fxmarty/20220711-h10m00s56_example_conll2003

task: token-classification
Backend: sagemaker-training
Backend args: {'instance_type': 'ml.m5.2xlarge', 'supported_instructions': 'avx512'}
Number of evaluation samples: 100

Fixed parameters:

• model_name_or_path: elastic/distilbert-base-uncased-finetuned-conll03-english
• dataset:
  • path: conll2003
  • eval_split: validation
  • data_keys: {'primary': 'tokens'}
  • ref_keys: ['ner_tags']
  • calibration_split: train
• node_exclusion: []
• per_channel: False
• calibration:
  • method: minmax
  • num_calibration_samples: 100
• framework: onnxruntime
• framework_args:
  • opset: 11
  • optimization_level: 1
• aware_training: False

Benchmarked parameters:

• quantization_approach: dynamic, static
• operators_to_quantize: ['Add', 'MatMul'], ['Add']

Evaluation

Non-time metrics

quantization_approach	operators_to_quantize		precision (original)	precision (optimized)		recall (original)	recall (optimized)		f1 (original)	f1 (optimized)		accuracy (original)	accuracy (optimized)
dynamic	['Add', 'MatMul']	|	0.974	0.974	|	0.955	0.949	|	0.964	0.962	|	0.990	0.989
dynamic	['Add']	|	0.974	0.974	|	0.955	0.955	|	0.964	0.964	|	0.990	0.990
static	['Add', 'MatMul']	|	0.974	0.081	|	0.955	0.222	|	0.964	0.118	|	0.990	0.467
static	['Add']	|	0.974	0.073	|	0.955	0.182	|	0.964	0.105	|	0.990	0.290

Time metrics

Time benchmarks were run for 3 seconds per config.

Below, time metrics for batch size = 1, input length = 64.

quantization_approach	operators_to_quantize		latency_mean (original, ms)	latency_mean (optimized, ms)		throughput (original, /s)	throughput (optimized, /s)
dynamic	['Add', 'MatMul']	|	59.35	21.91	|	17.00	45.67
dynamic	['Add']	|	59.18	29.24	|	17.00	34.33
static	['Add', 'MatMul']	|	59.25	28.31	|	17.00	35.33
static	['Add']	|	58.77	31.80	|	17.33	31.67