Tensorflow 2.0 can't use GPU, something wrong in cuDNN? :Failed to get convolution algorithm. This is probably because cuDNN failed to initialize

Question

I am trying to understand and debug my code. I try to predict with a CNN model developed under tf2.0/tf.keras on GPU, but get those error messages. could someone help me to fix it?

here is my environmental configuration

enviroments:
python 3.6.8
tensorflow-gpu 2.0.0-rc0
nvidia 418.x
CUDA 10.0
cuDNN 7.6+**

and the log file,

2019-09-28 13:10:59.833892: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcublas.so.10.0
2019-09-28 13:11:00.228025: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudnn.so.7
2019-09-28 13:11:00.957534: E tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR
2019-09-28 13:11:00.963310: E tensorflow/stream_executor/cuda/cuda_dnn.cc:329] Could not create cudnn handle: CUDNN_STATUS_INTERNAL_ERROR
2019-09-28 13:11:00.963416: W tensorflow/core/common_runtime/base_collective_executor.cc:216] 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 mobilenetv2_1.00_192/Conv1/Conv2D}}]]
mobilenetv2_1.00_192/block_15_expand_BN/cond/then/_630/Const: (Const): /job:localhost/replica:0/task:0/device:GPU:0=====>GPU Available:  True
=====> 4 Physical GPUs, 1 Logical GPUs

mobilenetv2_1.00_192/block_15_expand_BN/cond/then/_630/Const_1: (Const): /job:localhost/replica:0/task:0/device:GPU:0
mobilenetv2_1.00_192/block_15_depthwise_BN/cond/then/_644/Const: (Const): /job:localhost/replica:0/task:0/device:GPU:0
mobilenetv2_1.00_192/block_15_depthwise_BN/cond/then/_644/Const_1: (Const): /job:localhost/replica:0/task:0/device:GPU:0
mobilenetv2_1.00_192/block_15_project_BN/cond/then/_658/Const: (Const): /job:localhost/replica:0/task:0/device:GPU:0
mobilenetv2_1.00_192/block_15_project_BN/cond/then/_658/Const_1: (Const): /job:localhost/replica:0/task:0/device:GPU:0
mobilenetv2_1.00_192/block_16_expand_BN/cond/then/_672/Const: (Const): /job:localhost/replica:0/task:0/device:GPU:0
mobilenetv2_1.00_192/block_16_expand_BN/cond/then/_672/Const_1: (Const): /job:localhost/replica:0/task:0/device:GPU:0
mobilenetv2_1.00_192/block_16_depthwise_BN/cond/then/_686/Const: (Const): /job:localhost/replica:0/task:0/device:GPU:0
mobilenetv2_1.00_192/block_16_depthwise_BN/cond/then/_686/Const_1: (Const): /job:localhost/replica:0/task:0/device:GPU:0
mobilenetv2_1.00_192/block_16_project_BN/cond/then/_700/Const: (Const): /job:localhost/replica:0/task:0/device:GPU:0
mobilenetv2_1.00_192/block_16_project_BN/cond/then/_700/Const_1: (Const): /job:localhost/replica:0/task:0/device:GPU:0
mobilenetv2_1.00_192/Conv_1_bn/cond/then/_714/Const: (Const): /job:localhost/replica:0/task:0/device:GPU:0
mobilenetv2_1.00_192/Conv_1_bn/cond/then/_714/Const_1: (Const): /job:localhost/replica:0/task:0/device:GPU:0
Traceback (most recent call last):
  File "NSFW_Server.py", line 162, in <module>
    model.predict(initial_tensor)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training.py", line 915, in predict
    use_multiprocessing=use_multiprocessing)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training_arrays.py", line 722, in predict
    callbacks=callbacks)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/engine/training_arrays.py", line 393, in model_iteration
    batch_outs = f(ins_batch)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/keras/backend.py", line 3625, in __call__
    outputs = self._graph_fn(*converted_inputs)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/eager/function.py", line 1081, in __call__
    return self._call_impl(args, kwargs)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/eager/function.py", line 1121, in _call_impl
    return self._call_flat(args, self.captured_inputs, cancellation_manager)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/eager/function.py", line 1224, in _call_flat
    ctx, args, cancellation_manager=cancellation_manager)
  File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/eager/function.py", line 511, in call
    ctx=ctx)
  File "/usr/local/lib/python3.6/dist-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 mobilenetv2_1.00_192/Conv1/Conv2D (defined at /usr/local/lib/python3.6/dist-packages/tensorflow_core/python/framework/ops.py:1751) ]] [Op:__inference_keras_scratch_graph_10727]

Function call stack:
keras_scratch_graph

The code

if __name__ == "__main__":

    print("=====>GPU Available: ", tf.test.is_gpu_available())
    tf.debugging.set_log_device_placement(True)

    gpus = tf.config.experimental.list_physical_devices('GPU')
    if gpus:
        try:
            # Currently, memory growth needs to be the same across GPUs

            tf.config.experimental.set_visible_devices(gpus[0], 'GPU')
            tf.config.experimental.set_memory_growth(gpus[0], True)
            logical_gpus = tf.config.experimental.list_logical_devices('GPU')
            print("=====>", len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs")
        except RuntimeError as e:
            # Memory growth must be set before GPUs have been initialized
            print(e)

    paras_path = "./paras/{}".format(int(2011))
    model = tf.keras.experimental.load_from_saved_model(paras_path)
    initial_tensor = np.zeros((1, INPUT_SHAPE, INPUT_SHAPE, 3))
    model.predict(initial_tensor)

Answer 1

You have to check that you have the right version of CUDA + CUDNN + TensorFlow (also ensure that you have all installed).

A couple of examples of running configurations are presented below(UPDATE FOR LATEST VERSIONS OF TENSORFLOW)

1. Cuda 11.3.1 + CuDNN 8.2.1.32 + TensorFlow 2.7.0

2. Cuda 11.0 + CuDNN 8.0.4 + TensorFlow 2.4.0

3. Cuda 10.1 + CuDNN 7.6.5 (normally > 7.6) + TensorFlow 2.2.0/TensorFlow 2.3.0 (TF >= 2.1 requires CUDA >=10.1)

4. Cuda 10.1 + CuDNN 7.6.5 (normally > 7.6) + TensorFlow 2.1.0 (TF >= 2.1 requires CUDA >= 10.1)

5. Cuda 10.0 + CuDNN 7.6.3 + / TensorFlow 1.13/1.14 / TensorFlow 2.0.

6. Cuda 9.0 + CuDNN 7.0.5 + TensorFlow 1.10

Usually this error appears when you have an incompatible version of TensorFlow/CuDNN installed. In my case, this appeared when I tried using an older TensorFlow with a newer version of CuDNN.

**If for some reason you get an error message like(and nothing happens afterwards) :

Relying on the driver to perform ptx compilation

Solution : Install the latest nvidia driver

[SEEMS TO BE SOLVED IN TF >= 2.5.0] (see below):

Only for Windows Users : Some late combintations of CUDA, CUDNN and TF may not work, due to a bug (a .dll extension named improperly). To handle that specific case, please consult this link: Tensorflow GPU Could not load dynamic library 'cusolver64_10.dll'; dlerror: cusolver64_10.dll not found