def _test_model(
model_path: str,
input_paths: str,
output_paths: str,
runner_constructor: Callable[[Any], ModelRunner],
rtol: float = 1e-5,
atol: float = 1e-8,
):
model = load_model(model_path)
dataloader = DataLoader(input_paths, output_paths, 2, 0)
model_runner = runner_constructor(model, batch_size=2)
outputs, _ = model_runner.run(dataloader)
_test_output(outputs, dataloader, batch_size=2)
dataloader = DataLoader(input_paths, output_paths, 1, 0)
model_runner = runner_constructor(
model,
batch_size=1,
)
outputs, _ = model_runner.run(dataloader, max_steps=1)
assert len(outputs) == 1
outputs, _ = model_runner.run(dataloader)
_test_output(outputs, dataloader)
def test_nm_analyze_model_runner(
onnx_models_with_data: OnnxModelDataFixture, ):
model = load_model(onnx_models_with_data.model_path)
# Sanity check, asserting model can run random input
dataloader = DataLoader.from_model_random(model, 5, 0, 10)
model_runner = DeepSparseAnalyzeModelRunner(model, batch_size=5)
outputs, _ = model_runner.run(dataloader, max_steps=5)
fields = [
"num_threads", "num_sockets", "average_total_time", "iteration_times"
]
layer_fields = [
"name",
"canonical_name",
"input_dims",
"output_dims",
"strides",
"required_flops",
"kernel_sparsity",
"activation_sparsity",
"average_run_time_in_ms",
"average_utilization",
"average_teraflops_per_second",
]
for out in outputs:
for field in fields:
assert field in out
for layer_info in out["layer_info"]:
for field in layer_fields:
assert field in layer_info
def _create_sensitivity_ks_data(
model_path: str,
input_paths: str,
output_paths: str,
loss_approx_path: str,
loss_one_shot_path: str,
perf_path: str,
sparsity_levels: Union[List[float], None],
):
dataloader = DataLoader(input_paths, output_paths, 1)
analysis = pruning_loss_sens_magnitude(model_path)
analysis.save_json(loss_approx_path)
analysis = pruning_loss_sens_one_shot(
model_path,
dataloader,
1,
1,
show_progress=False,
sparsity_levels=sparsity_levels,
)
analysis.save_json(loss_one_shot_path)
if deepsparse is not None:
analysis = pruning_perf_sens_one_shot(
model_path,
dataloader,
1,
None,
iterations_per_check=10,
warmup_iterations_per_check=5,
sparsity_levels=sparsity_levels,
show_progress=False,
)
analysis.save_json(perf_path)
def test_one_shot_ks_perf_sensitivity(
onnx_models_with_analysis: OnnxModelAnalysisFixture, ):
expected_analysis = PruningLossSensitivityAnalysis.load_json(
onnx_models_with_analysis.perf_path)
dataloader = DataLoader(onnx_models_with_analysis.input_paths,
onnx_models_with_analysis.output_paths, 1)
analysis = pruning_perf_sens_one_shot(
onnx_models_with_analysis.model_path,
dataloader,
1,
None,
iterations_per_check=10,
warmup_iterations_per_check=3,
sparsity_levels=onnx_models_with_analysis.sparsity_levels,
show_progress=False,
)
expected_layers = sorted(expected_analysis.dict()["results"],
key=lambda x: x["index"])
actual_layers = sorted(
analysis.dict()["results"],
key=lambda x: x["index"],
)
_test_analysis_comparison(expected_layers, actual_layers)
