def infer_perf_pb(pb_model_file, val_data, inputs=["x:0"], outputs=["Identity:0"]):
x_test, y_test, label_test = val_data
q_model = alexnet.load_pb(pb_model_file)
concrete_function = get_concrete_function(graph_def=q_model.as_graph_def(),
inputs=inputs,
outputs=outputs,
print_graph=True)
bt = time.time()
_frozen_graph_predictions = concrete_function(x=tf.constant(x_test))
et = time.time()
accuracy = calc_accuracy(_frozen_graph_predictions[0], label_test)
print('accuracy:', accuracy)
throughput = x_test.shape[0] / (et - bt)
print('max throughput(fps):', throughput)
# latency when BS=1
times = 1000
single_test = x_test[:1]
bt = 0
warmup = 20
for i in range(times):
if i == warmup:
bt = time.time()
_frozen_graph_predictions = concrete_function(x=tf.constant(single_test))
et = time.time()
latency = (et - bt) * 1000 / (times - warmup)
print('latency(ms):', latency)
return accuracy, throughput, latency
def auto_tune(input_graph_path, yaml_config, batch_size):
fp32_graph = alexnet.load_pb(input_graph_path)
quan = inc.Quantization(yaml_config)
dataloader = Dataloader(batch_size)
q_model = quan(fp32_graph,
q_dataloader=dataloader,
eval_func=None,
eval_dataloader=dataloader)
return q_model
def auto_tune(input_graph_path, yaml_config, batch_size):
fp32_graph = alexnet.load_pb(input_graph_path)
tuner = ilit.Tuner(yaml_config)
dataloader = Dataloader(batch_size)
q_model = tuner.tune(
fp32_graph,
q_dataloader=dataloader,
eval_func=None,
eval_dataloader=dataloader)
return q_model