ubuntu18.04
tensorflow-gpu1.15.2
cuda=10.2
keras==2.1.5
不得不说版本交叉问题太太太太烦了
换了系统 新建虚拟环境 配的tf==1.15 , model搭建用的keras的API 即(大家烂熟的from keras.layers import Dense, Input) 跑图片训练都没问题,后来没有切环境为了多机器多GPU分布式训练,作死tensorflow升级到2.3 又降级回了1.15 开始报各种错:
#这种
Traceback (most recent call last):
File "/home/cy/PycharmProjects/grab/train.py", line 66, in <module>
callbacks=my_callbacks)
File "/home/cy/miniconda3/envs/tf-1.15/lib/python3.7/site-packages/keras/engine/training.py", line 1705, in fit
validation_steps=validation_steps)
File "/home/cy/miniconda3/envs/tf-1.15/lib/python3.7/site-packages/keras/engine/training.py", line 1155, in _fit_loop
callbacks.set_model(callback_model)
File "/home/cy/miniconda3/envs/tf-1.15/lib/python3.7/site-packages/keras/callbacks.py", line 52, in set_model
callback.set_model(model)
File "/home/cy/miniconda3/envs/tf-1.15/lib/python3.7/site-packages/tensorflow_core/python/keras/callbacks.py", line 899, in set_model
not model._is_graph_network and # pylint: disable=protected-access
AttributeError: 'Model' object has no attribute '_is_graph_network'
#或者这种
AttributeError: module ‘tensorflow.python.framework.ops‘ has no attribute ‘_TensorLike‘
网上教程让升级或者降级keras版本,本少女觉得版本适配的空间其实很大,tensorflow 2.0以下, keras从(2.0.8–2.3.1)都是okay的。
根源在于用的是tf.keras的 API 还是 keras的API
去年惯用了:
tf<=2.0
keras.layers.Dense
keras.models.Mode
tf>=2.0
tf.keras.layers.Dense
or
tf.keras.layers.core.Dense
一直很友好
现在tf 1的版本用tf.keras 也可以。。。有点迷了
总之遇到了不要先升级降级版本,先试试换个API吧