限度。

from __future__ import absolute_import, division, print_function, unicode_literals

from tensorflow import keras
from tensorflow.keras import layers, models
import tensorflow as tf
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split

# creating the dataset
X, y = make_classification(n_samples=10000, n_informative=10)
X, X_test, y, y_test = train_test_split(X, y)

# training the model
model = models.Sequential()
model.add(layers.Dense(10, activation='relu'))
model.add(layers.Dense(2, activation='softmax', name="output"))

model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

model.fit(X, y, 
          epochs=10,
          batch_size=32,
          verbose=0)

print(f"final accuracy is {model.evaluate(X_test, y_test, verbose=0)[1]}")

# saving and loading it
tf.saved_model.save(model, "/content/demo")
loaded_model = tf.saved_model.load("/content/demo")

inference_func = loaded_model.signatures["serving_default"]

# making the adversarial example
x = tf.random.normal([1, 20])
x = tf.Variable(x)

target = tf.convert_to_tensor([0, 1], dtype=tf.float32)
cce = tf.keras.losses.CategoricalCrossentropy()

with tf.GradientTape() as t:
    t.watch(x)
    y = inference_func(x)["output"]
    loss = cce(target, y)
dl_dx = t.gradient(loss, x)
x.assign_sub(learning_rate * dl_dx)

print(x.numpy())

我收到以下错误:

final accuracy is 0.8899999856948853
INFO:tensorflow:Assets written to: /content/demo/assets
---------------------------------------------------------------------------
InvalidArgumentError                      Traceback (most recent call last)
<ipython-input-31-1b61c316b9dc> in <module>()
     40 with tf.GradientTape() as t:
     41     t.watch(x)
---> 42     y = inference_func(x)["output"]
     43     loss = cce(target, y)
     44 dl_dx = t.gradient(loss, x)

6 frames
/usr/local/lib/python3.6/dist-packages/six.py in raise_from(value, from_value)

InvalidArgumentError: cannot compute __forward_signature_wrapper_147414 as input #0(zero-based) was expected to be a float tensor but is a resource tensor [Op:__forward_signature_wrapper_147414]

我对低级 tensorflow 很陌生，我还不太了解它是如何工作的。我相信这个问题与我的推理函数不是真正的 @tf.function 有关因为它的类型是 tensorflow.python.saved_model.load._WrapperFunction .但是我怎样才能检索到真正的功能呢？

Answer 1

我想到了 !所以我正在寻找的功能是loaded_model.__call__ .我不知道为什么 tensorflow doc 没有解释清楚。