联合张量流的训练和评估期间的MSE错误有所不同

时间：2019-12-05 10:27:07

标签： tensorflow machine-learning tensorflow2.0 federated tensorflow-federated

我正在联邦中的张量流中实现回归模型。我从本教程中使用的用于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合学习。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的mse较小时，训练集的10 ms的评估mse会更高。我在这里做错了什么？在tensorflow的fml的实现中隐藏了什么吗？有人可以向我解释吗？

1 个答案:

答案 0 :(得分：4)

您实际上在联盟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？

培训指标通常是在本地培训期间计算的；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-在here发生在前向传递调用期间。如果您设想一种极端情况，即仅在每个客户的一轮培训的结束时计算指标，您会清楚地看到一件事-客户正在报告代表的指标他的本地数据拟合得如何。

但是，

联合学习必须在每轮培训结束时生成一个单一的全局模型-在联合平均中，这些局部模型在参数空间中被平均在一起。在一般情况下，尚不清楚如何直观地解释这一步骤-参数空间中非线性模型的平均值无法提供平均预测或类似的结果。

联合评估采用这种平均模型，并在每个客户端上进行本地评估，而根本不适应本地数据。因此，如果您处于客户数据集分布完全不同的情况下，则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-联合平均报告的指标是在期间收集的适应本地数据的过程，而联合评估则是报告将所有这些本地训练的模型平均在一起后收集的指标。

实际上，如果将对迭代过程的next函数和评估函数的调用进行交错，则将看到以下模式：

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，您的联合评估也在下降，仅比您的训练指标慢得多-有效地衡量了客户数据集中的差异。您可以通过运行以下命令进行验证：

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))

，您将看到类似的结果

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>

因此您可以看到，在这三个数据集中，平均模型的性能差异很大。

最后一点：您可能会注意到evaluate函数的最终结果不是三个损失的平均值，这是因为evaluate函数会是示例加权的，而不是客户端加权的-也就是说，拥有更多数据的客户端的平均权重更大。

希望这会有所帮助！

相关问题
最新问题