Based on the information above, we made a list of operators to proceed with the test.
11 operators
[0] AVERAGE_POOL_2D
[1] CONV_2D
[2] DEPTHWISE_CONV_2D
[3] SOFTMAX
[4] RESHAPE
[5] ADD
[6] MAX_POOL_2D
[7] MUL
[8] PAD
[9] MEAN
[10] CONCATENATION
We use the input circles and qparam required for q-implant as follows.
input circle
Use the circle file generated during the 'common-artifacts' build process.
(ex: Conv2D_000.circle)
qparam
There is a directory called "import" and there are folders by module in it.
Run init.py to create a qparam.
(a manual code for each operator with Python to create param.json and .npy files.)
About op-level-test of q-implant.
We want to check if we can proceed with q-implant on the operator used in the frequently used MobileNet, ResNet, and Inception models.
We converted the model to circle and checked the operator through the circledump command.
Model Source
MobileNet : ONE/runtime/contrib/TFLiteSharp/TFLiteTestApp/res/mobilenet_v1_1.0_224.tflite
ResNet : ResNet_V2_101
Inception : Inception_V3
Operators
MobileNet
ResNet
Inception
Based on the information above, we made a list of operators to proceed with the test.
We use the input circles and qparam required for q-implant as follows.
input circle Use the circle file generated during the 'common-artifacts' build process. (ex: Conv2D_000.circle)
qparam There is a directory called "import" and there are folders by module in it. Run init.py to create a qparam. (a manual code for each operator with Python to create param.json and .npy files.)