Tensorflow 2.1 无法获取卷积算法。这可能是因为 cuDNN 初始化失败

Question

我正在使用 anaconda python 3.7 和tensorflow 2.1以及cuda 10.1和cudnn 7.6.5，并尝试运行retinaset（https://github.com/fizyr/keras-retinanet）：

python keras_retinanet/bin/train.py --freeze-backbone --random-transform --batch-size 8 --steps 500 --epochs 10 csv annotations.csv classes.csv

下面是由此产生的错误：

Epoch 1/10
2020-02-10 20:34:37.807590: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudnn64_7.dll
2020-02-10 20:34:38.835777: E tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR
2020-02-10 20:34:39.753051: E tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR
2020-02-10 20:34:39.776706: W tensorflow/core/common_runtime/base_collective_executor.cc:217] BaseCollectiveExecutor::StartAbort Unknown: Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.
         [[{{node conv1/convolution}}]]
Traceback (most recent call last):
  File "keras_retinanet/bin/train.py", line 530, in <module>
    main()
  File "keras_retinanet/bin/train.py", line 525, in main
    initial_epoch=args.initial_epoch
  File "C:\Anaconda\Anaconda3.7\lib\site-packages\keras\legacy\interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)
  File "C:\Anaconda\Anaconda3.7\lib\site-packages\keras\engine\training.py", line 1732, in fit_generator
    initial_epoch=initial_epoch)
  File "C:\Anaconda\Anaconda3.7\lib\site-packages\keras\engine\training_generator.py", line 220, in fit_generator
    reset_metrics=False)
  File "C:\Anaconda\Anaconda3.7\lib\site-packages\keras\engine\training.py", line 1514, in train_on_batch
    outputs = self.train_function(ins)
  File "C:\Anaconda\Anaconda3.7\lib\site-packages\tensorflow_core\python\keras\backend.py", line 3727, in __call__
    outputs = self._graph_fn(*converted_inputs)
  File "C:\Anaconda\Anaconda3.7\lib\site-packages\tensorflow_core\python\eager\function.py", line 1551, in __call__
    return self._call_impl(args, kwargs)
  File "C:\Anaconda\Anaconda3.7\lib\site-packages\tensorflow_core\python\eager\function.py", line 1591, in _call_impl
    return self._call_flat(args, self.captured_inputs, cancellation_manager)
  File "C:\Anaconda\Anaconda3.7\lib\site-packages\tensorflow_core\python\eager\function.py", line 1692, in _call_flat
    ctx, args, cancellation_manager=cancellation_manager))
  File "C:\Anaconda\Anaconda3.7\lib\site-packages\tensorflow_core\python\eager\function.py", line 545, in call
    ctx=ctx)
  File "C:\Anaconda\Anaconda3.7\lib\site-packages\tensorflow_core\python\eager\execute.py", line 67, in quick_execute
    six.raise_from(core._status_to_exception(e.code, message), None)
  File "<string>", line 3, in raise_from
tensorflow.python.framework.errors_impl.UnknownError:  Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.
         [[node conv1/convolution (defined at C:\Anaconda\Anaconda3.7\lib\site-packages\keras\backend\tensorflow_backend.py:3009) ]] [Op:__inference_keras_scratch_graph_12376]

Function call stack:
keras_scratch_graph

有人遇到过类似的问题吗？

Answer 1

当我尝试使用tf.distribute.MirroredStrategy(). 我现在找到了一个解决方法，允许我同时使用它们（尽管在单个 GPU 上进行训练效果很好）。尝试将以下内容放在应用程序的开头：

config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = True
session =tf.compat.v1.InteractiveSession(config=config)

希望有帮助！