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

[tests] remove TF tests (uses of require_tf) (#38944)

Browse Source 
* remove uses of require_tf

* remove redundant import guards

* this class has no tests

* nits

* del tf rng comment
This commit is contained in:
Joao Gante
2025-06-25 18:29:10 +01:00
committed by GitHub
parent d37f751797
commit 1d45d90e5d
44 changed files with 21 additions and 2504 deletions
Download Patch File Download Diff File Expand all files Collapse all files

106
tests/models/bert/test_tokenization_bert_tf.py
View File

@@ -1,106 +0,0 @@
import unittest
from pathlib import Path
from tempfile import TemporaryDirectory
from transformers import AutoConfig, TFAutoModel, is_tensorflow_text_available, is_tf_available
from transformers.models.bert.tokenization_bert import BertTokenizer
from transformers.testing_utils import require_tensorflow_text, require_tf, slow
if is_tf_available():
import tensorflow as tf
from transformers.modeling_tf_utils import keras
if is_tensorflow_text_available():
from transformers.models.bert import TFBertTokenizer
TOKENIZER_CHECKPOINTS = ["google-bert/bert-base-uncased", "google-bert/bert-base-cased"]
TINY_MODEL_CHECKPOINT = "hf-internal-testing/tiny-bert-tf-only"
if is_tf_available():
from transformers.modeling_tf_utils import keras
class ModelToSave(keras.Model):
def __init__(self, tokenizer):
super().__init__()
self.tokenizer = tokenizer
config = AutoConfig.from_pretrained(TINY_MODEL_CHECKPOINT)
self.bert = TFAutoModel.from_config(config)
def call(self, inputs):
tokenized = self.tokenizer(inputs)
out = self.bert(tokenized)
return out["pooler_output"]
@require_tf
@require_tensorflow_text
class BertTokenizationTest(unittest.TestCase):
# The TF tokenizers are usually going to be used as pretrained tokenizers from existing model checkpoints,
# so that's what we focus on here.
def setUp(self):
super().setUp()
self.tokenizers = [BertTokenizer.from_pretrained(checkpoint) for checkpoint in TOKENIZER_CHECKPOINTS]
self.tf_tokenizers = [TFBertTokenizer.from_pretrained(checkpoint) for checkpoint in TOKENIZER_CHECKPOINTS]
assert len(self.tokenizers) == len(self.tf_tokenizers)
self.test_sentences = [
"This is a straightforward English test sentence.",
"This one has some weird characters\rto\nsee\r\nif those\u00e9break things.",
"Now we're going to add some Chinese: 一 二 三 一二三",
"And some much more rare Chinese: 齉 堃 齉堃",
"Je vais aussi écrire en français pour tester les accents",
"Classical Irish also has some unusual characters, so in they go: Gaelaċ, ꝼ",
]
self.paired_sentences = list(zip(self.test_sentences, self.test_sentences[::-1]))
def test_output_equivalence(self):
for tokenizer, tf_tokenizer in zip(self.tokenizers, self.tf_tokenizers):
for test_inputs in (self.test_sentences, self.paired_sentences):
python_outputs = tokenizer(test_inputs, return_tensors="tf", padding="longest")
tf_outputs = tf_tokenizer(test_inputs)
for key in python_outputs.keys():
self.assertTrue(tf.reduce_all(python_outputs[key].shape == tf_outputs[key].shape))
self.assertTrue(tf.reduce_all(tf.cast(python_outputs[key], tf.int64) == tf_outputs[key]))
@slow
def test_different_pairing_styles(self):
for tf_tokenizer in self.tf_tokenizers:
merged_outputs = tf_tokenizer(self.paired_sentences)
separated_outputs = tf_tokenizer(
text=[sentence[0] for sentence in self.paired_sentences],
text_pair=[sentence[1] for sentence in self.paired_sentences],
)
for key in merged_outputs.keys():
self.assertTrue(tf.reduce_all(tf.cast(merged_outputs[key], tf.int64) == separated_outputs[key]))
@slow
def test_graph_mode(self):
for tf_tokenizer in self.tf_tokenizers:
compiled_tokenizer = tf.function(tf_tokenizer)
for test_inputs in (self.test_sentences, self.paired_sentences):
test_inputs = tf.constant(test_inputs)
compiled_outputs = compiled_tokenizer(test_inputs)
eager_outputs = tf_tokenizer(test_inputs)
for key in eager_outputs.keys():
self.assertTrue(tf.reduce_all(eager_outputs[key] == compiled_outputs[key]))
@slow
def test_export_for_inference(self):
for tf_tokenizer in self.tf_tokenizers:
model = ModelToSave(tokenizer=tf_tokenizer)
test_inputs = tf.convert_to_tensor(self.test_sentences)
out = model(test_inputs) # Build model with some sample inputs
with TemporaryDirectory() as tempdir:
save_path = Path(tempdir) / "saved.model"
model.export(save_path)
loaded_model = tf.saved_model.load(save_path)
loaded_output = loaded_model.serve(test_inputs)
# We may see small differences because the loaded model is compiled, so we need an epsilon for the test
self.assertLessEqual(tf.reduce_max(tf.abs(out - loaded_output)), 1e-5)

131
tests/models/gpt2/test_tokenization_gpt2_tf.py
View File

@@ -1,131 +0,0 @@
import unittest
from pathlib import Path
from tempfile import TemporaryDirectory
from transformers import AutoConfig, TFGPT2LMHeadModel, is_keras_nlp_available, is_tf_available
from transformers.models.gpt2.tokenization_gpt2 import GPT2Tokenizer
from transformers.testing_utils import require_keras_nlp, require_tf, slow
if is_tf_available():
import tensorflow as tf
if is_keras_nlp_available():
from transformers.models.gpt2 import TFGPT2Tokenizer
TOKENIZER_CHECKPOINTS = ["openai-community/gpt2"]
TINY_MODEL_CHECKPOINT = "openai-community/gpt2"
if is_tf_available():
class ModelToSave(tf.Module):
def __init__(self, tokenizer):
super().__init__()
self.tokenizer = tokenizer
config = AutoConfig.from_pretrained(TINY_MODEL_CHECKPOINT)
self.model = TFGPT2LMHeadModel.from_config(config)
@tf.function(input_signature=(tf.TensorSpec((None,), tf.string, name="text"),))
def serving(self, text):
tokenized = self.tokenizer(text)
input_ids_dense = tokenized["input_ids"].to_tensor()
input_mask = tf.cast(input_ids_dense > 0, tf.int32)
# input_mask = tf.reshape(input_mask, [-1, MAX_SEQ_LEN])
outputs = self.model(input_ids=input_ids_dense, attention_mask=input_mask)["logits"]
return outputs
@require_tf
@require_keras_nlp
class GPTTokenizationTest(unittest.TestCase):
# The TF tokenizers are usually going to be used as pretrained tokenizers from existing model checkpoints,
# so that's what we focus on here.
def setUp(self):
super().setUp()
self.tokenizers = [GPT2Tokenizer.from_pretrained(checkpoint) for checkpoint in (TOKENIZER_CHECKPOINTS)]
self.tf_tokenizers = [TFGPT2Tokenizer.from_pretrained(checkpoint) for checkpoint in TOKENIZER_CHECKPOINTS]
assert len(self.tokenizers) == len(self.tf_tokenizers)
self.test_sentences = [
"This is a straightforward English test sentence.",
"This one has some weird characters\rto\nsee\r\nif those\u00e9break things.",
"Now we're going to add some Chinese: 一 二 三 一二三",
"And some much more rare Chinese: 齉 堃 齉堃",
"Je vais aussi écrire en français pour tester les accents",
"Classical Irish also has some unusual characters, so in they go: Gaelaċ, ꝼ",
]
self.paired_sentences = list(zip(self.test_sentences, self.test_sentences[::-1]))
def test_output_equivalence(self):
for tokenizer, tf_tokenizer in zip(self.tokenizers, self.tf_tokenizers):
for test_inputs in self.test_sentences:
python_outputs = tokenizer([test_inputs], return_tensors="tf")
tf_outputs = tf_tokenizer([test_inputs])
for key in python_outputs.keys():
# convert them to numpy to avoid messing with ragged tensors
python_outputs_values = python_outputs[key].numpy()
tf_outputs_values = tf_outputs[key].numpy()
self.assertTrue(tf.reduce_all(python_outputs_values.shape == tf_outputs_values.shape))
self.assertTrue(tf.reduce_all(tf.cast(python_outputs_values, tf.int64) == tf_outputs_values))
@slow
def test_graph_mode(self):
for tf_tokenizer in self.tf_tokenizers:
compiled_tokenizer = tf.function(tf_tokenizer)
for test_inputs in self.test_sentences:
test_inputs = tf.constant(test_inputs)
compiled_outputs = compiled_tokenizer(test_inputs)
eager_outputs = tf_tokenizer(test_inputs)
for key in eager_outputs.keys():
self.assertTrue(tf.reduce_all(eager_outputs[key] == compiled_outputs[key]))
@slow
def test_saved_model(self):
for tf_tokenizer in self.tf_tokenizers:
model = ModelToSave(tokenizer=tf_tokenizer)
test_inputs = tf.convert_to_tensor([self.test_sentences[0]])
out = model.serving(test_inputs) # Build model with some sample inputs
with TemporaryDirectory() as tempdir:
save_path = Path(tempdir) / "saved.model"
tf.saved_model.save(model, save_path, signatures={"serving_default": model.serving})
loaded_model = tf.saved_model.load(save_path)
loaded_output = loaded_model.signatures["serving_default"](test_inputs)["output_0"]
# We may see small differences because the loaded model is compiled, so we need an epsilon for the test
self.assertTrue(tf.reduce_all(out == loaded_output))
@slow
def test_from_config(self):
for tf_tokenizer in self.tf_tokenizers:
test_inputs = tf.convert_to_tensor([self.test_sentences[0]])
out = tf_tokenizer(test_inputs) # Build model with some sample inputs
config = tf_tokenizer.get_config()
model_from_config = TFGPT2Tokenizer.from_config(config)
from_config_output = model_from_config(test_inputs)
for key in from_config_output.keys():
self.assertTrue(tf.reduce_all(from_config_output[key] == out[key]))
@slow
def test_padding(self):
for tf_tokenizer in self.tf_tokenizers:
# for the test to run
tf_tokenizer.pad_token_id = 123123
for max_length in [3, 5, 1024]:
test_inputs = tf.convert_to_tensor([self.test_sentences[0]])
out = tf_tokenizer(test_inputs, max_length=max_length)
out_length = out["input_ids"].numpy().shape[1]
assert out_length == max_length

37
tests/models/layoutlmv3/test_tokenization_layoutlmv3.py
View File

@@ -34,7 +34,6 @@ from transformers import (
from transformers.models.layoutlmv3.tokenization_layoutlmv3 import VOCAB_FILES_NAMES, LayoutLMv3Tokenizer
from transformers.testing_utils import (
require_pandas,
require_tf,
require_tokenizers,
require_torch,
slow,
@@ -2306,42 +2305,6 @@ class LayoutLMv3TokenizationTest(TokenizerTesterMixin, unittest.TestCase):
def test_np_encode_plus_sent_to_model(self):
pass
@require_tf
@slow
def test_tf_encode_plus_sent_to_model(self):
from transformers import TF_MODEL_MAPPING, TOKENIZER_MAPPING
MODEL_TOKENIZER_MAPPING = merge_model_tokenizer_mappings(TF_MODEL_MAPPING, TOKENIZER_MAPPING)
tokenizers = self.get_tokenizers(do_lower_case=False)
for tokenizer in tokenizers:
with self.subTest(f"{tokenizer.__class__.__name__}"):
if tokenizer.__class__ not in MODEL_TOKENIZER_MAPPING:
self.skipTest(f"{tokenizer.__class__} is not in the MODEL_TOKENIZER_MAPPING")
config_class, model_class = MODEL_TOKENIZER_MAPPING[tokenizer.__class__]
config = config_class()
if config.is_encoder_decoder or config.pad_token_id is None:
self.skipTest(reason="Model is an encoder-decoder or has no pad token id set.")
model = model_class(config)
# Make sure the model contains at least the full vocabulary size in its embedding matrix
self.assertGreaterEqual(model.config.vocab_size, len(tokenizer))
# Build sequence
first_ten_tokens = list(tokenizer.get_vocab().keys())[:10]
boxes = [[1000, 1000, 1000, 1000] for _ in range(len(first_ten_tokens))]
encoded_sequence = tokenizer.encode_plus(first_ten_tokens, boxes=boxes, return_tensors="tf")
batch_encoded_sequence = tokenizer.batch_encode_plus(
[first_ten_tokens, first_ten_tokens], boxes=[boxes, boxes], return_tensors="tf"
)
# This should not fail
model(encoded_sequence)
model(batch_encoded_sequence)
@unittest.skip(reason="Chat is not supported")
def test_chat_template(self):
pass

23
tests/models/pop2piano/test_feature_extraction_pop2piano.py
View File

@@ -24,7 +24,6 @@ from transformers.testing_utils import (
require_essentia,
require_librosa,
require_scipy,
require_tf,
require_torch,
)
from transformers.utils.import_utils import (
@@ -231,28 +230,6 @@ class Pop2PianoFeatureExtractionTest(SequenceFeatureExtractionTestMixin, unittes
# check shape
self.assertEqual(len(input_features["input_features"].shape), 3)
@require_tf
def test_batch_feature_tf(self):
import tensorflow as tf
feature_extractor = self.feature_extraction_class(**self.feat_extract_tester.prepare_feat_extract_dict())
speech_input1 = np.zeros([1_000_000], dtype=np.float32)
speech_input2 = np.ones([2_000_000], dtype=np.float32)
speech_input3 = np.random.randint(low=0, high=10, size=500_000).astype(np.float32)
input_features = feature_extractor(
[speech_input1, speech_input2, speech_input3],
sampling_rate=[44_100, 16_000, 48_000],
return_tensors="tf",
return_attention_mask=True,
)
# check tf tensor or not
self.assertTrue(tf.is_tensor(input_features["input_features"]))
# check shape
self.assertEqual(len(input_features["input_features"].shape), 3)
@unittest.skip(
"Pop2PianoFeatureExtractor does not supports padding externally (while processing audios in batches padding is automatically applied to max_length)"
)

156
tests/models/sam/test_processor_sam.py
View File

@@ -17,15 +17,10 @@ import unittest
import numpy as np
from transformers.testing_utils import (
require_tf,
require_torch,
require_torchvision,
require_vision,
)
from transformers.utils import is_tf_available, is_torch_available, is_vision_available
from transformers.testing_utils import require_torch, require_torchvision, require_vision
from transformers.utils import is_torch_available, is_vision_available
from ...test_processing_common import ProcessorTesterMixin, prepare_image_inputs
from ...test_processing_common import ProcessorTesterMixin
if is_vision_available():
@@ -38,11 +33,6 @@ if is_torch_available():
from transformers.models.sam.image_processing_sam import _mask_to_rle_pytorch
if is_tf_available():
import tensorflow as tf
from transformers.models.sam.image_processing_sam import _mask_to_rle_tf
@require_vision
@require_torchvision
@@ -202,143 +192,3 @@ class SamProcessorTest(ProcessorTesterMixin, unittest.TestCase):
self.assertEqual(len(rle), 1)
self.assertEqual(rle[0]["size"], [2, 2])
self.assertEqual(rle[0]["counts"], [1, 3]) # 1 zero, followed by 3 ones
@require_vision
@require_tf
class TFSamProcessorTest(unittest.TestCase):
def setUp(self):
self.tmpdirname = tempfile.mkdtemp()
image_processor = SamImageProcessor()
processor = SamProcessor(image_processor)
processor.save_pretrained(self.tmpdirname)
def get_image_processor(self, **kwargs):
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
def tearDown(self):
shutil.rmtree(self.tmpdirname)
# This is to avoid repeating the skipping of the common tests
def prepare_image_inputs(self):
"""This function prepares a list of PIL images."""
return prepare_image_inputs()
def test_save_load_pretrained_additional_features(self):
processor = SamProcessor(image_processor=self.get_image_processor())
processor.save_pretrained(self.tmpdirname)
image_processor_add_kwargs = self.get_image_processor(do_normalize=False, padding_value=1.0)
processor = SamProcessor.from_pretrained(self.tmpdirname, do_normalize=False, padding_value=1.0)
self.assertEqual(processor.image_processor.to_json_string(), image_processor_add_kwargs.to_json_string())
self.assertIsInstance(processor.image_processor, SamImageProcessor)
def test_image_processor(self):
image_processor = self.get_image_processor()
processor = SamProcessor(image_processor=image_processor)
image_input = self.prepare_image_inputs()
input_feat_extract = image_processor(image_input, return_tensors="np")
input_processor = processor(images=image_input, return_tensors="np")
input_feat_extract.pop("original_sizes") # pop original_sizes as it is popped in the processor
input_feat_extract.pop("reshaped_input_sizes") # pop reshaped_input_sizes as it is popped in the processor
for key in input_feat_extract.keys():
self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
@require_tf
def test_post_process_masks(self):
image_processor = self.get_image_processor()
processor = SamProcessor(image_processor=image_processor)
dummy_masks = [tf.ones((1, 3, 5, 5))]
original_sizes = [[1764, 2646]]
reshaped_input_size = [[683, 1024]]
masks = processor.post_process_masks(dummy_masks, original_sizes, reshaped_input_size, return_tensors="tf")
self.assertEqual(masks[0].shape, (1, 3, 1764, 2646))
masks = processor.post_process_masks(
dummy_masks,
tf.convert_to_tensor(original_sizes),
tf.convert_to_tensor(reshaped_input_size),
return_tensors="tf",
)
self.assertEqual(masks[0].shape, (1, 3, 1764, 2646))
# should also work with np
dummy_masks = [np.ones((1, 3, 5, 5))]
masks = processor.post_process_masks(
dummy_masks, np.array(original_sizes), np.array(reshaped_input_size), return_tensors="tf"
)
self.assertEqual(masks[0].shape, (1, 3, 1764, 2646))
dummy_masks = [[1, 0], [0, 1]]
with self.assertRaises(tf.errors.InvalidArgumentError):
masks = processor.post_process_masks(
dummy_masks, np.array(original_sizes), np.array(reshaped_input_size), return_tensors="tf"
)
def test_rle_encoding(self):
"""
Test the run-length encoding function.
"""
# Test that a mask of all zeros returns a single run [height * width].
input_mask = tf.zeros((1, 2, 2), dtype=tf.int64) # shape: 1 x 2 x 2
rle = _mask_to_rle_tf(input_mask)
self.assertEqual(len(rle), 1)
self.assertEqual(rle[0]["size"], [2, 2])
# For a 2x2 all-zero mask, we expect a single run of length 4:
self.assertEqual(rle[0]["counts"], [4])
# Test that a mask of all ones returns [0, height * width].
input_mask = tf.ones((1, 2, 2), dtype=tf.int64) # shape: 1 x 2 x 2
rle = _mask_to_rle_tf(input_mask)
self.assertEqual(len(rle), 1)
self.assertEqual(rle[0]["size"], [2, 2])
# For a 2x2 all-one mask, we expect two runs: [0, 4].
self.assertEqual(rle[0]["counts"], [0, 4])
# Test a mask with mixed 0s and 1s to ensure the run-length encoding is correct.
# Example mask:
# Row 0: [0, 1]
# Row 1: [1, 1]
# This is shape (1, 2, 2).
# Flattened in Fortran order -> [0, 1, 1, 1].
# The RLE for [0,1,1,1] is [1, 3].
input_mask = tf.constant([[[0, 1], [1, 1]]], dtype=tf.int64)
rle = _mask_to_rle_tf(input_mask)
self.assertEqual(len(rle), 1)
self.assertEqual(rle[0]["size"], [2, 2])
self.assertEqual(rle[0]["counts"], [1, 3]) # 1 zero, followed by 3 ones
@require_vision
@require_torchvision
class SamProcessorEquivalenceTest(unittest.TestCase):
def setUp(self):
self.tmpdirname = tempfile.mkdtemp()
image_processor = SamImageProcessor()
processor = SamProcessor(image_processor)
processor.save_pretrained(self.tmpdirname)
def get_image_processor(self, **kwargs):
return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
def tearDown(self):
shutil.rmtree(self.tmpdirname)
# This is to avoid repeating the skipping of the common tests
def prepare_image_inputs(self):
"""This function prepares a list of PIL images."""
return prepare_image_inputs()

11
tests/models/whisper/test_tokenization_whisper.py
View File

@@ -18,7 +18,7 @@ import numpy as np
from transformers.models.whisper import WhisperTokenizer, WhisperTokenizerFast
from transformers.models.whisper.tokenization_whisper import _combine_tokens_into_words, _find_longest_common_sequence
from transformers.testing_utils import require_flax, require_tf, require_torch, slow
from transformers.testing_utils import require_flax, require_torch, slow
from ...test_tokenization_common import TokenizerTesterMixin
@@ -588,15 +588,6 @@ class SpeechToTextTokenizerMultilinguialTest(unittest.TestCase):
self.assertListEqual(WhisperTokenizer._convert_to_list(np_array), test_list)
self.assertListEqual(WhisperTokenizerFast._convert_to_list(np_array), test_list)
@require_tf
def test_convert_to_list_tf(self):
import tensorflow as tf
test_list = [[1, 2, 3], [4, 5, 6]]
tf_tensor = tf.constant(test_list)
self.assertListEqual(WhisperTokenizer._convert_to_list(tf_tensor), test_list)
self.assertListEqual(WhisperTokenizerFast._convert_to_list(tf_tensor), test_list)
@require_flax
def test_convert_to_list_jax(self):
import jax.numpy as jnp