tensorflow - tensorflow federated中训练和评估期间的MSE误差不同

Question

我在 tensorflow federated 中实现回归模型。我从本教程中用于 keras 的简单模型开始：https ://www.tensorflow.org/tutorials/keras/regression

我将模型更改为使用联邦学习。这是我的模型：

import pandas as pd
import tensorflow as tf

from tensorflow import keras
from tensorflow.keras import layers
import tensorflow_federated as tff

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']
raw_dataset = pd.read_csv(dataset_path, names=column_names,
                      na_values = "?", comment='\t',
                      sep=" ", skipinitialspace=True)

df = raw_dataset.copy()
df = df.dropna()
dfs = [x for _, x in df.groupby('Origin')]


datasets = []
targets = []
for dataframe in dfs:
    target = dataframe.pop('MPG')

    from sklearn.preprocessing import StandardScaler
    standard_scaler_x = StandardScaler(with_mean=True, with_std=True)
    normalized_values = standard_scaler_x.fit_transform(dataframe.values)

    dataset = tf.data.Dataset.from_tensor_slices(({ 'x': normalized_values, 'y': target.values}))
    train_dataset = dataset.shuffle(len(dataframe)).repeat(10).batch(20)
    test_dataset = dataset.shuffle(len(dataframe)).batch(1)
    datasets.append(train_dataset)


def build_model():
  model = keras.Sequential([
    layers.Dense(64, activation='relu', input_shape=[7]),
    layers.Dense(64, activation='relu'),
    layers.Dense(1)
  ])
  return model
dataset_path


import collections


model = build_model()

sample_batch = tf.nest.map_structure(
    lambda x: x.numpy(), iter(datasets[0]).next())

def loss_fn_Federated(y_true, y_pred):
    return tf.reduce_mean(tf.keras.losses.MSE(y_true, y_pred))

def create_tff_model():
  keras_model_clone = tf.keras.models.clone_model(model)
#   adam = keras.optimizers.Adam()
  adam = tf.keras.optimizers.SGD(0.002)
  keras_model_clone.compile(optimizer=adam, loss='mse', metrics=[tf.keras.metrics.MeanSquaredError()])
  return tff.learning.from_compiled_keras_model(keras_model_clone, sample_batch)

print("Create averaging process")
# This command builds all the TensorFlow graphs and serializes them: 
iterative_process = tff.learning.build_federated_averaging_process(model_fn=create_tff_model)

print("Initzialize averaging process")
state = iterative_process.initialize()

print("Start iterations")
for _ in range(10):
  state, metrics = iterative_process.next(state, datasets)
  print('metrics={}'.format(metrics))

Start iterations
metrics=<mean_squared_error=95.8644027709961,loss=96.28633880615234>
metrics=<mean_squared_error=9.511247634887695,loss=9.522096633911133>
metrics=<mean_squared_error=8.26853084564209,loss=8.277074813842773>
metrics=<mean_squared_error=7.975323677062988,loss=7.9771647453308105>
metrics=<mean_squared_error=7.618809700012207,loss=7.644164562225342>
metrics=<mean_squared_error=7.347906112670898,loss=7.340310096740723>
metrics=<mean_squared_error=7.210267543792725,loss=7.210223197937012>
metrics=<mean_squared_error=7.045553207397461,loss=7.045469760894775>
metrics=<mean_squared_error=6.861278533935547,loss=6.878870487213135>
metrics=<mean_squared_error=6.80275297164917,loss=6.817670822143555>

evaluation = tff.learning.build_federated_evaluation(model_fn=create_tff_model)


test_metrics = evaluation(state.model, datasets)
print(test_metrics)

<mean_squared_error=27.308320999145508,loss=27.19877052307129>

我很困惑为什么当迭代过程返回一个小得多的 mse 时，训练集的 10 次迭代后评估的 mse 更高。我在这里做错了什么？是不是在 tensorflow 中 fml 的实现中隐藏了什么？有人可以向我解释吗？

标签： tensorflowmachine-learningtensorflow2.0tensorflow-federatedfederated-learning

Answer 1

您实际上在联邦学习中遇到了一个非常有趣的现象。特别是，这里需要问的问题是：如何计算训练指标？

训练指标通常在本地训练期间计算；因此，它们是在客户端拟合其本地数据时计算的；在 TFF 中，它们是在执行每个本地步骤之前计算的——这发生在前向传递调用期间。如果您想象极端情况，即仅在每个客户端的一轮培训结束时计算指标，您会清楚地看到一件事——客户端报告的指标代表了它与本地数据的匹配程度。

然而，联邦学习必须在每轮训练结束时产生一个单一的全局模型——在联邦平均中，这些局部模型在参数空间中一起平均。在一般情况下，不清楚如何直观地解释这样的步骤——参数空间中非线性模型的平均值并没有给你一个平均预测或类似的东西。

联合评估采用这个平均模型，并在每个客户端上运行本地评估，根本不拟合本地数据。因此，如果您的客户端数据集具有完全不同的分布，您应该期望从联合评估返回的指标与从一轮联合训练返回的指标完全不同——联合平均是报告在处理过程中收集的指标适应本地数据，而联合评估是在对所有这些本地训练的模型进行平均后收集的报告指标。

实际上，如果您next将迭代过程的函数和评估函数的调用交错，您将看到如下模式：

train metrics=<mean_squared_error=88.22489929199219,loss=88.6319351196289>
eval metrics=<mean_squared_error=33.69473648071289,loss=33.55160140991211>
train metrics=<mean_squared_error=8.873666763305664,loss=8.882776260375977>
eval metrics=<mean_squared_error=29.235883712768555,loss=29.13833236694336>
train metrics=<mean_squared_error=7.932246208190918,loss=7.918393611907959>
eval metrics=<mean_squared_error=27.9038028717041,loss=27.866817474365234>
train metrics=<mean_squared_error=7.573018550872803,loss=7.576478958129883>
eval metrics=<mean_squared_error=27.600923538208008,loss=27.561887741088867>
train metrics=<mean_squared_error=7.228050708770752,loss=7.224897861480713>
eval metrics=<mean_squared_error=27.46322250366211,loss=27.36537742614746>
train metrics=<mean_squared_error=7.049572944641113,loss=7.03688907623291>
eval metrics=<mean_squared_error=26.755760192871094,loss=26.719152450561523>
train metrics=<mean_squared_error=6.983217716217041,loss=6.954374313354492>
eval metrics=<mean_squared_error=26.756895065307617,loss=26.647253036499023>
train metrics=<mean_squared_error=6.909178256988525,loss=6.923810005187988>
eval metrics=<mean_squared_error=27.047882080078125,loss=26.86684799194336>
train metrics=<mean_squared_error=6.8190460205078125,loss=6.79202938079834>
eval metrics=<mean_squared_error=26.209386825561523,loss=26.10053062438965>
train metrics=<mean_squared_error=6.7200140953063965,loss=6.737307071685791>
eval metrics=<mean_squared_error=26.682661056518555,loss=26.64984703063965>

也就是说，你的联合评估也在下降，只是比你的训练指标慢得多——有效地衡量你的客户数据集的变化。您可以通过运行来验证这一点：

eval_metrics = evaluation(state.model, [datasets[0]])
print('eval metrics on 0th dataset={}'.format(eval_metrics))
eval_metrics = evaluation(state.model, [datasets[1]])
print('eval metrics on 1st dataset={}'.format(eval_metrics))
eval_metrics = evaluation(state.model, [datasets[2]])
print('eval metrics on 2nd dataset={}'.format(eval_metrics))

你会看到类似的结果

eval metrics on 0th dataset=<mean_squared_error=9.426984786987305,loss=9.431192398071289>
eval metrics on 1st dataset=<mean_squared_error=34.96992111206055,loss=34.96992492675781>
eval metrics on 2nd dataset=<mean_squared_error=72.94075775146484,loss=72.88787841796875>

所以你可以看到你的平均模型在这三个数据集中的表现有很大的不同。

最后一点：您可能会注意到您的evaluate函数的最终结果不是您的三个损失的平均值——这是因为该evaluate函数将是示例加权的，而不是客户加权的——也就是说，拥有更多数据的客户获得更多平均体重。

希望这可以帮助！