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

Extend Transformers Trainer Class to Enable PyTorch Torchscript for Inference (#17153)

Browse Source 
* add jit mode option and model wrap

* Update src/transformers/training_args.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/training_args.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* refine code

* Update src/transformers/trainer.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/trainer.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* add ut and refine code

* code refine

* refine code

* add inference doc

* Update src/transformers/trainer.py

Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>

* Update src/transformers/trainer.py

Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>

* add cpu inference performance doc

* Update perf_infer_cpu.mdx

* Update perf_infer_cpu.mdx

* Update performance.mdx

* Update _toctree.yml

* refine jit func naming

* Update _toctree.yml

* Delete perf_infer_gpu_one.mdx

* Update perf_infer_cpu.mdx

* Update docs/source/en/perf_infer_cpu.mdx

Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>

* add none check before jit

* Update docs/source/en/perf_infer_cpu.mdx

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update docs/source/en/perf_infer_cpu.mdx

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Stas Bekman <stas@stason.org>
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
This commit is contained in:
jianan-gu
2022-06-14 19:56:47 +08:00
committed by GitHub
parent df15703b42
commit 3b29c9fdb7
6 changed files with 169 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

75
tests/trainer/test_trainer.py
View File

@@ -844,6 +844,47 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
expected_acc = AlmostAccuracy()((pred + 1, y))["accuracy"]
self.assertAlmostEqual(results["eval_accuracy"], expected_acc)
def test_evaluate_with_jit(self):
trainer = get_regression_trainer(a=1.5, b=2.5, compute_metrics=AlmostAccuracy(), jit_mode_eval=True)
results = trainer.evaluate()
x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
pred = 1.5 * x + 2.5
expected_loss = ((pred - y) ** 2).mean()
self.assertAlmostEqual(results["eval_loss"], expected_loss)
expected_acc = AlmostAccuracy()((pred, y))["accuracy"]
self.assertAlmostEqual(results["eval_accuracy"], expected_acc)
# With a number of elements not a round multiple of the batch size
trainer = get_regression_trainer(
a=1.5, b=2.5, eval_len=66, compute_metrics=AlmostAccuracy(), jit_mode_eval=True
)
results = trainer.evaluate()
x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
pred = 1.5 * x + 2.5
expected_loss = ((pred - y) ** 2).mean()
self.assertAlmostEqual(results["eval_loss"], expected_loss)
expected_acc = AlmostAccuracy()((pred, y))["accuracy"]
self.assertAlmostEqual(results["eval_accuracy"], expected_acc)
# With logits preprocess
trainer = get_regression_trainer(
a=1.5,
b=2.5,
compute_metrics=AlmostAccuracy(),
preprocess_logits_for_metrics=lambda logits, labels: logits + 1,
jit_mode_eval=True,
)
results = trainer.evaluate()
x, y = trainer.eval_dataset.x, trainer.eval_dataset.ys[0]
pred = 1.5 * x + 2.5
expected_loss = ((pred - y) ** 2).mean()
self.assertAlmostEqual(results["eval_loss"], expected_loss)
expected_acc = AlmostAccuracy()((pred + 1, y))["accuracy"]
self.assertAlmostEqual(results["eval_accuracy"], expected_acc)
@require_torch_bf16
@require_intel_extension_for_pytorch
def test_evaluate_with_ipex(self):
@@ -930,6 +971,40 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
self.assertTrue(np.array_equal(labels[0], trainer.eval_dataset.ys[0]))
self.assertTrue(np.array_equal(labels[1], trainer.eval_dataset.ys[1]))
def test_predict_with_jit(self):
trainer = get_regression_trainer(a=1.5, b=2.5, jit_mode_eval=True)
preds = trainer.predict(trainer.eval_dataset).predictions
x = trainer.eval_dataset.x
self.assertTrue(np.allclose(preds, 1.5 * x + 2.5))
# With a number of elements not a round multiple of the batch size
trainer = get_regression_trainer(a=1.5, b=2.5, eval_len=66, jit_mode_eval=True)
preds = trainer.predict(trainer.eval_dataset).predictions
x = trainer.eval_dataset.x
self.assertTrue(np.allclose(preds, 1.5 * x + 2.5))
# With more than one output of the model
trainer = get_regression_trainer(a=1.5, b=2.5, double_output=True, jit_mode_eval=True)
preds = trainer.predict(trainer.eval_dataset).predictions
x = trainer.eval_dataset.x
self.assertEqual(len(preds), 2)
self.assertTrue(np.allclose(preds[0], 1.5 * x + 2.5))
self.assertTrue(np.allclose(preds[1], 1.5 * x + 2.5))
# With more than one output/label of the model
trainer = get_regression_trainer(
a=1.5, b=2.5, double_output=True, label_names=["labels", "labels_2"], jit_mode_eval=True
)
outputs = trainer.predict(trainer.eval_dataset)
preds = outputs.predictions
labels = outputs.label_ids
x = trainer.eval_dataset.x
self.assertEqual(len(preds), 2)
self.assertTrue(np.allclose(preds[0], 1.5 * x + 2.5))
self.assertTrue(np.allclose(preds[1], 1.5 * x + 2.5))
self.assertTrue(np.array_equal(labels[0], trainer.eval_dataset.ys[0]))
self.assertTrue(np.array_equal(labels[1], trainer.eval_dataset.ys[1]))
@require_torch_bf16
@require_intel_extension_for_pytorch
def test_predict_with_ipex(self):