def test_trace_memory(self):
MODEL_ID = "sshleifer/tiny-gpt2"
def _check_summary_is_not_empty(summary):
self.assertTrue(hasattr(summary, "sequential"))
self.assertTrue(hasattr(summary, "cumulative"))
self.assertTrue(hasattr(summary, "current"))
self.assertTrue(hasattr(summary, "total"))
with tempfile.TemporaryDirectory() as tmp_dir:
benchmark_args = TensorFlowBenchmarkArguments(
models=[MODEL_ID],
inference=True,
sequence_lengths=[8],
batch_sizes=[1],
log_filename=os.path.join(tmp_dir, "log.txt"),
log_print=True,
trace_memory_line_by_line=True,
eager_mode=True,
multi_process=False,
)
benchmark = TensorFlowBenchmark(benchmark_args)
result = benchmark.run()
_check_summary_is_not_empty(result.inference_summary)
self.assertTrue(Path(os.path.join(tmp_dir, "log.txt")).exists())
def test_inference_no_configs_graph(self):
MODEL_ID = "sshleifer/tiny-gpt2"
benchmark_args = TensorFlowBenchmarkArguments(
models=[MODEL_ID],
training=False,
inference=True,
sequence_lengths=[8],
batch_sizes=[1],
multi_process=False,
)
benchmark = TensorFlowBenchmark(benchmark_args)
results = benchmark.run()
self.check_results_dict_not_empty(results.time_inference_result)
self.check_results_dict_not_empty(results.memory_inference_result)
def test_inference_no_configs_only_pretrain(self):
MODEL_ID = "sshleifer/tiny-distilbert-base-uncased-finetuned-sst-2-english"
benchmark_args = TensorFlowBenchmarkArguments(
models=[MODEL_ID],
training=False,
inference=True,
sequence_lengths=[8],
batch_sizes=[1],
multi_process=False,
only_pretrain_model=True,
)
benchmark = TensorFlowBenchmark(benchmark_args)
results = benchmark.run()
self.check_results_dict_not_empty(results.time_inference_result)
self.check_results_dict_not_empty(results.memory_inference_result)
def test_inference_encoder_decoder_with_configs(self):
MODEL_ID = "patrickvonplaten/t5-tiny-random"
config = AutoConfig.from_pretrained(MODEL_ID)
benchmark_args = TensorFlowBenchmarkArguments(
models=[MODEL_ID],
training=False,
inference=True,
sequence_lengths=[8],
batch_sizes=[1],
multi_process=False,
)
benchmark = TensorFlowBenchmark(benchmark_args, configs=[config])
results = benchmark.run()
self.check_results_dict_not_empty(results.time_inference_result)
self.check_results_dict_not_empty(results.memory_inference_result)
def test_train_with_configs(self):
MODEL_ID = "sshleifer/tiny-gpt2"
config = AutoConfig.from_pretrained(MODEL_ID)
benchmark_args = TensorFlowBenchmarkArguments(
models=[MODEL_ID],
training=True,
inference=False,
sequence_lengths=[8],
batch_sizes=[1],
multi_process=False,
)
benchmark = TensorFlowBenchmark(benchmark_args, [config])
results = benchmark.run()
self.check_results_dict_not_empty(results.time_train_result)
self.check_results_dict_not_empty(results.memory_train_result)
def test_save_csv_files(self):
MODEL_ID = "sshleifer/tiny-gpt2"
with tempfile.TemporaryDirectory() as tmp_dir:
benchmark_args = TensorFlowBenchmarkArguments(
models=[MODEL_ID],
inference=True,
save_to_csv=True,
sequence_lengths=[8],
batch_sizes=[1],
inference_time_csv_file=os.path.join(tmp_dir, "inf_time.csv"),
inference_memory_csv_file=os.path.join(tmp_dir, "inf_mem.csv"),
env_info_csv_file=os.path.join(tmp_dir, "env.csv"),
multi_process=False,
)
benchmark = TensorFlowBenchmark(benchmark_args)
benchmark.run()
self.assertTrue(Path(os.path.join(tmp_dir, "inf_time.csv")).exists())
self.assertTrue(Path(os.path.join(tmp_dir, "inf_mem.csv")).exists())
self.assertTrue(Path(os.path.join(tmp_dir, "env.csv")).exists())
def main():
parser = HfArgumentParser(TensorFlowBenchmarkArguments)
benchmark_args = parser.parse_args_into_dataclasses()[0]
benchmark = TensorFlowBenchmark(args=benchmark_args)
try:
benchmark_args = parser.parse_args_into_dataclasses()[0]
except ValueError as e:
arg_error_msg = "Arg --no_{0} is no longer used, please use --no-{0} instead."
begin_error_msg = " ".join(str(e).split(" ")[:-1])
full_error_msg = ""
depreciated_args = eval(str(e).split(" ")[-1])
wrong_args = []
for arg in depreciated_args:
# arg[2:] removes '--'
if arg[2:] in TensorFlowBenchmark.deprecated_args:
# arg[5:] removes '--no_'
full_error_msg += arg_error_msg.format(arg[5:])
else:
wrong_args.append(arg)
if len(wrong_args) > 0:
full_error_msg = full_error_msg + begin_error_msg + str(wrong_args)
raise ValueError(full_error_msg)
benchmark.run()
def main():
parser = HfArgumentParser(TensorFlowBenchmarkArguments)
benchmark_args = parser.parse_args_into_dataclasses()[0]
benchmark = TensorFlowBenchmark(args=benchmark_args)
benchmark.run()
