[TensorFlow] Adding GroupViT (#18020)

* chore: initial commit

* chore: adding util methods

yet to work on the nn.functional.interpolate port with align_corener=True

* chore: refactor the utils

* used tf.compat.v1.image.resize to align the F.interpolate function
* added type hints to the method signatures
* added references to the gists where one 2 one alignment of torch and tf has been shown

* chore: adding the layers

* chore: porting all the layers from torch to tf

This is the initial draft, nothing is tested yet.

* chore: aligning the layers with reference to tf clip

* chore: aligning the modules

* added demaraction comments
* added copied and adapted from comments

* chore: aligning with CLIP

* chore: wrangling the layers to keep it tf compatible

* chore: aligning the names of the layers for porting

* chore: style changes

* chore: adding docs and inits

* chore: adding tfp dependencis

the code is taken from TAPAS

* chore: initial commit for testing

* chore: aligning the vision embeddings with the vit implementatino

* chore: changing model prefix

* chore: fixing the name of the model and the layer normalization test case

* chore: every test passes but the slow ones

* chore: fix style and integration test

* chore: moving comments below decorators

* chore: make fixup and fix-copies changes

* chore: adding the Vision and Text Model to check_repo

* chore: modifying the prefix name to align it with the torch implementation

* chore: fix typo in configuration

* choer: changing the name of the model variable

* chore: adding segmentation flag

* chore: gante's review

* chore: style refactor

* chore: amy review

* chore: adding shape_list to parts that have been copied from other snippets

* chore: init batchnorm with torch defaults

* chore: adding shape_list to pass the tests

* test fix: adding seed as 0

* set seed

* chore: changing the straight through trick to fix -ve dimensinos

* chore: adding a dimension to the loss

* chore: adding reviewers and contributors names to the docs

* chore: added changes after review

* chore: code quality fixup

* chore: fixing the segmentation snippet

* chore: adding  to the layer calls

* chore: changing int32 to int64 for inputs of serving

* chore: review changes

* chore: style changes

* chore: remove from_pt=True

* fix: repo consistency

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

tests/models/groupvit/test_modeling_groupvit.py

@@ -17,6 +17,7 @@
 
 import inspect
 import os
+import random
 import tempfile
 import unittest
 
@@ -24,7 +25,7 @@ import numpy as np
 
 import requests
 from transformers import GroupViTConfig, GroupViTTextConfig, GroupViTVisionConfig
-from transformers.testing_utils import require_torch, require_vision, slow, torch_device
+from transformers.testing_utils import is_pt_tf_cross_test, require_torch, require_vision, slow, torch_device
 from transformers.utils import is_torch_available, is_vision_available
 
 from ...test_configuration_common import ConfigTester
@@ -95,7 +96,8 @@ class GroupViTVisionModelTester:
         self.seq_length = num_patches
 
     def prepare_config_and_inputs(self):
-        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+        rng = random.Random(0)
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size], rng=rng)
         config = self.get_config()
 
         return config, pixel_values
@@ -161,6 +163,18 @@ class GroupViTVisionModelTest(ModelTesterMixin, unittest.TestCase):
     def test_inputs_embeds(self):
         pass
 
+    @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_common_attributes(self):
         config, _ = self.model_tester.prepare_config_and_inputs_for_common()
 
@@ -368,7 +382,8 @@ class GroupViTTextModelTester:
         self.scope = scope
 
     def prepare_config_and_inputs(self):
-        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+        rng = random.Random(0)
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size, rng=rng)
 
         input_mask = None
         if self.use_input_mask:
@@ -532,6 +547,18 @@ class GroupViTModelTest(ModelTesterMixin, unittest.TestCase):
     def test_model_common_attributes(self):
         pass
 
+    @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()