keras 构建模型很简单,上手很方便,同时又是 tensorflow 的高级 API,所以学学也挺好。
模型复现在我们的实验中也挺重要的,跑出了一个模型,虽然我们可以将模型的 checkpoint 保存,但再跑一遍,怎么都得不到相同的结果。
用 keras 实现模型,想要能够复现,首先需要设置各个可能的随机过程的 seed,如 np.random.seed(1),然后代码不要在 GPU 上跑,而是限制在 CPU 上跑。
(当使用 conv2D 层时,似乎在 GPU 上跑没法复现,即使设置 batch_size=1,只在 CPU 上跑才能复现。)
我的 tensorflow+keras 版本:
print(tf.VERSION) # '1.10.0'
print(tf.keras.__version__) # '2.1.6-tf'
keras 模型可复现的配置:
import numpy as np
import tensorflow as tf
import random as rn
import os
# run on CPU only, if you want to run code on GPU, you should delete the following line.
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
os.environ["PYTHONHASHSEED"] = '0'
# The below is necessary for starting Numpy generated random numbers
# in a well-defined initial state.
np.random.seed(42)
# The below is necessary for starting core Python generated random numbers
# in a well-defined state.
rn.seed(12345)
# Force TensorFlow to use single thread.
# Multiple threads are a potential source of non-reproducible results.
# For further details, see: https://stackoverflow.com/questions/42022950/
session_conf = tf.ConfigProto(intra_op_parallelism_threads=1,
inter_op_parallelism_threads=1)
from keras import backend as K
# The below tf.set_random_seed() will make random number generation
# in the TensorFlow backend have a well-defined initial state.
# For further details, see:
# https://www.tensorflow.org/api_docs/python/tf/set_random_seed
tf.set_random_seed(1234)
sess = tf.Session(graph=tf.get_default_graph(), config=session_conf)
K.set_session(sess)
# Rest of code follows ...
对于 tensorflow low-level API,即用 tf.variable_scope() 和 tf.get_variable() 自行构建 layers,同样会出现这种问题。
keras 文档 对此的解释是:
而 pytorch 是怎么保证可复现:(cudnn中对卷积操作进行了优化,牺牲了精度来换取计算效率。可以看到,下面的代码强制 cudnn 产生确定性的结果,但会牺牲效率。具体参见博客 PyTorch的可重复性问题 (如何使实验结果可复现))
from torch.backends import cudnn
cudnn.benchmark = False # if benchmark=True, deterministic will be False
cudnn.deterministic = True
References
How can I obtain reproducible results using Keras during development? -- Keras Documentation
具有Tensorflow后端的Keras可以随意使用CPU或GPU吗?
PyTorch的可重复性问题 (如何使实验结果可复现)-- hyk_1996