Add BLIP-2 (#21441)

* First draft

* More improvements

* More improvements

* Improve conversion script

* Convert all weights

* Make forward pass work

* Make logits match

* More improvements

* More improvements

* More improvements

* Use get_input_embeddings

* Improve some more

* Improve model tests

* Improve model tests

* More improvements

* Fix processor

* Update files

* Update prepare_inputs_for_generation

* More improvements

* Fix copies

* More fixes

* Make fixup

* More improvements

* Add support for seq2seq language model

* More improvements

* Fix test

* More improvements

* Improve conversion script

* Remove some todo's

* Fix README's

* Improve conversion script

* Fix generation

* Fix style and remove Blip2Model

* Fix model outputs

* More improvements

* Set eos_token_id in config

* Fix quality

* Small improvements

* Add processor tests

* More improvements

* Apply suggestions

* Apply suggestions

* Add integration test

* Update image URL

* Add integration test

* Fix model_type

* Update style

* Improve docs

* Add doc tests

* Fix copies

* Remove tests which are passing

* Improve some more

* Add tests for seq2seq language models

* Minor fix

* Convert more checkpoints

* finalize CI

* Fix blip and blip2 processors

* add `accelerate` support for `blip2`

* clean up

* make style

* Update conversion script

* Update conversion script some more

* Update organization

* revert toc file

* add blip-2 to toc file

* Some more improvements

* Fix docstring

* Improve docs

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

src/transformers/models/blip_2/processing_blip_2.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+#
+# 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.
+"""
+Processor class for BLIP-2.
+"""
+
+from typing import List, Optional, Union
+
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import BatchEncoding, PaddingStrategy, PreTokenizedInput, TextInput, TruncationStrategy
+from ...utils import TensorType
+
+
+class Blip2Processor(ProcessorMixin):
+    r"""
+    Constructs a BLIP-2 processor which wraps a BLIP image processor and an OPT/T5 tokenizer into a single processor.
+
+    [`BlipProcessor`] offers all the functionalities of [`BlipImageProcessor`] and [`AutoTokenizer`]. See the docstring
+    of [`~BlipProcessor.__call__`] and [`~BlipProcessor.decode`] for more information.
+
+    Args:
+        image_processor (`BlipImageProcessor`):
+            An instance of [`BlipImageProcessor`]. The image processor is a required input.
+        tokenizer (`AutoTokenizer`):
+            An instance of ['PreTrainedTokenizer`]. The tokenizer is a required input.
+    """
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "BlipImageProcessor"
+    tokenizer_class = "AutoTokenizer"
+
+    # Copied from transformers.models.blip.processing_blip.BlipProcessor.__init__
+    def __init__(self, image_processor, tokenizer):
+        tokenizer.return_token_type_ids = False
+        super().__init__(image_processor, tokenizer)
+        self.current_processor = self.image_processor
+
+    # Copied from transformers.models.blip.processing_blip.BlipProcessor.__call__
+    def __call__(
+        self,
+        images=None,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+        add_special_tokens: bool = True,
+        padding: Union[bool, str, PaddingStrategy] = False,
+        truncation: Union[bool, str, TruncationStrategy] = None,
+        max_length: Optional[int] = None,
+        stride: int = 0,
+        pad_to_multiple_of: Optional[int] = None,
+        return_attention_mask: Optional[bool] = None,
+        return_overflowing_tokens: bool = False,
+        return_special_tokens_mask: bool = False,
+        return_offsets_mapping: bool = False,
+        return_token_type_ids: bool = False,
+        return_length: bool = False,
+        verbose: bool = True,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        **kwargs,
+    ) -> BatchEncoding:
+        """
+        This method uses [`BlipImageProcessor.__call__`] method to prepare image(s) for the model, and
+        [`BertTokenizerFast.__call__`] to prepare text for the model.
+
+        Please refer to the docstring of the above two methods for more information.
+        """
+        if images is None and text is None:
+            raise ValueError("You have to specify either images or text.")
+
+        # Get only text
+        if images is None:
+            self.current_processor = self.tokenizer
+            text_encoding = self.tokenizer(
+                text=text,
+                add_special_tokens=add_special_tokens,
+                padding=padding,
+                truncation=truncation,
+                max_length=max_length,
+                stride=stride,
+                pad_to_multiple_of=pad_to_multiple_of,
+                return_attention_mask=return_attention_mask,
+                return_overflowing_tokens=return_overflowing_tokens,
+                return_special_tokens_mask=return_special_tokens_mask,
+                return_offsets_mapping=return_offsets_mapping,
+                return_token_type_ids=return_token_type_ids,
+                return_length=return_length,
+                verbose=verbose,
+                return_tensors=return_tensors,
+                **kwargs,
+            )
+            return text_encoding
+
+        # add pixel_values
+        encoding_image_processor = self.image_processor(images, return_tensors=return_tensors)
+
+        if text is not None:
+            text_encoding = self.tokenizer(
+                text=text,
+                add_special_tokens=add_special_tokens,
+                padding=padding,
+                truncation=truncation,
+                max_length=max_length,
+                stride=stride,
+                pad_to_multiple_of=pad_to_multiple_of,
+                return_attention_mask=return_attention_mask,
+                return_overflowing_tokens=return_overflowing_tokens,
+                return_special_tokens_mask=return_special_tokens_mask,
+                return_offsets_mapping=return_offsets_mapping,
+                return_token_type_ids=return_token_type_ids,
+                return_length=return_length,
+                verbose=verbose,
+                return_tensors=return_tensors,
+                **kwargs,
+            )
+        else:
+            text_encoding = None
+
+        if text_encoding is not None:
+            encoding_image_processor.update(text_encoding)
+
+        return encoding_image_processor
+
+    # Copied from transformers.models.blip.processing_blip.BlipProcessor.batch_decode with BertTokenizerFast->PreTrainedTokenizer
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to PreTrainedTokenizer's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    # Copied from transformers.models.blip.processing_blip.BlipProcessor.decode with BertTokenizerFast->PreTrainedTokenizer
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to PreTrainedTokenizer's [`~PreTrainedTokenizer.decode`]. Please refer
+        to the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    # Copied from transformers.models.blip.processing_blip.BlipProcessor.model_input_names
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))

src/transformers/utils/constants.py

@@ -2,3 +2,5 @@ IMAGENET_DEFAULT_MEAN = [0.485, 0.456, 0.406]
 IMAGENET_DEFAULT_STD = [0.229, 0.224, 0.225]
 IMAGENET_STANDARD_MEAN = [0.5, 0.5, 0.5]
 IMAGENET_STANDARD_STD = [0.5, 0.5, 0.5]
+OPENAI_CLIP_MEAN = [0.48145466, 0.4578275, 0.40821073]
+OPENAI_CLIP_STD = [0.26862954, 0.26130258, 0.27577711]

src/transformers/utils/dummy_pt_objects.py

@@ -1211,6 +1211,37 @@ class BlipVisionModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+BLIP_2_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class Blip2ForConditionalGeneration(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Blip2PreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Blip2QFormerModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class Blip2VisionModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 BLOOM_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

tests/models/blip_2/test_modeling_blip_2.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
+#
+# 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.
+""" Testing suite for the PyTorch BLIP-2 model. """
+
+
+import inspect
+import tempfile
+import unittest
+
+import numpy as np
+import requests
+
+from transformers import CONFIG_MAPPING, Blip2Config, Blip2QFormerConfig, Blip2VisionConfig
+from transformers.testing_utils import require_torch, require_vision, slow, torch_device
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import (
+    ModelTesterMixin,
+    _config_zero_init,
+    floats_tensor,
+    ids_tensor,
+    random_attention_mask,
+)
+
+
+if is_torch_available():
+    import torch
+    from torch import nn
+
+    from transformers import Blip2ForConditionalGeneration, Blip2VisionModel
+    from transformers.models.blip_2.modeling_blip_2 import BLIP_2_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import Blip2Processor
+
+
+class Blip2VisionModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        image_size=30,
+        patch_size=2,
+        num_channels=3,
+        is_training=True,
+        hidden_size=32,
+        projection_dim=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        initializer_range=1e-10,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_channels = num_channels
+        self.is_training = is_training
+        self.hidden_size = hidden_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.initializer_range = initializer_range
+        self.scope = scope
+
+        # in ViT, the seq length equals the number of patches + 1 (we add 1 for the [CLS] token)
+        num_patches = (image_size // patch_size) ** 2
+        self.seq_length = num_patches + 1
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+        config = self.get_config()
+
+        return config, pixel_values
+
+    def get_config(self):
+        return Blip2VisionConfig(
+            image_size=self.image_size,
+            patch_size=self.patch_size,
+            num_channels=self.num_channels,
+            hidden_size=self.hidden_size,
+            projection_dim=self.projection_dim,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(self, config, pixel_values):
+        model = Blip2VisionModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(pixel_values)
+        # expected sequence length = num_patches + 1 (we add 1 for the [CLS] token)
+        image_size = (self.image_size, self.image_size)
+        patch_size = (self.patch_size, self.patch_size)
+        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, num_patches + 1, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_torch
+class Blip2VisionModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as BLIP-2's vision encoder does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (Blip2VisionModel,) if is_torch_available() else ()
+    fx_compatible = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = Blip2VisionModelTester(self)
+        self.config_tester = ConfigTester(
+            self, config_class=Blip2VisionConfig, has_text_modality=False, hidden_size=37
+        )
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    @unittest.skip(reason="BLIP-2's vision encoder does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    def test_model_common_attributes(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            self.assertIsInstance(model.get_input_embeddings(), (nn.Module))
+            x = model.get_output_embeddings()
+            self.assertTrue(x is None or isinstance(x, nn.Linear))
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_training(self):
+        pass
+
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="Blip2VisionModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    @unittest.skip(reason="Blip2VisionModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_to_base(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in BLIP_2_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = Blip2VisionModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+class Blip2QFormerModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        projection_dim=32,
+        num_hidden_layers=6,
+        num_attention_heads=4,
+        intermediate_size=37,
+        dropout=0.1,
+        attention_dropout=0.1,
+        max_position_embeddings=512,
+        initializer_range=0.02,
+        bos_token_id=0,
+        scope=None,
+    ):
+        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_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.projection_dim = projection_dim
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.scope = scope
+        self.bos_token_id = bos_token_id
+
+    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 = random_attention_mask([self.batch_size, self.seq_length])
+
+        if input_mask is not None:
+            batch_size, seq_length = input_mask.shape
+            rnd_start_indices = np.random.randint(1, seq_length - 1, size=(batch_size,))
+            for batch_idx, start_index in enumerate(rnd_start_indices):
+                input_mask[batch_idx, :start_index] = 1
+                input_mask[batch_idx, start_index:] = 0
+
+        config = self.get_config()
+
+        return config, input_ids, input_mask
+
+    def get_config(self):
+        return Blip2QFormerConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            projection_dim=self.projection_dim,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            dropout=self.dropout,
+            attention_dropout=self.attention_dropout,
+            max_position_embeddings=self.max_position_embeddings,
+            initializer_range=self.initializer_range,
+            bos_token_id=self.bos_token_id,
+        )
+
+
+# this class is based on `OPTModelTester` found in tests/models/opt/test_modeling_opt.py
+class Blip2TextModelDecoderOnlyTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=12,
+        seq_length=7,
+        is_training=True,
+        use_labels=False,
+        vocab_size=99,
+        hidden_size=16,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=4,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=20,
+        eos_token_id=2,
+        pad_token_id=1,
+        bos_token_id=0,
+        embed_dim=16,
+        num_labels=3,
+        word_embed_proj_dim=16,
+        type_sequence_label_size=2,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        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.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.embed_dim = embed_dim
+        self.num_labels = num_labels
+        self.type_sequence_label_size = type_sequence_label_size
+        self.word_embed_proj_dim = word_embed_proj_dim
+        self.is_encoder_decoder = False
+
+    def prepare_config_and_inputs(self):
+        config = self.get_config()
+
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(
+            3,
+        )
+        input_ids[:, -1] = self.eos_token_id  # Eos Token
+
+        attention_mask = input_ids.ne(self.pad_token_id)
+
+        return config, input_ids, attention_mask
+
+    def get_config(self):
+        return CONFIG_MAPPING["opt"](
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            ffn_dim=self.intermediate_size,
+            dropout=self.hidden_dropout_prob,
+            attention_dropout=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.bos_token_id,
+            pad_token_id=self.pad_token_id,
+            embed_dim=self.embed_dim,
+            is_encoder_decoder=False,
+            word_embed_proj_dim=self.word_embed_proj_dim,
+        )
+
+
+# this model tester uses a decoder-only language model (OPT)
+class Blip2ForConditionalGenerationDecoderOnlyModelTester:
+    def __init__(
+        self, parent, vision_kwargs=None, qformer_kwargs=None, text_kwargs=None, is_training=True, num_query_tokens=10
+    ):
+        if vision_kwargs is None:
+            vision_kwargs = {}
+        if qformer_kwargs is None:
+            qformer_kwargs = {}
+        if text_kwargs is None:
+            text_kwargs = {}
+
+        self.parent = parent
+        self.vision_model_tester = Blip2VisionModelTester(parent, **vision_kwargs)
+        self.qformer_model_tester = Blip2QFormerModelTester(parent, **qformer_kwargs)
+        self.text_model_tester = Blip2TextModelDecoderOnlyTester(parent, **text_kwargs)
+        self.is_training = is_training
+        self.num_query_tokens = num_query_tokens
+
+    def prepare_config_and_inputs(self):
+        _, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
+        _, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
+
+        config = self.get_config()
+
+        return config, input_ids, attention_mask, pixel_values
+
+    def get_config(self):
+        return Blip2Config.from_vision_qformer_text_configs(
+            vision_config=self.vision_model_tester.get_config(),
+            qformer_config=self.qformer_model_tester.get_config(),
+            text_config=self.text_model_tester.get_config(),
+            num_query_tokens=self.num_query_tokens,
+        )
+
+    def create_and_check_for_conditional_generation(self, config, input_ids, attention_mask, pixel_values):
+        model = Blip2ForConditionalGeneration(config).to(torch_device).eval()
+        with torch.no_grad():
+            result = model(pixel_values, input_ids, attention_mask)
+
+        expected_seq_length = self.num_query_tokens + self.text_model_tester.seq_length
+        self.parent.assertEqual(
+            result.logits.shape,
+            (self.vision_model_tester.batch_size, expected_seq_length, self.text_model_tester.vocab_size),
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, attention_mask, pixel_values = config_and_inputs
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "labels": input_ids,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class Blip2ForConditionalGenerationDecoderOnlyTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (Blip2ForConditionalGeneration,) if is_torch_available() else ()
+    fx_compatible = False
+    test_head_masking = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_attention_outputs = False
+    test_torchscript = False
+
+    def setUp(self):
+        self.model_tester = Blip2ForConditionalGenerationDecoderOnlyModelTester(self)
+
+    def test_for_conditional_generation(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_conditional_generation(*config_and_inputs)
+
+    @unittest.skip(reason="Hidden_states is tested in individual model tests")
+    def test_hidden_states_output(self):
+        pass
+
+    @unittest.skip(reason="Inputs_embeds is tested in individual model tests")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Retain_grad is tested in individual model tests")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    @unittest.skip(reason="Blip2Model does not have input/output embeddings")
+    def test_model_common_attributes(self):
+        pass
+
+    @unittest.skip(reason="There's no base Blip2Model")
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    @unittest.skip(reason="There's no base Blip2Model")
+    def test_save_load_fast_init_to_base(self):
+        pass
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_load_vision_qformer_text_config(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # Save Blip2Config and check if we can load Blip2VisionConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            vision_config = Blip2VisionConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
+
+        # Save Blip2Config and check if we can load Blip2QFormerConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            qformer_config = Blip2QFormerConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.qformer_config.to_dict(), qformer_config.to_dict())
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in BLIP_2_PRETRAINED_MODEL_ARCHIVE_LIST:
+            model = Blip2ForConditionalGeneration.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+# this class is based on `T5ModelTester` found in tests/models/t5/test_modeling_t5.py
+class Blip2TextModelTester:
+    def __init__(
+        self,
+        parent,
+        vocab_size=99,
+        batch_size=12,
+        encoder_seq_length=7,
+        decoder_seq_length=9,
+        # For common tests
+        is_training=True,
+        use_attention_mask=True,
+        use_labels=True,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        d_ff=37,
+        relative_attention_num_buckets=8,
+        dropout_rate=0.1,
+        initializer_factor=0.002,
+        eos_token_id=1,
+        pad_token_id=0,
+        decoder_start_token_id=0,
+        scope=None,
+        decoder_layers=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.encoder_seq_length = encoder_seq_length
+        self.decoder_seq_length = decoder_seq_length
+        # For common tests
+        self.seq_length = self.decoder_seq_length
+        self.is_training = is_training
+        self.use_attention_mask = use_attention_mask
+        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.d_ff = d_ff
+        self.relative_attention_num_buckets = relative_attention_num_buckets
+        self.dropout_rate = dropout_rate
+        self.initializer_factor = initializer_factor
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.decoder_start_token_id = decoder_start_token_id
+        self.scope = None
+        self.decoder_layers = decoder_layers
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.encoder_seq_length], self.vocab_size)
+        decoder_input_ids = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        attention_mask = None
+        decoder_attention_mask = None
+        if self.use_attention_mask:
+            attention_mask = ids_tensor([self.batch_size, self.encoder_seq_length], vocab_size=2)
+            decoder_attention_mask = ids_tensor([self.batch_size, self.decoder_seq_length], vocab_size=2)
+
+        lm_labels = None
+        if self.use_labels:
+            lm_labels = ids_tensor([self.batch_size, self.decoder_seq_length], self.vocab_size)
+
+        config = self.get_config()
+
+        return (
+            config,
+            input_ids,
+            decoder_input_ids,
+            attention_mask,
+            decoder_attention_mask,
+            lm_labels,
+        )
+
+    def get_config(self):
+        return CONFIG_MAPPING["t5"](
+            vocab_size=self.vocab_size,
+            d_model=self.hidden_size,
+            d_ff=self.d_ff,
+            d_kv=self.hidden_size // self.num_attention_heads,
+            num_layers=self.num_hidden_layers,
+            num_decoder_layers=self.decoder_layers,
+            num_heads=self.num_attention_heads,
+            relative_attention_num_buckets=self.relative_attention_num_buckets,
+            dropout_rate=self.dropout_rate,
+            initializer_factor=self.initializer_factor,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.pad_token_id,
+            pad_token_id=self.pad_token_id,
+            decoder_start_token_id=self.decoder_start_token_id,
+        )
+
+
+# this model tester uses an encoder-decoder language model (T5)
+class Blip2ForConditionalGenerationModelTester:
+    def __init__(
+        self, parent, vision_kwargs=None, qformer_kwargs=None, text_kwargs=None, is_training=True, num_query_tokens=10
+    ):
+        if vision_kwargs is None:
+            vision_kwargs = {}
+        if qformer_kwargs is None:
+            qformer_kwargs = {}
+        if text_kwargs is None:
+            text_kwargs = {}
+
+        self.parent = parent
+        self.vision_model_tester = Blip2VisionModelTester(parent, **vision_kwargs)
+        self.qformer_model_tester = Blip2QFormerModelTester(parent, **qformer_kwargs)
+        self.text_model_tester = Blip2TextModelTester(parent, **text_kwargs)
+        self.is_training = is_training
+        self.num_query_tokens = num_query_tokens
+
+    def prepare_config_and_inputs(self):
+        _, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
+        (
+            _,
+            input_ids,
+            decoder_input_ids,
+            attention_mask,
+            decoder_attention_mask,
+            lm_labels,
+        ) = self.text_model_tester.prepare_config_and_inputs()
+
+        config = self.get_config()
+
+        return config, input_ids, attention_mask, pixel_values, decoder_input_ids, decoder_attention_mask, lm_labels
+
+    def get_config(self):
+        return Blip2Config.from_vision_qformer_text_configs(
+            vision_config=self.vision_model_tester.get_config(),
+            qformer_config=self.qformer_model_tester.get_config(),
+            text_config=self.text_model_tester.get_config(),
+            num_query_tokens=self.num_query_tokens,
+        )
+
+    def create_and_check_for_conditional_generation(
+        self, config, input_ids, attention_mask, pixel_values, decoder_input_ids, decoder_attention_mask, labels
+    ):
+        model = Blip2ForConditionalGeneration(config).to(torch_device).eval()
+        with torch.no_grad():
+            result = model(pixel_values, input_ids, attention_mask, decoder_input_ids, decoder_attention_mask)
+
+        self.parent.assertEqual(
+            result.logits.shape,
+            (
+                self.vision_model_tester.batch_size,
+                self.text_model_tester.seq_length,
+                self.text_model_tester.vocab_size,
+            ),
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            attention_mask,
+            pixel_values,
+            decoder_input_ids,
+            decoder_attention_mask,
+            labels,
+        ) = config_and_inputs
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "decoder_input_ids": decoder_input_ids,
+            "decoder_attention_mask": decoder_attention_mask,
+            "labels": labels,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class Blip2ForConditionalGenerationTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (Blip2ForConditionalGeneration,) if is_torch_available() else ()
+    fx_compatible = False
+    test_head_masking = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_attention_outputs = False
+    test_torchscript = False
+
+    def setUp(self):
+        self.model_tester = Blip2ForConditionalGenerationModelTester(self)
+
+    def test_for_conditional_generation(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_conditional_generation(*config_and_inputs)
+
+    @unittest.skip(reason="Hidden_states is tested in individual model tests")
+    def test_hidden_states_output(self):
+        pass
+
+    @unittest.skip(reason="Inputs_embeds is tested in individual model tests")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Retain_grad is tested in individual model tests")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    @unittest.skip(reason="Blip2Model does not have input/output embeddings")
+    def test_model_common_attributes(self):
+        pass
+
+    @unittest.skip(reason="There's no base Blip2Model")
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    @unittest.skip(reason="There's no base Blip2Model")
+    def test_save_load_fast_init_to_base(self):
+        pass
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_load_vision_qformer_text_config(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # Save Blip2Config and check if we can load Blip2VisionConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            vision_config = Blip2VisionConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
+
+        # Save Blip2Config and check if we can load Blip2QFormerConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            qformer_config = Blip2QFormerConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.qformer_config.to_dict(), qformer_config.to_dict())
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in BLIP_2_PRETRAINED_MODEL_ARCHIVE_LIST:
+            model = Blip2ForConditionalGeneration.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+    # override from common to deal with nested configurations (`vision_config`, `text_config` and `qformer_config`)
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for key in ["vision_config", "qformer_config", "text_config"]:
+            setattr(configs_no_init, key, _config_zero_init(getattr(configs_no_init, key)))
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    self.assertIn(
+                        ((param.data.mean() * 1e9).round() / 1e9).item(),
+                        [0.0, 1.0],
+                        msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                    )
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "https://huggingface.co/hf-internal-testing/blip-test-image/resolve/main/demo.jpg"
+    image = Image.open(requests.get(url, stream=True).raw)
+    return image
+
+
+@require_vision
+@require_torch
+@slow
+class Blip2ModelIntegrationTest(unittest.TestCase):
+    def test_inference_opt(self):
+        processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
+        model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-2.7b").to(torch_device)
+
+        # prepare image
+        image = prepare_img()
+        inputs = processor(images=image, return_tensors="pt").to(torch_device)
+
+        predictions = model.generate(**inputs)
+        generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
+
+        # Test output
+        self.assertEqual(predictions[0].tolist(), [2, 102, 693, 2828, 15, 5, 4105, 19, 10, 2335, 50118])
+        self.assertEqual("a woman sitting on the beach with a dog", generated_text)
+
+        # image and context
+        prompt = "Question: which city is this? Answer:"
+        inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device)
+
+        predictions = model.generate(**inputs)
+        generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
+
+        # Test output
+        self.assertEqual(
+            predictions[0].tolist(),
+            [2, 24, 18, 45, 10, 343, 6, 24, 18, 10, 4105, 50118],
+        )
+        self.assertEqual(generated_text, "it's not a city, it's a beach")
+
+    def test_inference_t5(self):
+        processor = Blip2Processor.from_pretrained("Salesforce/blip2-flan-t5-xl")
+        model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-flan-t5-xl").to(torch_device)
+
+        # prepare image
+        image = prepare_img()
+        inputs = processor(images=image, return_tensors="pt").to(torch_device)
+
+        predictions = model.generate(**inputs)
+        generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
+
+        # Test output
+        self.assertEqual(predictions[0].tolist(), [0, 2335, 1556, 28, 1782, 30, 8, 2608, 1])
+        self.assertEqual("woman playing with dog on the beach", generated_text)
+
+        # image and context
+        prompt = "Question: which city is this? Answer:"
+        inputs = processor(images=image, text=prompt, return_tensors="pt").to(torch_device)
+
+        predictions = model.generate(**inputs)
+        generated_text = processor.batch_decode(predictions, skip_special_tokens=True)[0].strip()
+
+        # Test output
+        self.assertEqual(
+            predictions[0].tolist(),
+            [0, 3, 7, 152, 67, 839, 1],
+        )
+        self.assertEqual(generated_text, "san diego")

tests/models/blip_2/test_processor_blip_2.py

+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# 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 shutil
+import tempfile
+import unittest
+
+import numpy as np
+import pytest
+
+from transformers.testing_utils import require_vision
+from transformers.utils import is_vision_available
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import AutoProcessor, Blip2Processor, BlipImageProcessor, GPT2Tokenizer, PreTrainedTokenizerFast
+
+
+@require_vision
+class Blip2ProcessorTest(unittest.TestCase):
+    def setUp(self):
+        self.tmpdirname = tempfile.mkdtemp()
+
+        image_processor = BlipImageProcessor()
+        tokenizer = GPT2Tokenizer.from_pretrained("hf-internal-testing/tiny-random-GPT2Model")
+
+        processor = Blip2Processor(image_processor, tokenizer)
+
+        processor.save_pretrained(self.tmpdirname)
+
+    def get_tokenizer(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).tokenizer
+
+    def get_image_processor(self, **kwargs):
+        return AutoProcessor.from_pretrained(self.tmpdirname, **kwargs).image_processor
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
+    def prepare_image_inputs(self):
+        """This function prepares a list of PIL images, or a list of numpy arrays if one specifies numpify=True,
+        or a list of PyTorch tensors if one specifies torchify=True.
+        """
+
+        image_inputs = [np.random.randint(255, size=(3, 30, 400), dtype=np.uint8)]
+
+        image_inputs = [Image.fromarray(np.moveaxis(x, 0, -1)) for x in image_inputs]
+
+        return image_inputs
+
+    def test_save_load_pretrained_additional_features(self):
+        processor = Blip2Processor(tokenizer=self.get_tokenizer(), image_processor=self.get_image_processor())
+        processor.save_pretrained(self.tmpdirname)
+
+        tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
+        image_processor_add_kwargs = self.get_image_processor(do_normalize=False, padding_value=1.0)
+
+        processor = Blip2Processor.from_pretrained(
+            self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
+        )
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
+        self.assertIsInstance(processor.tokenizer, PreTrainedTokenizerFast)
+
+        self.assertEqual(processor.image_processor.to_json_string(), image_processor_add_kwargs.to_json_string())
+        self.assertIsInstance(processor.image_processor, BlipImageProcessor)
+
+    def test_image_processor(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = Blip2Processor(tokenizer=tokenizer, 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")
+
+        for key in input_feat_extract.keys():
+            self.assertAlmostEqual(input_feat_extract[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_tokenizer(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = Blip2Processor(tokenizer=tokenizer, image_processor=image_processor)
+
+        input_str = "lower newer"
+
+        encoded_processor = processor(text=input_str)
+
+        encoded_tok = tokenizer(input_str, return_token_type_ids=False)
+
+        for key in encoded_tok.keys():
+            self.assertListEqual(encoded_tok[key], encoded_processor[key])
+
+    def test_processor(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = Blip2Processor(tokenizer=tokenizer, image_processor=image_processor)
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(text=input_str, images=image_input)
+
+        self.assertListEqual(list(inputs.keys()), ["pixel_values", "input_ids", "attention_mask"])
+
+        # test if it raises when no input is passed
+        with pytest.raises(ValueError):
+            processor()
+
+    def test_tokenizer_decode(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = Blip2Processor(tokenizer=tokenizer, image_processor=image_processor)
+
+        predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
+
+        decoded_processor = processor.batch_decode(predicted_ids)
+        decoded_tok = tokenizer.batch_decode(predicted_ids)
+
+        self.assertListEqual(decoded_tok, decoded_processor)
+
+    def test_model_input_names(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = Blip2Processor(tokenizer=tokenizer, image_processor=image_processor)
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(text=input_str, images=image_input)
+
+        # For now the processor supports only ['pixel_values', 'input_ids', 'attention_mask']
+        self.assertListEqual(list(inputs.keys()), ["pixel_values", "input_ids", "attention_mask"])

utils/check_repo.py

@@ -53,6 +53,7 @@ PRIVATE_MODELS = [
 # Being in this list is an exception and should **not** be the rule.
 IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
     # models to ignore for not tested
+    "Blip2QFormerModel",  # Building part of bigger (tested) model.
     "DetaEncoder",  # Building part of bigger (tested) model.
     "DetaDecoder",  # Building part of bigger (tested) model.
     "GraphormerEncoder",  # Building part of bigger (tested) model.
@@ -171,6 +172,9 @@ TEST_FILES_WITH_NO_COMMON_TESTS = [
 # should **not** be the rule.
 IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
     # models to ignore for model xxx mapping
+    "Blip2ForConditionalGeneration",
+    "Blip2QFormerModel",
+    "Blip2VisionModel",
     "GitVisionModel",
     "GraphormerModel",
     "GraphormerForGraphClassification",

utils/documentation_tests.txt

@@ -35,6 +35,7 @@ src/transformers/models/blenderbot/configuration_blenderbot.py
 src/transformers/models/blenderbot/modeling_blenderbot.py
 src/transformers/models/blenderbot_small/configuration_blenderbot_small.py
 src/transformers/models/blenderbot_small/modeling_blenderbot_small.py
+src/transformers/models/blip_2/modeling_blip_2.py
 src/transformers/models/blip/modeling_blip.py
 src/transformers/models/bloom/configuration_bloom.py
 src/transformers/models/camembert/configuration_camembert.py