The Conv2D layer in tensorflow is not working when GPU is turned on. The operating system is Windows (Windows Server 2019 Datacenter). The size is Standard NV6_Promo (6 vcpus, 56 GiB memory) in East US 2.

This means one cannot use Conv2D layers in tensorflow in training for image classification. Note that tensorflow Dense layers work with GPU.

I have created a simple stand-alone example.

tensorflow_test_Conv2D.py

compute application of Conv2D layer on tensorflow tensor

if following lines are uncommented, then code turns off GPU

if following lines are commented, then code uses GPU

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"

import tensorflow as tf

if name == "main":
# create tensorflow object start with 2d object and expand dims to get 4d object
x = tf.constant([[1.0, 2.02, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0]])
x = tf.expand_dims(x, axis=0)
x = tf.expand_dims(x, axis=3)
print("input tensor shape: {} and value \n{}".format(x.shape, x))
# apply convolutional layer
layer = tf.keras.layers.Conv2D(1, (2, 2), input_shape=(3, 3, 1))
y = layer(x)
print("Conv2D(x) shape: {} and value \n{}".format(y.shape, y))

Program runs when GPU turned off (uncomment the os lines). Program fails when GPU turned on. Here is error message:

2021-07-15 04:49:36.851512: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library cudnn64_8.dll
2021-07-15 04:49:37.199399: E tensorflow/stream_executor/cuda/cuda_dnn.cc:352] Loaded runtime CuDNN library: 8.0.5 but source was compiled with: 8.1.0. CuDNN library needs to have matching major version and equal or higher minor version. If using a binary install, upgrade your CuDNN library. If building from sources, make sure the library loaded at runtime is compatible with the version specified during compile configuration.
2021-07-15 04:49:37.215405: E tensorflow/stream_executor/cuda/cuda_dnn.cc:352] Loaded runtime CuDNN library: 8.0.5 but source was compiled with: 8.1.0. CuDNN library needs to have matching major version and equal or higher minor version. If using a binary install, upgrade your CuDNN library. If building from sources, make sure the library loaded at runtime is compatible with the version specified during compile configuration.
Traceback (most recent call last):
File "tensorflow_test_Conv2D.py", line 19, in <module>
y = layer(x)
File "c:\Miniconda\envs\py38_tensorflow\lib\site-packages\tensorflow\python\keras\engine\base_layer.py", line 1030, in call
outputs = call_fn(inputs, *args, **kwargs)
File "c:\Miniconda\envs\py38_tensorflow\lib\site-packages\tensorflow\python\keras\layers\convolutional.py", line 249, in call
outputs = self._convolution_op(inputs, self.kernel)
File "c:\Miniconda\envs\py38_tensorflow\lib\site-packages\tensorflow\python\util\dispatch.py", line 206, in wrapper
return target(*args, **kwargs)
File "c:\Miniconda\envs\py38_tensorflow\lib\site-packages\tensorflow\python\ops\nn_ops.py", line 1012, in convolution_v2
return convolution_internal(
File "c:\Miniconda\envs\py38_tensorflow\lib\site-packages\tensorflow\python\ops\nn_ops.py", line 1142, in convolution_internal
return op(
File "c:\Miniconda\envs\py38_tensorflow\lib\site-packages\tensorflow\python\ops\nn_ops.py", line 2596, in _conv2d_expanded_batch
return gen_nn_ops.conv2d(
File "c:\Miniconda\envs\py38_tensorflow\lib\site-packages\tensorflow\python\ops\gen_nn_ops.py", line 931, in conv2d
_ops.raise_from_not_ok_status(e, name)
File "c:\Miniconda\envs\py38_tensorflow\lib\site-packages\tensorflow\python\framework\ops.py", line 6897, in raise_from_not_ok_status
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. [Op:Conv2D]