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

use torch.testing.assertclose instead to get more details about error in cis (#35659)

Browse Source 
* use torch.testing.assertclose instead to get more details about error in cis

* fix

* style

* test_all

* revert for I bert

* fixes and updates

* more image processing fixes

* more image processors

* fix mamba and co

* style

* less strick

* ok I won't be strict

* skip and be done

* up
This commit is contained in:
Arthur
2025-01-24 16:55:28 +01:00
committed by GitHub
parent 72d1a4cd53
commit b912f5ee43
255 changed files with 1048 additions and 969 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
tests/trainer/test_trainer.py
View File

@@ -609,8 +609,8 @@ class TrainerIntegrationCommon:
state_dict = safetensors.torch.load_file(os.path.join(checkpoint, SAFE_WEIGHTS_NAME))
best_model.load_state_dict(state_dict)
best_model.to(trainer.args.device)
self.assertTrue(torch.allclose(best_model.a, trainer.model.a))
self.assertTrue(torch.allclose(best_model.b, trainer.model.b))
torch.testing.assert_close(best_model.a, trainer.model.a)
torch.testing.assert_close(best_model.b, trainer.model.b)
metrics = trainer.evaluate()
self.assertEqual(metrics[metric], best_value)
@@ -698,8 +698,8 @@ class TrainerIntegrationPrerunTest(TestCasePlus, TrainerIntegrationCommon):
def check_trained_model(self, model, alternate_seed=False):
# Checks a training seeded with learning_rate = 0.1
(a, b) = self.alternate_trained_model if alternate_seed else self.default_trained_model
self.assertTrue(torch.allclose(model.a, a))
self.assertTrue(torch.allclose(model.b, b))
torch.testing.assert_close(model.a, a)
torch.testing.assert_close(model.b, b)
def test_reproducible_training(self):
# Checks that training worked, model trained and seed made a reproducible training.
@@ -1567,8 +1567,7 @@ class TrainerIntegrationTest(TestCasePlus, TrainerIntegrationCommon):
# Check that we get identical embeddings just in case
emb1 = trainer.model.get_input_embeddings()(dummy_input)
emb2 = trainer.model.get_input_embeddings()(dummy_input)
self.assertTrue(torch.allclose(emb1, emb2), "Neftune noise is still applied!")
torch.testing.assert_close(emb1, emb2)
def test_logging_inf_nan_filter(self):
config = GPT2Config(vocab_size=100, n_positions=128, n_embd=32, n_layer=3, n_head=4)

4
tests/trainer/test_trainer_utils.py
View File

@@ -162,7 +162,7 @@ class TrainerUtilsTest(unittest.TestCase):
label_smoothed_loss = LabelSmoother(0.1)(model_output, random_labels)
log_probs = -nn.functional.log_softmax(random_logits, dim=-1)
expected_loss = (1 - epsilon) * loss + epsilon * log_probs.mean()
self.assertTrue(torch.allclose(label_smoothed_loss, expected_loss))
torch.testing.assert_close(label_smoothed_loss, expected_loss)
# With a few -100 labels
random_labels[0, 1] = -100
@@ -178,7 +178,7 @@ class TrainerUtilsTest(unittest.TestCase):
log_probs[2, 1] = 0.0
log_probs[2, 3] = 0.0
expected_loss = (1 - epsilon) * loss + epsilon * log_probs.sum() / (num_labels * 17)
self.assertTrue(torch.allclose(label_smoothed_loss, expected_loss))
torch.testing.assert_close(label_smoothed_loss, expected_loss)
def test_group_by_length(self):
# Get some inputs of random lengths