TF2-GradientTape与Model.fit-为什么GradientTape不起作用？

问题描述

晚上好

我想用tf2和Gradient Tape函数实现一个简单的回归问题的玩具示例。使用Model.fit，它可以正常学习，但是使用GradientTape可以做到，但与model.fit（）相比，损失不会增加。这是我的示例代码和结果。我找不到问题。

model_opt = tf.keras.optimizers.Adam() 
loss_fn = tf.keras.losses.MeanSquaredError()
with tf.GradientTape() as tape:
    y = model(X,training=True)
    loss_value = loss_fn(y_true,y)
grads = tape.gradient(loss_value,model.trainable_variables)
model_opt.apply_gradients(zip(grads,model.trainable_variables))

#Results:
42.47433806265809
42.63973672226078
36.687397360178586
38.744844324717526
36.59080452300609
...

这是带有model.fit（）的常规案例

model.compile(optimizer=tf.keras.optimizers.Adam(),loss=tf.keras.losses.MSE,metrics="mse")
...
model.fit(X,y_true,verbose=0)
#Results
[40.97759069299212]
[28.04145720307729]
[17.643483147375473]
[7.575242056454791]
[5.83682193867299]

精度应该大致相同，但看起来根本没有学习。输入X是张量，y_true也是。

编辑 进行测试

import pathlib
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers

dataset_path = keras.utils.get_file("auto-mpg.data","http://archive.ics.uci.edu/ml/machine-learning-databases/auto-mpg/auto-mpg.data")

column_names = ['MPG','Cylinders','displacement','Horsepower','Weight','acceleration','Model Year','Origin']
dataset = pd.read_csv(dataset_path,names=column_names,na_values = "?",comment='\t',sep=" ",skipinitialspace=True)

dataset = dataset.dropna()
dataset['Origin'] = dataset['Origin'].map({1: 'USA',2: 'Europe',3: 'Japan'})
dataset = pd.get_dummies(dataset,prefix='',prefix_sep='')

train_dataset = dataset.sample(frac=0.8,random_state=0)
test_dataset = dataset.drop(train_dataset.index)

train_stats = train_dataset.describe()
train_stats.pop("MPG")
train_stats = train_stats.transpose()

train_labels = train_dataset.pop('MPG')
test_labels = test_dataset.pop('MPG')

def norm(x):
  return (x - train_stats['mean']) / train_stats['std']

normed_train_data = norm(train_dataset)
normed_test_data = norm(test_dataset)

def build_model_fit():
  model = keras.Sequential([
    layers.Dense(64,activation='relu',input_shape=[len(train_dataset.keys())]),layers.Dense(64,activation='relu'),layers.Dense(1)])
  optimizer = tf.keras.optimizers.RMSprop(0.001)
  model.compile(loss='mse',optimizer=optimizer)
  return model

def build_model_tape():
  model = keras.Sequential([
    layers.Dense(64,layers.Dense(1)])
  opt = tf.keras.optimizers.RMSprop(0.001)
  return model,opt

model_f = build_model_fit()
model_g,opt_g = build_model_tape()

EPOCHS = 20

#Model.fit() - Test
history = model_f.fit(normed_train_data,train_labels,epochs=EPOCHS,verbose=2)

X = tf.convert_to_tensor(normed_train_data.to_numpy())
y_true = tf.convert_to_tensor(train_labels.to_numpy())

#GradientTape - Test
loss_fn = tf.keras.losses.MeanSquaredError()
for i in range(0,EPOCHS):
    with tf.GradientTape() as tape:
        y = model_g(X,training=True)
        loss_value = loss_fn(y_true,y)
    grads = tape.gradient(loss_value,model_g.trainable_variables)
    opt_g.apply_gradients(zip(grads,model_g.trainable_variables))
    print(loss_value)

解决方法

OP在损失值中看到的差异是由于在model.fit和tf.GradientTape训练循环中使用了不同的批次大小。如果未指定batch_size的{​​{1}}关键字参数，则将使用32的批处理大小。在model.fit训练循环中，批次大小等于训练集中的样本数量（即314）。

要解决此问题，请在训练循环中实施批处理。一种实现方法是使用tf.GradientTape API，如下所示。

tf.data

还请注意，loss_fn = tf.keras.losses.MeanSquaredError() for i in range(0,EPOCHS): epoch_losses = [] for x_batch,y_batch in tf.data.Dataset.from_tensor_slices((X,y_true)).batch(32): with tf.GradientTape() as tape: y = model_g(x_batch,training=True) loss_value = loss_fn(y_batch,y) epoch_losses.append(loss_value.numpy()) grads = tape.gradient(loss_value,model_g.trainable_variables) opt_g.apply_gradients(zip(grads,model_g.trainable_variables)) print(np.mean(loss_value)) 每次迭代都会对数据进行混洗，而自定义训练循环则不会（需要由开发人员实施）。