TF2：在非急切模式下计算 keras 回调中的梯度

Question

TF 版本：2.2.0-rc3（在 Colab 中）

我在回调中使用以下代码（来自tf.keras get computed gradient during training）来计算模型中所有参数的梯度。

def on_train_begin(self, logs=None):
        # Functions return weights of each layer
        self.layerweights = []
        for lndx, l in enumerate(self.model.layers):
            if hasattr(l, 'kernel'):
                self.layerweights.append(l.kernel)

        input_tensors = [self.model.inputs[0],
                        self.model.sample_weights[0],
                        self.model.targets[0],
                        K.learning_phase()]

        # Get gradients of all the relevant layers at once
        grads = self.model.optimizer.get_gradients(self.model.total_loss, self.layerweights)
        self.get_gradients = K.function(inputs=input_tensors,outputs=grads)

但是，当我运行它时，我收到以下错误。

AttributeError: 'Model' object has no attribute 'sample_weights'

model.targets 也出现同样的错误。

如何在回调中获取渐变？

在 Eager 模式下，解决方案 Get Gradients with Keras Tensorflow 2.0 有效。但是，我想在 Non-eager 模式下使用它。

Answer 1

这是使用 keras 后端捕获渐变的端到端代码。我已经从 model.fit 的回调中调用梯度捕获函数来捕获每个时期结束后的梯度。 此代码在 tensorflow 1.x 和 tensorflow 2.x 版本中都兼容，并且我已经在 colab 中运行过。 如果您想在 tensorflow 1.x 中运行，请将第一个语句替换为带有%tensorflow_version 1.x 的程序并重新启动运行时。

捕捉模型的梯度 -

# Importing dependency
%tensorflow_version 2.x
from tensorflow import keras
from tensorflow.keras import backend as K
from tensorflow.keras import datasets
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Activation, Dropout, Flatten, Conv2D, MaxPooling2D
from tensorflow.keras.layers import BatchNormalization
import numpy as np
import tensorflow as tf

tf.keras.backend.clear_session()  # For easy reset of notebook state.
tf.compat.v1.disable_eager_execution()

# Import Data
(train_images, train_labels), (test_images, test_labels) = datasets.cifar10.load_data()

# Build Model
model = Sequential()
model.add(Conv2D(32, (3, 3), activation='relu', input_shape=(32, 32, 3)))
model.add(MaxPooling2D((2, 2)))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D((2, 2)))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(Flatten())
model.add(Dense(64, activation='relu'))
model.add(Dense(10))

# Model Summary
model.summary()

# Model Compile 
model.compile(optimizer='adam',
              loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

# Define the Gradient Fucntion
epoch_gradient = []

# Define the Gradient Function
def get_gradient_func(model):
    grads = K.gradients(model.total_loss, model.trainable_weights)
    inputs = model._feed_inputs + model._feed_targets + model._feed_sample_weights
    func = K.function(inputs, grads)
    return func

# Define the Required Callback Function
class GradientCalcCallback(keras.callbacks.Callback):
  def on_epoch_end(self, epoch, logs=None):
      get_gradient = get_gradient_func(model)
      grads = get_gradient([train_images, train_labels, np.ones(len(train_labels))])
      epoch_gradient.append(grads)

epoch = 4

model.fit(train_images, train_labels, epochs=epoch, validation_data=(test_images, test_labels), callbacks=[GradientCalcCallback()])


# (7) Convert to a 2 dimensiaonal array of (epoch, gradients) type
gradient = np.asarray(epoch_gradient)
print("Total number of epochs run:", epoch)
print("Gradient Array has the shape:",gradient.shape)

输出 -

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 30, 30, 32)        896       
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 15, 15, 32)        0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 13, 13, 64)        18496     
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 6, 6, 64)          0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 4, 4, 64)          36928     
_________________________________________________________________
flatten (Flatten)            (None, 1024)              0         
_________________________________________________________________
dense (Dense)                (None, 64)                65600     
_________________________________________________________________
dense_1 (Dense)              (None, 10)                650       
=================================================================
Total params: 122,570
Trainable params: 122,570
Non-trainable params: 0
_________________________________________________________________
Train on 50000 samples, validate on 10000 samples
Epoch 1/4
50000/50000 [==============================] - 73s 1ms/sample - loss: 1.8199 - accuracy: 0.3834 - val_loss: 1.4791 - val_accuracy: 0.4548
Epoch 2/4
50000/50000 [==============================] - 357s 7ms/sample - loss: 1.3590 - accuracy: 0.5124 - val_loss: 1.2661 - val_accuracy: 0.5520
Epoch 3/4
50000/50000 [==============================] - 377s 8ms/sample - loss: 1.1981 - accuracy: 0.5787 - val_loss: 1.2625 - val_accuracy: 0.5674
Epoch 4/4
50000/50000 [==============================] - 345s 7ms/sample - loss: 1.0838 - accuracy: 0.6183 - val_loss: 1.1302 - val_accuracy: 0.6083
Total number of epochs run: 4
Gradient Array has the shape: (4, 10)

希望这能回答您的问题。快乐学习。