Support more versions of added Operators
This issue is to discuss the fix for the ignored tests of nodes import.
The tests that are currently ignored:
-
"averagepool_3d_dilations_." && "averagepool_2d_(precomputed_)?pads_."
count_include_pads attribute cannot be supported
-> Cannot addcount_include_padsbecause padding is handled as a separate Op in paddedAvgPooling and paddedMaxPooling
See aidge_core#244, aidge_backend_cpu#46 -
"maxpool_.with_argmax_2d._strides"
storage_order != 0
-> To add support for attrstorage_orderthe second output (indices) needs to be added to MaxPool
See aidge_core#254 -
"pad/.*axes"
pad axes not supported
-> To fully support the tests of "pad", the template on DIM has to be removed -
"sum_(one|example)"
support only 2 inputs
-> Eithernb_inputsneeds to be put back in Add op or a new operator needs to be added that does Add for different input number -
"equal.*" && "^and/"
unsupported bool input
-> DataType bool needs to be added to Aidge -
"gemm_.*(transposeA|all_attr)")
supports only transA == 0 -
"gemm.*matrix_bias"
2D bias not supported -
"gemm.*(beta|alpha)"
unsuppored attribute "alpha" and "beta" -
"batchnorm_.*training_mode"
does not support training_mode -> When training_mode is true: 2 extra outputs (running_mean and running_var)
See aidge_core#254 -
"split_equal_.*opset13"
nb_outputs/split attributes -
"lstm_batchwise"
does not support layout attribute -
"lstm_with_peepholes"
does not support input4 -
"training_dropout"
unsupported bool input -
"resize*.*"
attributes unknown -
"dequantizelinear*.*"
-> All inputs are dynamic so we cannot retrieve data type for cast operator
See #63 (closed) -
"qlinearconv.*"
-> All inputs are dynamic so we cannot retrieve conv's kernel_dims from inputs when it's not provided as attr
See #63 (closed) -
"quantizelinear*.*"
-> All inputs are dynamic so we cannot retrieve data type for cast operator
See #63 (closed)