ivere27's picture
TinyReceiptVQA structured-QA 21M — split KV ONNX, FP32 and static W8A8
f536c6e
Raw History Blame Contribute Delete
2.96 kB
{
"format": "tiny_receipt_vqa_split_kv_onnx_v2",
"canonical_heldout_report": "eval/fp32_int8_heldout_summary.json",
"heldout_summary": {
"fp32": {
"address": {
"n": 1032,
"answer_exact": 0.9534883720930233,
"recomputed_n": 1032,
"recomputed_answer_exact": 0.9525193798449613,
"target_exact": 0.015503875968992248
},
"overall": {
"n": 2000,
"answer_exact": 0.9725,
"recomputed_n": 2000,
"recomputed_answer_exact": 0.972,
"target_exact": 0.4735
},
"phone_number": {
"n": 968,
"answer_exact": 0.9927685950413223,
"recomputed_n": 968,
"recomputed_answer_exact": 0.9927685950413223,
"target_exact": 0.9617768595041323
}
},
"int8": {
"address": {
"n": 1032,
"answer_exact": 0.9505813953488372,
"recomputed_n": 1032,
"recomputed_answer_exact": 0.9505813953488372,
"target_exact": 0.014534883720930232
},
"overall": {
"n": 2000,
"answer_exact": 0.971,
"recomputed_n": 2000,
"recomputed_answer_exact": 0.971,
"target_exact": 0.472
},
"phone_number": {
"n": 968,
"answer_exact": 0.9927685950413223,
"recomputed_n": 968,
"recomputed_answer_exact": 0.9927685950413223,
"target_exact": 0.9597107438016529
}
}
},
"fp32_int8_comparison": {
"n": 2000,
"prediction_exact_agreement": 0.7445,
"answer_agreement": 0.9795,
"selected_family_agreement": 1.0,
"changed_predictions": 511
},
"heldout_timing": {
"fp32": {
"encoder_seconds": 243.63899111794308,
"decoder_seconds": 29.627758088114206,
"total_seconds": 273.2667492060573,
"decoder_calls": 2953,
"providers": [
"CPUExecutionProvider"
],
"encoder": "encoder_model.onnx",
"decoder": "decoder_model.onnx"
},
"int8": {
"encoder_seconds": 195.71248534298502,
"decoder_seconds": 61.439259737788234,
"total_seconds": 257.15174508077325,
"decoder_calls": 2934,
"providers": [
"CPUExecutionProvider"
],
"encoder": "encoder_model_int8.onnx",
"decoder": "decoder_model_int8.onnx"
}
},
"runtime_benchmark": {
"method": "20 heldout VQAs, CPUExecutionProvider, one thread, one warmup and three measured repetitions per VQA; session creation and input preprocessing excluded",
"fp32": {
"report": "eval/runtime_benchmark_fp32_cpu.json",
"report_sha256": "5aab96400c0ff6892dca8b8e7eebff71f45e419485c2e557b57aa9f4907b571f",
"e2e_median_of_vqa_medians_ms": 288.50183452595957
},
"int8_w8a8": {
"report": "eval/runtime_benchmark_int8_cpu.json",
"report_sha256": "0389b525014e8af0771050a1526815b8692889ca66b8810bdf43585434d96625",
"e2e_median_of_vqa_medians_ms": 173.67218298022635
}
}
}