MOKT is a library for data-driven testing of OpenCL kernels. It obtains valid inputs and outputs from TensorFlow models — this way, you can easily get test data for a wide variety of machine learning operations, ranging from primitives such as ReLU and element-wise addition to whole subgraphs, e.g. ResNet's bottleneck blocks.
Install using pip:
pip install mokt
Note that only Python 3 is supported.
Check out examples to see MOKT in action.
It is recommended that you read the data extraction design note to get familiar with the way MOKT interacts with TensorFlow.
The high-level data API is used as follows:
@TestData(
tf_checkpoint_dir='/path/to/checkpoint/dir',
tf_values={'input': 'operation/name:0', 'output': 'another/op:0'})
def my_test_func(test_data):
print(type(test_data['input'])) # <class 'numpy.ndarray'>
my_test_func()Choosing the correct nodes for your tests is easier with TensorBoard, which visualizes the computational graph and shows helpful info, such as tensor shapes, operation names, etc.
Execution is performed in a TestEnvironment, which conveniently wraps host state and handles data conversion (read the class documentation for more information).
You may of course choose to write your own specialized implementation and use this library for data extraction only.
python3 com.py --some_keysKeys: --input/--first_input, --second_input*, --output, --cl_source*, --kernel_name*, --file*
'*' - optionally
If you do not specify 'cl_source' and 'kernel name' then they will be identified automatically by the output operation name
If your output ends with '.../Relu:0', then 'cl_source' will be 'relu.cl' and 'kernel_name' - 'relu'