eval_sampler = SequentialSampler(eval_dataset)
eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, \
batch_size=args.eval_batch_size)
def eval_func(model):
return evaluate_onnxrt(args, model, tokenizer, eval_dataloader)
if args.benchmark:
model = onnx.load(args.model_path)
from lpot.experimental.data.datasets.dummy_dataset import DummyDataset
from lpot.experimental.data.dataloaders.onnxrt_dataloader import ONNXRTDataLoader
shapes, lows, highs = parse_dummy_input(model, args.benchmark_nums,
args.max_seq_length)
dummy_dataset = DummyDataset(shapes,
low=lows,
high=highs,
dtype="int64")
dummy_dataloader = ONNXRTDataLoader(dummy_dataset)
print(
'---------------------------------------------------------------')
if args.accuracy_only:
results = evaluate_onnxrt(args, model, tokenizer, eval_dataloader)
print("Accuracy: %.5f" % results)
else:
results = evaluate_onnxrt(args,
model,
tokenizer,
dummy_dataloader,
benchmark=True)
latency = np.array(results).mean() / args.eval_batch_size
help='Shapes of input to model, like 3x224x224 or 128x768, 128x256')
args = parser.parse_args()
model = onnx.load(args.model_path)
shapes, lows, highs, dtypes = parse_dummy_input(model, args.benchmark_nums)
if args.input_shape:
input_shape = args.input_shape.replace(' ', '')
input_shapes = input_shape.split(',')
input_shapes = [input_shapes] if type(input_shapes)!=list else input_shapes
input_shapes = [shape.split('x') for shape in input_shapes]
shapes = [tuple([args.benchmark_nums] + [int(dim) for dim in shape]) for shape in input_shapes]
from lpot.experimental.data.datasets.dummy_dataset import DummyDataset
from lpot.experimental.data.dataloaders.onnxrt_dataloader import ONNXRTDataLoader
dummy_dataset = DummyDataset(shapes, low=lows, high=highs, dtype=dtypes, label=True)
dummy_dataloader = ONNXRTDataLoader(dummy_dataset, batch_size=args.eval_batch_size)
def eval_func(model):
return evaluate_onnxrt(model, dummy_dataloader, reference)
if args.benchmark:
from lpot.experimental import Benchmark, common
evaluator = Benchmark(args.config)
evaluator.model = common.Model(model)
evaluator.b_dataloader = dummy_dataloader
results = evaluator()
for mode, result in results.items():
acc, batch_size, result_list = result
latency = np.array(result_list).mean() / batch_size