Hi,
Our EVM execution output is not matching with the PC TIDL execution output using the same quantized model. Not sure if this is a bug in the VH's SDK. Our model inference results are shown below (pseudo color visualization, upper left - pc result, lower left - evm result):
After performing a layer level check, we found that multiple concat, add and LeakyReLU ops have different outputs on PC and EVM. These ops are listed below, along with their corresponding onnx netron screenshots:
/model/feature_extractor/up_3/Concat_output_0
/model/cost_volume/Concat_95_output_0
/model/aggregation/fusions.0/relu_1/LeakyRelu_output_0
/model/aggregation/fusions.1/Add_3_output_0
I've also attached the inputs and outputs of each of the listed ops, in the format of *.y file generated by PC and EVM tidl execution.
and the model quantization config we used is
tidl_tools_path : /home/root/tidl_tools artifacts_folder : /home/root/metoak-demo/out/model-artifacts/metoak-mostereonet-aanet-singlergb-2/ platform : J7 version : 7.2 tensor_bits : 16 debug_level : 1 max_num_subgraphs : 16 deny_list : deny_list:layer_type : deny_list:layer_name : model_type : accuracy_level : 1 advanced_options:calibration_frames : 2 advanced_options:calibration_iterations : 5 advanced_options:output_feature_16bit_names_list : advanced_options:params_16bit_names_list : advanced_options:mixed_precision_factor : -1 advanced_options:quantization_scale_type : 4 advanced_options:high_resolution_optimization : 0 advanced_options:pre_batchnorm_fold : 1 ti_internal_nc_flag : 1601 advanced_options:activation_clipping : 1 advanced_options:weight_clipping : 1 advanced_options:bias_calibration : 1 advanced_options:add_data_convert_ops : 3 advanced_options:channel_wise_quantization : 0 advanced_options:inference_mode : 0 advanced_options:num_cores : 1
