RQ2

Through experimental analysis, we classify the obtained bugs. The classification results are shown in the following table:

RQ2

Examples:
implementation bug

Conv2dTranspose

initialisation differences

[{'name': 'Transpose', 'params': {'output_shape': [0, 2, 3, 1], 'tensor_space': 4}}, {'name': 'ReduceSum', 'params': {'keep_dims': False, 'dim': 1, 'tensor_space': 4}}, {'name': 'Flatten'}, {'name': 'Dense', 'params': {'in_features': 28, 'out_features': 10}}, {'name': 'Softmax'}]

data format differences

Sigmoid + RSqrt / lambda(1/(x^2))

operator setting differences

MXNet uses unity cuda

precision

GlobalAvgPool2d

random error

When using Lambda, the output of the 1/x process is very likely to have nan/inf.

Finding 2.1. The existence of implementation differences leads to a large amount of false positives by existing testing approaches.

To reduce the false positives suffer from these implementation differences, we can consider strategies like initialization transformations, data format transformations and setting transformation for specific operators.

Page updated

Google Sites

Report abuse