def test_onnx_quantize_acc(cfg, rec_val, batch_size=1, original=False):
qconfig = qtz.qconfig(
skip_conv_layers=[0],
skip_dense_layer=False,
nbit_input=cfg.nbit_input,
nbit_weight=cfg.nbit_input,
dtype_input=cfg.dtype_input,
dtype_weight=cfg.dtype_input,
dtype_activation=cfg.dtype_output,
debug_enabled_ops=None,
calibrate_mode="percentile",
calibrate_chunk_by=8,
)
dataset = list(calibrate_dataset(cfg.model, rec_val, batch_size, 64))
model, logfile = get_onnx_model(cfg.model,
batch_size,
qconfig,
tvm.target.cuda(),
original=original,
dataset=dataset)
val_data, batch_fn = get_val_data(cfg.model,
rec_val=rec_val,
batch_size=batch_size)
with tvm.autotvm.apply_history_best(logfile):
acc = eval_acc(model, val_data, batch_fn, log_interval=1000)
assert acc > cfg.expected_acc
return acc
def calibrate_dataset(model_name, rec_val, batch_size, calibration_samples):
val_data, _ = get_val_data(model_name, rec_val=rec_val, batch_size=batch_size)
val_data.reset()
for i, batch in enumerate(val_data):
if i * batch_size >= calibration_samples:
break
data = batch.data[0].asnumpy()
yield {"data": data}