lstm - CuDNNLSTM : UnknownError: Fail to find the dnn implementation

Question

我已经成功运行了以 LSTM 作为第一层的模型。但出于好奇，我用 CuDNNLSTM 替换了 LSTM。
但是在model.fit之后，它回复了以下错误信息:

UnknownError: Fail to find the dnn implementation.
    [[{{node cu_dnnlstm_5/CudnnRNN}} = CudnnRNN[T=DT_FLOAT, _class=["loc:@training_2/Adam/gradients/cu_dnnlstm_5/CudnnRNN_grad/CudnnRNNBackprop"], direction="unidirectional", dropout=0, input_mode="linear_input", is_training=true, rnn_mode="lstm", seed=87654321, seed2=0, _device="/job:localhost/replica:0/task:0/device:GPU:0"](cu_dnnlstm_5/transpose, cu_dnnlstm_5/ExpandDims_1, cu_dnnlstm_5/ExpandDims_1, cu_dnnlstm_5/concat_1)]]
    [[{{node metrics_3/mean_squared_error/Mean_1/_1877}} = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device_incarnation=1, tensor_name="edge_4852_metrics_3/mean_squared_error/Mean_1", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:CPU:0"]()]]

我已经在这个 discussion 上尝试过 TestCudnnLSTM()并成功通过测试:

Keras 版本:2.2.4
Tensorflow 版本:1.12.0
创建模型
_______________________________________________________
层(类型)输出形状参数#
================================================== ================
cu_dnnlstm_1 (CuDNNLSTM) (无, 1000, 1) 16
================================================== ================
总参数:16
可训练参数:16
不可训练的参数:0
_______________________________________________________
没有任何
模型编译

看来问题是在模型拟合的时候出现的。但我不知道到底是什么问题？

Answer 1

对于 TensorFlow v2，一种解决方案是 -

import tensorflow as tf
physical_devices = tf.config.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(physical_devices[0], enable=True)

然后你也可以使用 keras 模型 -

from tensorflow.keras.models import Model

Documentation

这个解决方案对我有用，它只支持一个 GPU 的内存增长。