def test_sparsity_algo(test_models, tmp_path, models):
model_name, model_framework, algorithm, sparsity_level, normed_threshold, ref_name = test_models
algorithm_config = Dict({
'algorithms': [{
'name': algorithm,
'params': {
'sparsity_level': sparsity_level,
'normed_threshold': normed_threshold,
}
}]
})
model = models.get(model_name, model_framework, tmp_path)
engine_config = get_engine_config(model_name)
config = merge_configs(model.model_params, engine_config, algorithm_config)
config.engine.evaluate = False
config.engine.type = 'accuracy_checker'
_ = optimize(config)
output_dir = os.path.join(config.model.exec_log_dir, 'optimized')
xml_path = os.path.join(output_dir, config.model.model_name + '.xml')
bin_path = os.path.join(output_dir, config.model.model_name + '.bin')
output_model, meta = stdout_redirect(restore_graph_from_ir, xml_path,
bin_path)
output_model.meta_data = meta
assert check_sparsity_level(NXModel(graph=output_model), config,
sparsity_level)
check_graph(tmp_path,
output_model,
model_name + ref_name,
model_framework,
check_weights=True)
def test_sparsity(test_models, tmp_path, models):
model_name, model_framework, algorithm, sparsity_level, normed_threshold, expected_accuracy = test_models
algorithm_config = Dict({
'algorithms': [{
'name': algorithm,
'params': {
'sparsity_level': sparsity_level,
'normed_threshold': normed_threshold,
}
}]
})
if algorithm == 'WeightSparsity':
bias_config = Dict({'target_device': 'CPU', 'stat_subset_size': 300})
algorithm_config['algorithms'][0]['params'].update(bias_config)
model = models.get(model_name, model_framework, tmp_path)
engine_config = get_engine_config(model_name)
config = merge_configs(model.model_params, engine_config, algorithm_config)
config.engine.models[0].datasets[0].subsample_size = 1000
metrics = optimize(config)
output_dir = os.path.join(config.model.exec_log_dir, 'optimized')
for metric_name in metrics:
print('{}: {:.4f}'.format(metric_name, metrics[metric_name]))
assert metrics == pytest.approx(expected_accuracy, abs=0.006)
xml_path = os.path.join(output_dir, config.model.model_name + '.xml')
bin_path = os.path.join(output_dir, config.model.model_name + '.bin')
assert os.path.exists(xml_path)
assert os.path.exists(bin_path)
# Check resulting sparsity level
model, _ = stdout_redirect(restore_graph_from_ir, xml_path, bin_path)
assert check_sparsity_level(NXModel(graph=model), config, sparsity_level)