[tests] remove pt_tf equivalence tests (#36253)

tests/models/groupvit/test_modeling_groupvit.py

@@ -24,7 +24,7 @@ import numpy as np
 import requests
 
 from transformers import GroupViTConfig, GroupViTTextConfig, GroupViTVisionConfig
-from transformers.testing_utils import is_flaky, is_pt_tf_cross_test, require_torch, require_vision, slow, torch_device
+from transformers.testing_utils import is_flaky, require_torch, require_vision, slow, torch_device
 from transformers.utils import is_torch_available, is_vision_available
 
 from ...test_configuration_common import ConfigTester
@@ -166,18 +166,6 @@ class GroupViTVisionModelTest(ModelTesterMixin, unittest.TestCase):
     def test_batching_equivalence(self):
         super().test_batching_equivalence()
 
-    @is_pt_tf_cross_test
-    def test_pt_tf_model_equivalence(self):
-        import tensorflow as tf
-
-        seed = 338
-        random.seed(seed)
-        np.random.seed(seed)
-        torch.manual_seed(seed)
-        torch.cuda.manual_seed_all(seed)
-        tf.random.set_seed(seed)
-        return super().test_pt_tf_model_equivalence()
-
     def test_model_get_set_embeddings(self):
         config, _ = self.model_tester.prepare_config_and_inputs_for_common()
 
@@ -595,22 +583,6 @@ class GroupViTModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase
     def test_model_get_set_embeddings(self):
         pass
 
-    # overwritten from parent as this equivalent test needs a specific `seed` and hard to get a good one!
-    def check_pt_tf_outputs(self, tf_outputs, pt_outputs, model_class, tol=2e-5, name="outputs", attributes=None):
-        super().check_pt_tf_outputs(tf_outputs, pt_outputs, model_class, tol=tol, name=name, attributes=attributes)
-
-    @is_pt_tf_cross_test
-    def test_pt_tf_model_equivalence(self):
-        import tensorflow as tf
-
-        seed = 163
-        random.seed(seed)
-        np.random.seed(seed)
-        torch.manual_seed(seed)
-        torch.cuda.manual_seed_all(seed)
-        tf.random.set_seed(seed)
-        return super().test_pt_tf_model_equivalence()
-
     # override as the `logit_scale` parameter initilization is different for GROUPVIT
     def test_initialization(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

tests/models/groupvit/test_modeling_tf_groupvit.py

@@ -23,12 +23,10 @@ import tempfile
 import unittest
 from importlib import import_module
 
-import numpy as np
 import requests
 
 from transformers import GroupViTConfig, GroupViTTextConfig, GroupViTVisionConfig
 from transformers.testing_utils import (
-    is_pt_tf_cross_test,
     require_tensorflow_probability,
     require_tf,
     require_vision,
@@ -149,10 +147,6 @@ class TFGroupViTVisionModelTest(TFModelTesterMixin, unittest.TestCase):
     test_head_masking = False
     test_onnx = False
 
-    def check_pt_tf_outputs(self, tf_outputs, pt_outputs, model_class, tol=1e-4, name="outputs", attributes=None):
-        # We override with a slightly higher tol value, as this model tends to diverge a bit more
-        super().check_pt_tf_outputs(tf_outputs, pt_outputs, model_class, tol, name, attributes)
-
     def setUp(self):
         self.model_tester = TFGroupViTVisionModelTester(self)
         self.config_tester = ConfigTester(
@@ -291,25 +285,6 @@ class TFGroupViTVisionModelTest(TFModelTesterMixin, unittest.TestCase):
 
             check_hidden_states_output(inputs_dict, config, model_class)
 
-    @is_pt_tf_cross_test
-    def test_pt_tf_model_equivalence(self):
-        # `GroupViT` computes some indices using argmax, uses them as
-        # one-hot encoding for further computation. The problem is
-        # while PT/TF have very small difference in `y_soft` (~ 1e-9),
-        # the argmax could be totally different, if there are at least
-        # 2 indices with almost identical values. This leads to very
-        # large difference in the outputs. We need specific seeds to
-        # avoid almost identical values happening in `y_soft`.
-        import torch
-
-        seed = 338
-        random.seed(seed)
-        np.random.seed(seed)
-        torch.manual_seed(seed)
-        torch.cuda.manual_seed_all(seed)
-        tf.random.set_seed(seed)
-        return super().test_pt_tf_model_equivalence()
-
     @slow
     def test_model_from_pretrained(self):
         model_name = "nvidia/groupvit-gcc-yfcc"
@@ -462,10 +437,6 @@ class TFGroupViTTextModelTest(TFModelTesterMixin, unittest.TestCase):
     test_head_masking = False
     test_onnx = False
 
-    def check_pt_tf_outputs(self, tf_outputs, pt_outputs, model_class, tol=1e-4, name="outputs", attributes=None):
-        # We override with a slightly higher tol value, as this model tends to diverge a bit more
-        super().check_pt_tf_outputs(tf_outputs, pt_outputs, model_class, tol, name, attributes)
-
     def setUp(self):
         self.model_tester = TFGroupViTTextModelTester(self)
         self.config_tester = ConfigTester(self, config_class=GroupViTTextConfig, hidden_size=37)
@@ -588,10 +559,6 @@ class TFGroupViTModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.Test
     test_attention_outputs = False
     test_onnx = False
 
-    def check_pt_tf_outputs(self, tf_outputs, pt_outputs, model_class, tol=1e-4, name="outputs", attributes=None):
-        # We override with a slightly higher tol value, as this model tends to diverge a bit more
-        super().check_pt_tf_outputs(tf_outputs, pt_outputs, model_class, tol, name, attributes)
-
     def setUp(self):
         self.model_tester = TFGroupViTModelTester(self)
 
@@ -616,25 +583,6 @@ class TFGroupViTModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.Test
     def test_keras_fit(self):
         super().test_keras_fit()
 
-    @is_pt_tf_cross_test
-    def test_pt_tf_model_equivalence(self):
-        # `GroupViT` computes some indices using argmax, uses them as
-        # one-hot encoding for further computation. The problem is
-        # while PT/TF have very small difference in `y_soft` (~ 1e-9),
-        # the argmax could be totally different, if there are at least
-        # 2 indices with almost identical values. This leads to very
-        # large difference in the outputs. We need specific seeds to
-        # avoid almost identical values happening in `y_soft`.
-        import torch
-
-        seed = 158
-        random.seed(seed)
-        np.random.seed(seed)
-        torch.manual_seed(seed)
-        torch.cuda.manual_seed_all(seed)
-        tf.random.set_seed(seed)
-        return super().test_pt_tf_model_equivalence()
-
     # overwrite from common since `TFGroupViTModelTester` set `return_loss` to `True` and causes the preparation of
     # `symbolic_inputs` failed.
     def test_keras_save_load(self):