Add TFDPR (#8203)

* Create modeling_tf_dpr.py

* Add TFDPR

* Add back TFPegasus, TFMarian, TFMBart, TFBlenderBot

last commit accidentally deleted these 4 lines, so I recover them back

* Add TFDPR

* Add TFDPR

* clean up some comments, add TF input-style doc string

* Add TFDPR

* Make return_dict=False as default

* Fix return_dict bug (in .from_pretrained)

* Add get_input_embeddings()

* Create test_modeling_tf_dpr.py

The current version is already passed all 27 tests!
Please see the test run at : 
https://colab.research.google.com/drive/1czS_m9zy5k-iSJbzA_DP1k1xAAC_sdkf?usp=sharing



* fix quality

* delete init weights

* run fix copies

* fix repo consis

* del config_class, load_tf_weights

They shoud be 'pytorch only'

* add config_class back

after removing it, test failed ... so totally only removing "use_tf_weights = None" on Lysandre suggestion

* newline after .. note::

* import tf, np (Necessary for ModelIntegrationTest)

* slow_test from_pretrained with from_pt=True

At the moment we don't have TF weights (since we don't have official official TF model)
Previously, I did not run slow test, so I missed this bug

* Add simple TFDPRModelIntegrationTest

Note that this is just a test that TF and Pytorch gives approx. the same output.
However, I could not test with the official DPR repo's output yet

* upload correct tf model

* remove position_ids as missing keys
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: patrickvonplaten <patrick@huggingface.co>

docs/source/model_doc/dpr.rst

@@ -99,3 +99,22 @@ DPRReader
 
 .. autoclass:: transformers.DPRReader
     :members: forward
+
+TFDPRContextEncoder
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFDPRContextEncoder
+    :members: call
+
+TFDPRQuestionEncoder
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFDPRQuestionEncoder
+    :members: call
+
+
+TFDPRReader
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.TFDPRReader
+    :members: call

src/transformers/__init__.py

@@ -406,6 +406,9 @@ if is_torch_available():
         DistilBertPreTrainedModel,
     )
     from .modeling_dpr import (
+        DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
+        DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
+        DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST,
         DPRContextEncoder,
         DPRPretrainedContextEncoder,
         DPRPretrainedQuestionEncoder,
@@ -713,6 +716,17 @@ if is_tf_available():
         TFDistilBertModel,
         TFDistilBertPreTrainedModel,
     )
+    from .modeling_tf_dpr import (
+        TF_DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TFDPRContextEncoder,
+        TFDPRPretrainedContextEncoder,
+        TFDPRPretrainedQuestionEncoder,
+        TFDPRPretrainedReader,
+        TFDPRQuestionEncoder,
+        TFDPRReader,
+    )
     from .modeling_tf_electra import (
         TF_ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST,
         TFElectraForMaskedLM,

src/transformers/convert_pytorch_checkpoint_to_tf2.py

@@ -25,6 +25,9 @@ from transformers import (
     CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP,
     DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
+    DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
+    DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
+    DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST,
     ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP,
     FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -43,6 +46,7 @@ from transformers import (
     CamembertConfig,
     CTRLConfig,
     DistilBertConfig,
+    DPRConfig,
     ElectraConfig,
     FlaubertConfig,
     GPT2Config,
@@ -59,6 +63,9 @@ from transformers import (
     TFCTRLLMHeadModel,
     TFDistilBertForMaskedLM,
     TFDistilBertForQuestionAnswering,
+    TFDPRContextEncoder,
+    TFDPRQuestionEncoder,
+    TFDPRReader,
     TFElectraForPreTraining,
     TFFlaubertWithLMHeadModel,
     TFGPT2LMHeadModel,
@@ -98,6 +105,9 @@ if is_torch_available():
         CTRLLMHeadModel,
         DistilBertForMaskedLM,
         DistilBertForQuestionAnswering,
+        DPRContextEncoder,
+        DPRQuestionEncoder,
+        DPRReader,
         ElectraForPreTraining,
         FlaubertWithLMHeadModel,
         GPT2LMHeadModel,
@@ -147,6 +157,18 @@ MODEL_CLASSES = {
         BertForSequenceClassification,
         BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
     ),
+    "dpr": (
+        DPRConfig,
+        TFDPRQuestionEncoder,
+        TFDPRContextEncoder,
+        TFDPRReader,
+        DPRQuestionEncoder,
+        DPRContextEncoder,
+        DPRReader,
+        DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
+        DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
+        DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST,
+    ),
     "gpt2": (
         GPT2Config,
         TFGPT2LMHeadModel,

src/transformers/modeling_tf_auto.py

@@ -43,6 +43,7 @@ from .configuration_auto import (
     replace_list_option_in_docstrings,
 )
 from .configuration_blenderbot import BlenderbotConfig
+from .configuration_dpr import DPRConfig
 from .configuration_marian import MarianConfig
 from .configuration_mbart import MBartConfig
 from .configuration_pegasus import PegasusConfig
@@ -87,6 +88,7 @@ from .modeling_tf_distilbert import (
     TFDistilBertForTokenClassification,
     TFDistilBertModel,
 )
+from .modeling_tf_dpr import TFDPRQuestionEncoder
 from .modeling_tf_electra import (
     TFElectraForMaskedLM,
     TFElectraForMultipleChoice,
@@ -192,6 +194,7 @@ TF_MODEL_MAPPING = OrderedDict(
         (CTRLConfig, TFCTRLModel),
         (ElectraConfig, TFElectraModel),
         (FunnelConfig, TFFunnelModel),
+        (DPRConfig, TFDPRQuestionEncoder),
     ]
 )
 

src/transformers/utils/dummy_pt_objects.py

@@ -735,6 +735,15 @@ class DistilBertPreTrainedModel:
         requires_pytorch(self)
 
 
+DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
 class DPRContextEncoder:
     def __init__(self, *args, **kwargs):
         requires_pytorch(self)

src/transformers/utils/dummy_tf_objects.py

@@ -495,6 +495,45 @@ class TFDistilBertPreTrainedModel:
         requires_tf(self)
 
 
+TF_DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+TF_DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+TF_DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class TFDPRContextEncoder:
+    def __init__(self, *args, **kwargs):
+        requires_tf(self)
+
+
+class TFDPRPretrainedContextEncoder:
+    def __init__(self, *args, **kwargs):
+        requires_tf(self)
+
+
+class TFDPRPretrainedQuestionEncoder:
+    def __init__(self, *args, **kwargs):
+        requires_tf(self)
+
+
+class TFDPRPretrainedReader:
+    def __init__(self, *args, **kwargs):
+        requires_tf(self)
+
+
+class TFDPRQuestionEncoder:
+    def __init__(self, *args, **kwargs):
+        requires_tf(self)
+
+
+class TFDPRReader:
+    def __init__(self, *args, **kwargs):
+        requires_tf(self)
+
+
 TF_ELECTRA_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

tests/test_modeling_dpr.py

@@ -24,6 +24,8 @@ from .test_modeling_common import ModelTesterMixin, ids_tensor, random_attention
 
 
 if is_torch_available():
+    import torch
+
     from transformers import BertConfig, DPRConfig, DPRContextEncoder, DPRQuestionEncoder, DPRReader
     from transformers.modeling_dpr import (
         DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
@@ -227,3 +229,36 @@ class DPRModelTest(ModelTesterMixin, unittest.TestCase):
         for model_name in DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
             model = DPRReader.from_pretrained(model_name)
             self.assertIsNotNone(model)
+
+
+@require_torch
+class DPRModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference_no_head(self):
+        model = DPRQuestionEncoder.from_pretrained("facebook/dpr-question_encoder-single-nq-base", return_dict=False)
+        model.to(torch_device)
+
+        input_ids = torch.tensor(
+            [[101, 7592, 1010, 2003, 2026, 3899, 10140, 1029, 102]], dtype=torch.long, device=torch_device
+        )  # [CLS] hello, is my dog cute? [SEP]
+        output = model(input_ids)[0]  # embedding shape = (1, 768)
+        # compare the actual values for a slice.
+        expected_slice = torch.tensor(
+            [
+                [
+                    0.03236253,
+                    0.12753335,
+                    0.16818509,
+                    0.00279786,
+                    0.3896933,
+                    0.24264945,
+                    0.2178971,
+                    -0.02335227,
+                    -0.08481959,
+                    -0.14324117,
+                ]
+            ],
+            dtype=torch.float,
+            device=torch_device,
+        )
+        self.assertTrue(torch.allclose(output[:, :10], expected_slice, atol=1e-4))

tests/test_modeling_tf_dpr.py

+# coding=utf-8
+# Copyright 2020 Huggingface
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import unittest
+
+from transformers import is_tf_available
+from transformers.testing_utils import require_tf, slow
+
+from .test_configuration_common import ConfigTester
+from .test_modeling_tf_common import TFModelTesterMixin, ids_tensor
+
+
+if is_tf_available():
+    import numpy
+    import tensorflow as tf
+
+    from transformers import (
+        TF_DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST,
+        TF_DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST,
+        BertConfig,
+        DPRConfig,
+        TFDPRContextEncoder,
+        TFDPRQuestionEncoder,
+        TFDPRReader,
+    )
+
+
+class TFDPRModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_token_type_ids=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=16,
+        type_sequence_label_size=2,
+        initializer_range=0.02,
+        num_labels=3,
+        num_choices=4,
+        scope=None,
+        projection_dim=0,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_token_type_ids = use_token_type_ids
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.num_labels = num_labels
+        self.num_choices = num_choices
+        self.scope = scope
+        self.projection_dim = projection_dim
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = ids_tensor(
+                [self.batch_size, self.seq_length], vocab_size=2
+            )  # follow test_modeling_tf_ctrl.py
+
+        token_type_ids = None
+        if self.use_token_type_ids:
+            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+        sequence_labels = None
+        token_labels = None
+        choice_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+        config = BertConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            type_vocab_size=self.type_vocab_size,
+            is_decoder=False,
+            initializer_range=self.initializer_range,
+            # MODIFY
+            return_dict=False,
+        )
+        config = DPRConfig(projection_dim=self.projection_dim, **config.to_dict())
+
+        return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+    def create_and_check_dpr_context_encoder(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = TFDPRContextEncoder(config=config)
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+        result = model(input_ids, token_type_ids=token_type_ids)
+        result = model(input_ids, return_dict=True)  # MODIFY
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.projection_dim or self.hidden_size))
+
+    def create_and_check_dpr_question_encoder(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = TFDPRQuestionEncoder(config=config)
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+        result = model(input_ids, token_type_ids=token_type_ids)
+        result = model(input_ids, return_dict=True)  # MODIFY
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.projection_dim or self.hidden_size))
+
+    def create_and_check_dpr_reader(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = TFDPRReader(config=config)
+        result = model(input_ids, attention_mask=input_mask, return_dict=True)  # MODIFY
+
+        self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
+        self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
+        self.parent.assertEqual(result.relevance_logits.shape, (self.batch_size,))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = config_and_inputs
+        inputs_dict = {"input_ids": input_ids}
+        return config, inputs_dict
+
+
+@require_tf
+class TFDPRModelTest(TFModelTesterMixin, unittest.TestCase):
+
+    all_model_classes = (
+        (
+            TFDPRContextEncoder,
+            TFDPRQuestionEncoder,
+            TFDPRReader,
+        )
+        if is_tf_available()
+        else ()
+    )
+
+    test_resize_embeddings = False
+    test_missing_keys = False
+    test_pruning = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = TFDPRModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=DPRConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_dpr_context_encoder_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_dpr_context_encoder(*config_and_inputs)
+
+    def test_dpr_question_encoder_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_dpr_question_encoder(*config_and_inputs)
+
+    def test_dpr_reader_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_dpr_reader(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in TF_DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = TFDPRContextEncoder.from_pretrained(model_name, from_pt=True)
+            self.assertIsNotNone(model)
+
+        for model_name in TF_DPR_CONTEXT_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = TFDPRContextEncoder.from_pretrained(model_name, from_pt=True)
+            self.assertIsNotNone(model)
+
+        for model_name in TF_DPR_QUESTION_ENCODER_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = TFDPRQuestionEncoder.from_pretrained(model_name, from_pt=True)
+            self.assertIsNotNone(model)
+
+        for model_name in TF_DPR_READER_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = TFDPRReader.from_pretrained(model_name, from_pt=True)
+            self.assertIsNotNone(model)
+
+
+@require_tf
+class TFDPRModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference_no_head(self):
+        model = TFDPRQuestionEncoder.from_pretrained("facebook/dpr-question_encoder-single-nq-base", return_dict=False)
+
+        input_ids = tf.constant(
+            [[101, 7592, 1010, 2003, 2026, 3899, 10140, 1029, 102]]
+        )  # [CLS] hello, is my dog cute? [SEP]
+        output = model(input_ids)[0]  # embedding shape = (1, 768)
+        # compare the actual values for a slice.
+        expected_slice = tf.constant(
+            [
+                [
+                    0.03236253,
+                    0.12753335,
+                    0.16818509,
+                    0.00279786,
+                    0.3896933,
+                    0.24264945,
+                    0.2178971,
+                    -0.02335227,
+                    -0.08481959,
+                    -0.14324117,
+                ]
+            ]
+        )
+        self.assertTrue(numpy.allclose(output[:, :10].numpy(), expected_slice.numpy(), atol=1e-4))

utils/check_repo.py

@@ -33,6 +33,8 @@ IGNORE_NON_TESTED = [
     "DPRSpanPredictor",  # Building part of bigger (tested) model.
     "ReformerForMaskedLM",  # Needs to be setup as decoder.
     "T5Stack",  # Building part of bigger (tested) model.
+    "TFDPREncoder",  # Building part of bigger (tested) model.
+    "TFDPRSpanPredictor",  # Building part of bigger (tested) model.
     "TFElectraMainLayer",  # Building part of bigger (tested) model (should it be a TFPreTrainedModel ?)
     "TFRobertaForMultipleChoice",  # TODO: fix
 ]
@@ -57,6 +59,8 @@ IGNORE_NON_DOCUMENTED = [
     "DPREncoder",  # Building part of bigger (documented) model.
     "DPRSpanPredictor",  # Building part of bigger (documented) model.
     "T5Stack",  # Building part of bigger (tested) model.
+    "TFDPREncoder",  # Building part of bigger (documented) model.
+    "TFDPRSpanPredictor",  # Building part of bigger (documented) model.
     "TFElectraMainLayer",  # Building part of bigger (documented) model (should it be a TFPreTrainedModel ?)
 ]
 
@@ -87,6 +91,10 @@ IGNORE_NON_AUTO_CONFIGURED = [
     "RagSequenceForGeneration",
     "RagTokenForGeneration",
     "T5Stack",
+    "TFDPRContextEncoder",
+    "TFDPREncoder",
+    "TFDPRReader",
+    "TFDPRSpanPredictor",
     "TFFunnelBaseModel",
     "TFGPT2DoubleHeadsModel",
     "TFOpenAIGPTDoubleHeadsModel",