input_data = np.array(random_input, dtype=np.uint8)
benchmark_time = timeit.timeit(
lambda: engine.run_inference(input_data),
number=iterations)
# Time consumed for each iteration (milliseconds).
time_per_inference = (benchmark_time / iterations) * 1000
print(time_per_inference, 'ms (iterations = ', iterations, ')')
return time_per_inference
if __name__ == '__main__':
args = test_utils.parse_args()
machine = test_utils.machine_info()
test_utils.check_cpu_scaling_governor_status()
# Read references from csv file.
model_list, reference = test_utils.read_reference(
'basic_engine_reference_%s.csv' % machine)
total_models = len(model_list)
# Put column names in first row.
results = [('MODEL', 'INFERENCE_TIME')]
for cnt, model in enumerate(model_list, start=1):
print('-------------- Model ', cnt, '/', total_models, ' ---------------')
results.append((model, _run_benchmark_for_model(model)))
test_utils.save_as_csv(
'basic_engine_benchmarks_%s_%s.csv' % (
machine, time.strftime('%Y%m%d-%H%M%S')),
results)
test_utils.check_result(reference, results, args.enable_assertion)
learning_rate = 0.01
batch_size = 100
print('Start backprop')
start_time = time.perf_counter()
model.train_with_sgd(dataset,
num_iter,
learning_rate,
batch_size,
print_every=-1)
training_time = time.perf_counter() - start_time
print('Backprop time: ', training_time, 's')
return training_time
if __name__ == '__main__':
args = test_utils.parse_args()
machine = test_utils.machine_info()
# cases are defined by parameter pairs [num_classes, feature_dim].
cases = [[4, 256], [16, 256], [4, 1024], [16, 1024]]
results = [('CASE', 'TRAINING_TIME(s)')]
for params in cases:
num_classes = params[0]
feature_dim = params[1]
print('-------- num_classes=%d / feature_dim=%d --------' %
(num_classes, feature_dim))
results.append((":".join(str(i) for i in params),
_benchmark_for_training(num_classes, feature_dim)))
test_utils.save_as_csv(
'softmax_regression_benchmarks_%s_%s.csv' %
(machine, time.strftime('%Y%m%d-%H%M%S')), results)
print('Image pre-processing time: ', end_time - start_time, 's')
start_time = end_time
for class_id, tensors in enumerate(image_list_by_category.values()):
engine.train(tensors, class_id)
engine.save_model(output_model_path)
training_time = time.monotonic() - start_time
print('Model: ', model)
print('Data set : ', data_set)
print('Training time : ', training_time, 's')
# Remove the model.
subprocess.call(['rm', output_model_path])
return training_time
if __name__ == '__main__':
args = test_utils.parse_args()
machine = test_utils.machine_info()
models, reference = test_utils.read_reference(
'imprinting_reference_%s.csv' % machine)
model_num = len(models)
results = [('MODEL', 'DATA_SET', 'INFERENCE_TIME')]
for cnt, name in enumerate(models, start=1):
# 10 Categories, each has 20 images.
data = 'open_image_v4_subset'
print('---------------- ', cnt, '/', model_num, ' ----------------')
results.append((name, data, _benchmark_for_training(name, data)))
test_utils.save_as_csv(
'imprinting_benchmarks_%s_%s.csv' %
(machine, time.strftime('%Y%m%d-%H%M%S')), results)
test_utils.check_result(reference, results, args.enable_assertion)
engines.append(engine)
input_data_list.append(input_data)
benchmark_time = timeit.timeit(
lambda: _run_inferences(engines, input_data_list),
number=iterations)
# Time consumed for each iteration (milliseconds).
time_per_inference = (benchmark_time / iterations) * 1000
print(time_per_inference, 'ms (iterations = ', iterations, ')')
return time_per_inference
if __name__ == '__main__':
args = test_utils.parse_args()
machine = test_utils.machine_info()
test_utils.check_cpu_scaling_governor_status()
# Read references from csv file.
modelsets_list, reference = test_utils.read_reference(
'cocompilation_reference_%s.csv' % machine)
total_modelsets = len(modelsets_list)
# Put column names in first row.
results = [('MODELS', 'INFERENCE_TIME')]
for cnt, modelsets in enumerate(modelsets_list, start=1):
print('-------------- Models ', cnt, '/', total_modelsets, ' ---------------')
results.append((modelsets, _run_benchmark_for_cocompiled_models(modelsets.split(','))))
test_utils.save_as_csv(
'cocompilation_benchmarks_%s_%s.csv' % (
machine, time.strftime('%Y%m%d-%H%M%S')),
results)
test_utils.check_result(reference, results, args.enable_assertion)
