File size: 5,778 Bytes

ed3aeeb

model_id,task,source_quantization_scheme,source_weight_dtype,source_activation_dtype,source_input_dtype,source_output_dtype,onnx_preservation,onnx_actual_representation,mlir_preservation,mlir_actual_representation,mlir_float_fallback
AD01,anomaly_detection,FULL_INTEGER static PTQ performed upstream; published artifact consumed unchanged; TFLITE_BUILTINS_INT8 with int8 I/O,int8 weights; int32 biases,int8,int8; scale=0.3910152316093445; zero_point=89,int8; scale=0.36449846625328064; zero_point=96,PASS,NOT_RETAINED,FAIL,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,TRUE
LM04,language_modeling/text_classification,dynamic ONNX quantization; MatMulInteger/DynamicQuantizeLinear with quantized embeddings,INT8 MatMul weights; UINT8 embedding tables,dynamic UINT8 on selected MatMul paths plus FLOAT remainder,INT64 input_ids/attention_mask/token_type_ids,FLOAT single logit,PASS,DYNAMIC_QOPERATOR,FAIL,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,TRUE
OD06,object_detection,FULL_INTEGER core with dequantized float raw-head outputs,int8 (plus int32 bias),int8,uint8; scale=0.0078125; zero_point=127,float32 class scores and box regressions,PASS,NOT_RETAINED,NOT_ESTABLISHED,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,UNKNOWN_NOT_REACHED
OD07,object_detection,FULL_INTEGER core with dequantized float raw-head outputs,int8 (plus int32 bias),int8,uint8; scale=0.0078125; zero_point=127,float32 class scores and box regressions,PASS,NOT_RETAINED,NOT_ESTABLISHED,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,UNKNOWN_NOT_REACHED
SG06,semantic_segmentation,FULL_INTEGER,uint8 (plus int32 bias),uint8,uint8; scale=0.0078125; zero_point=128,int64 category IDs,PASS,NOT_RETAINED,FAIL,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,TRUE
SG07,semantic_segmentation,FULL_INTEGER,uint8 (plus int32 bias),uint8,uint8; scale=0.0078125; zero_point=128,int64 category IDs,PASS,NOT_RETAINED,FAIL,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,TRUE
SG08,semantic_segmentation,FULL_INTEGER,uint8 (plus int32 bias),uint8,uint8; scale=0.0078125; zero_point=128,int64 category IDs (paired FP32 GraphDef outputs int32 category IDs),PASS,NOT_RETAINED,FAIL,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,TRUE
SP01,keyword spotting,full-integer affine TFLite; upstream PTQ artifact used unchanged,int8 (bias int32),int8,int8; scale=0.5847029089927673; zero_point=83,int8; scale=0.00390625; zero_point=-128,PASS,NOT_RETAINED,NOT_ESTABLISHED,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,UNKNOWN_NOT_REACHED
SP02,keyword spotting,full-integer affine TFLite public artifact; upstream training/conversion may use QAT and representative calibration,int8 (bias int32),int8,int8; scale=0.003701042616739869; zero_point=-128,int8; scale=0.00390625; zero_point=-128,PASS,NOT_RETAINED,NOT_ESTABLISHED,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,UNKNOWN_NOT_REACHED
SP08,keyword spotting,full-integer affine TFLite; README says strictly int8 but measured legacy artifact tensors are uint8/int32,uint8 (bias int32),uint8,uint8; scale=0.10196070373058319; zero_point=0,uint8; scale=0.00390625; zero_point=0,PASS,QDQ,FAIL,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,TRUE
SP09,keyword spotting,clustered (32 clusters; kmeans++) retrained FP32 then upstream post-training full-integer quantization,int8 (bias int32),int8,int8; scale=1.1071635484695435; zero_point=95,int8; scale=0.00390625; zero_point=-128,PASS,NOT_RETAINED,NOT_ESTABLISHED,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,UNKNOWN_NOT_REACHED
VC01,vision_classification,upstream static full-integer TFLite,int8,int8,int8 scale=0.003921568859 zp=-128,int8 scale=0.00390625 zp=-128,PASS,NOT_RETAINED,NOT_ESTABLISHED,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,UNKNOWN_NOT_REACHED
VC02,vision_classification,upstream static full-integer TFLite,int8,int8,int8 scale=1.0 zp=-128,int8 scale=0.00390625 zp=-128,PASS,NOT_RETAINED,NOT_ESTABLISHED,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,UNKNOWN_NOT_REACHED
VC03,vision_classification,upstream quantization-aware training/FakeQuant then fully-quantized TFLite,uint8,uint8/int32 accumulator,uint8 scale=0.0078125 zp=128,uint8 scale=0.00390625 zp=0,PASS,NOT_RETAINED,FAIL,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,TRUE
VC04,vision_classification,upstream quantization-aware training/FakeQuant then fully-quantized TFLite,uint8,uint8/int32 accumulator,uint8 scale=0.0078125 zp=128,uint8 scale=0.00390625 zp=0,PASS,NOT_RETAINED,FAIL,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,TRUE
VC05,vision_classification,upstream static INT8 TFLite,int8,int8 with int32 accumulators,uint8 scale=0.007843137719 zp=127,float32,PASS,NOT_RETAINED,NOT_ESTABLISHED,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,UNKNOWN_NOT_REACHED
VC06,vision_classification,upstream static QDQ INT8 ONNX,int8,int8 represented by Q/DQ; float boundaries,float32,float32,PASS,NOT_RETAINED,NOT_ESTABLISHED,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,UNKNOWN_NOT_REACHED
VC09,vision_classification,upstream static PTQ via Intel Neural Compressor/ONNX Runtime,int8,uint8/int8 internal; float32 boundaries,float32,float32,PASS,NOT_RETAINED,NOT_ESTABLISHED,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,UNKNOWN_NOT_REACHED
VC11,vision_classification,static full-integer TFLite distributed by Google,int8,int8/uint8 with int32 accumulators,uint8 (runtime quantization parameters recorded in JSON),uint8,PASS,NOT_RETAINED,NOT_ESTABLISHED,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,UNKNOWN_NOT_REACHED
VC12,vision_classification,upstream static PTQ via Intel Neural Compressor/ONNX Runtime,int8,uint8/int8 internal; float32 boundaries,float32,float32,PASS,NOT_RETAINED,NOT_ESTABLISHED,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,UNKNOWN_NOT_REACHED
VC13,vision_classification,upstream static PTQ via Intel Neural Compressor/ONNX Runtime,int8,uint8/int8 internal; float32 boundaries,float32,float32,PASS,MIXED_QOPERATOR_QDQ,NOT_ESTABLISHED,LOW_LEVEL_INTEGER_COMPUTE_NOT_PROVEN,UNKNOWN_NOT_REACHED