SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Benchmarks (#4912)

Browse Source 
* finish benchmark

* fix isort

* fix setup cfg

* retab

* fix time measuring of tf graph mode

* fix tf cuda

* clean code

* better error message
This commit is contained in:
Patrick von Platen
2020-06-22 12:06:56 +02:00
committed by GitHub
parent 18a0150bfa
commit fa0be6d761
18 changed files with 1040 additions and 363 deletions
Download Patch File Download Diff File Expand all files Collapse all files

87
tests/test_benchmark.py
View File

@@ -5,7 +5,7 @@ from pathlib import Path
from transformers import AutoConfig, is_torch_available
from .utils import require_torch
from .utils import require_torch, torch_device
if is_torch_available():
@@ -26,7 +26,12 @@ class BenchmarkTest(unittest.TestCase):
def test_inference_no_configs(self):
MODEL_ID = "sshleifer/tiny-gpt2"
benchmark_args = PyTorchBenchmarkArguments(
models=[MODEL_ID], training=False, no_inference=False, sequence_lengths=[8], batch_sizes=[1]
models=[MODEL_ID],
training=False,
no_inference=False,
sequence_lengths=[8],
batch_sizes=[1],
no_multi_process=True,
)
benchmark = PyTorchBenchmark(benchmark_args)
results = benchmark.run()
@@ -42,6 +47,24 @@ class BenchmarkTest(unittest.TestCase):
torchscript=True,
sequence_lengths=[8],
batch_sizes=[1],
no_multi_process=True,
)
benchmark = PyTorchBenchmark(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)
@unittest.skipIf(torch_device == "cpu", "Cant do half precision")
def test_inference_fp16(self):
MODEL_ID = "sshleifer/tiny-gpt2"
benchmark_args = PyTorchBenchmarkArguments(
models=[MODEL_ID],
training=False,
no_inference=False,
fp16=True,
sequence_lengths=[8],
batch_sizes=[1],
no_multi_process=True,
)
benchmark = PyTorchBenchmark(benchmark_args)
results = benchmark.run()
@@ -51,7 +74,29 @@ class BenchmarkTest(unittest.TestCase):
def test_train_no_configs(self):
MODEL_ID = "sshleifer/tiny-gpt2"
benchmark_args = PyTorchBenchmarkArguments(
models=[MODEL_ID], training=True, no_inference=True, sequence_lengths=[8], batch_sizes=[1]
models=[MODEL_ID],
training=True,
no_inference=True,
sequence_lengths=[8],
batch_sizes=[1],
no_multi_process=True,
)
benchmark = PyTorchBenchmark(benchmark_args)
results = benchmark.run()
self.check_results_dict_not_empty(results.time_train_result)
self.check_results_dict_not_empty(results.memory_train_result)
@unittest.skipIf(torch_device == "cpu", "Cant do half precision")
def test_train_no_configs_fp16(self):
MODEL_ID = "sshleifer/tiny-gpt2"
benchmark_args = PyTorchBenchmarkArguments(
models=[MODEL_ID],
training=True,
no_inference=True,
sequence_lengths=[8],
batch_sizes=[1],
fp16=True,
no_multi_process=True,
)
benchmark = PyTorchBenchmark(benchmark_args)
results = benchmark.run()
@@ -62,7 +107,12 @@ class BenchmarkTest(unittest.TestCase):
MODEL_ID = "sshleifer/tiny-gpt2"
config = AutoConfig.from_pretrained(MODEL_ID)
benchmark_args = PyTorchBenchmarkArguments(
models=[MODEL_ID], training=False, no_inference=False, sequence_lengths=[8], batch_sizes=[1]
models=[MODEL_ID],
training=False,
no_inference=False,
sequence_lengths=[8],
batch_sizes=[1],
no_multi_process=True,
)
benchmark = PyTorchBenchmark(benchmark_args, configs=[config])
results = benchmark.run()
@@ -73,7 +123,12 @@ class BenchmarkTest(unittest.TestCase):
MODEL_ID = "sshleifer/tinier_bart"
config = AutoConfig.from_pretrained(MODEL_ID)
benchmark_args = PyTorchBenchmarkArguments(
models=[MODEL_ID], training=False, no_inference=False, sequence_lengths=[8], batch_sizes=[1]
models=[MODEL_ID],
training=False,
no_inference=False,
sequence_lengths=[8],
batch_sizes=[1],
no_multi_process=True,
)
benchmark = PyTorchBenchmark(benchmark_args, configs=[config])
results = benchmark.run()
@@ -81,26 +136,15 @@ class BenchmarkTest(unittest.TestCase):
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 = PyTorchBenchmarkArguments(
models=[MODEL_ID], training=True, no_inference=True, sequence_lengths=[8], batch_sizes=[1]
)
benchmark = PyTorchBenchmark(benchmark_args, configs=[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_train_with_configs_torchscript(self):
MODEL_ID = "sshleifer/tiny-gpt2"
config = AutoConfig.from_pretrained(MODEL_ID)
benchmark_args = PyTorchBenchmarkArguments(
models=[MODEL_ID],
training=True,
no_inference=True,
torchscript=True,
sequence_lengths=[8],
batch_sizes=[1],
no_multi_process=True,
)
benchmark = PyTorchBenchmark(benchmark_args, configs=[config])
results = benchmark.run()
@@ -111,7 +155,12 @@ class BenchmarkTest(unittest.TestCase):
MODEL_ID = "sshleifer/tinier_bart"
config = AutoConfig.from_pretrained(MODEL_ID)
benchmark_args = PyTorchBenchmarkArguments(
models=[MODEL_ID], training=True, no_inference=True, sequence_lengths=[8], batch_sizes=[1]
models=[MODEL_ID],
training=True,
no_inference=True,
sequence_lengths=[8],
batch_sizes=[1],
no_multi_process=True,
)
benchmark = PyTorchBenchmark(benchmark_args, configs=[config])
results = benchmark.run()
@@ -133,6 +182,7 @@ class BenchmarkTest(unittest.TestCase):
inference_memory_csv_file=os.path.join(tmp_dir, "inf_mem.csv"),
train_time_csv_file=os.path.join(tmp_dir, "train_time.csv"),
env_info_csv_file=os.path.join(tmp_dir, "env.csv"),
no_multi_process=True,
)
benchmark = PyTorchBenchmark(benchmark_args)
benchmark.run()
@@ -161,6 +211,7 @@ class BenchmarkTest(unittest.TestCase):
log_filename=os.path.join(tmp_dir, "log.txt"),
log_print=True,
trace_memory_line_by_line=True,
no_multi_process=True,
)
benchmark = PyTorchBenchmark(benchmark_args)
result = benchmark.run()

165
tests/test_benchmark_tf.py Normal file
View File

@@ -0,0 +1,165 @@
import os
import tempfile
import unittest
from pathlib import Path
from transformers import AutoConfig, is_tf_available
from .utils import require_tf
if is_tf_available():
import tensorflow as tf
from transformers import TensorflowBenchmark, TensorflowBenchmarkArguments
@require_tf
class TFBenchmarkTest(unittest.TestCase):
def check_results_dict_not_empty(self, results):
for model_result in results.values():
for batch_size, sequence_length in zip(model_result["bs"], model_result["ss"]):
result = model_result["result"][batch_size][sequence_length]
self.assertIsNotNone(result)
def test_inference_no_configs_eager(self):
MODEL_ID = "sshleifer/tiny-gpt2"
benchmark_args = TensorflowBenchmarkArguments(
models=[MODEL_ID],
training=False,
no_inference=False,
sequence_lengths=[8],
batch_sizes=[1],
eager_mode=True,
no_multi_process=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_no_configs_graph(self):
MODEL_ID = "sshleifer/tiny-gpt2"
benchmark_args = TensorflowBenchmarkArguments(
models=[MODEL_ID],
training=False,
no_inference=False,
sequence_lengths=[8],
batch_sizes=[1],
no_multi_process=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_with_configs_eager(self):
MODEL_ID = "sshleifer/tiny-gpt2"
config = AutoConfig.from_pretrained(MODEL_ID)
benchmark_args = TensorflowBenchmarkArguments(
models=[MODEL_ID],
training=False,
no_inference=False,
sequence_lengths=[8],
batch_sizes=[1],
eager_mode=True,
no_multi_process=True,
)
benchmark = TensorflowBenchmark(benchmark_args, [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_inference_with_configs_graph(self):
MODEL_ID = "sshleifer/tiny-gpt2"
config = AutoConfig.from_pretrained(MODEL_ID)
benchmark_args = TensorflowBenchmarkArguments(
models=[MODEL_ID],
training=False,
no_inference=False,
sequence_lengths=[8],
batch_sizes=[1],
no_multi_process=True,
)
benchmark = TensorflowBenchmark(benchmark_args, [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_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,
no_inference=False,
sequence_lengths=[8],
batch_sizes=[1],
no_multi_process=True,
)
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)
@unittest.skipIf(is_tf_available() and len(tf.config.list_physical_devices("GPU")) == 0, "Cannot do xla on CPU.")
def test_inference_no_configs_xla(self):
MODEL_ID = "sshleifer/tiny-gpt2"
benchmark_args = TensorflowBenchmarkArguments(
models=[MODEL_ID],
training=False,
no_inference=False,
sequence_lengths=[8],
batch_sizes=[1],
use_xla=True,
no_multi_process=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_save_csv_files(self):
MODEL_ID = "sshleifer/tiny-gpt2"
with tempfile.TemporaryDirectory() as tmp_dir:
benchmark_args = TensorflowBenchmarkArguments(
models=[MODEL_ID],
no_inference=False,
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"),
no_multi_process=True,
)
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 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],
no_inference=False,
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,
no_multi_process=True,
)
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())