`blip` support for training (#21021)

* `blip` support for training

* remove labels creation

* remove unneeded `decoder_input_ids` creation

* final changes

- add colab link to documentation
- reduction = mean for loss

* fix nits

* update link

* clearer error message

docs/source/en/model_doc/blip.mdx

@@ -31,6 +31,10 @@ However, most existing pre-trained models only excel in either understanding-bas
 This model was contributed by [ybelkada](https://huggingface.co/ybelkada).
 The original code can be found [here](https://github.com/salesforce/BLIP).
 
+## Resources
+
+- [Jupyter notebook](https://github.com/huggingface/notebooks/blob/main/examples/image_captioning_blip.ipynb) on how to fine-tune BLIP for image captioning on a custom dataset
+
 
 ## BlipConfig
 

src/transformers/models/blip/modeling_blip.py

@@ -1014,6 +1014,7 @@ class BlipForConditionalGeneration(BlipPreTrainedModel):
             encoder_hidden_states=image_embeds,
             labels=labels,
             return_dict=return_dict,
+            reduction="mean",
         )
 
         if not return_dict:
@@ -1125,7 +1126,7 @@ class BlipForQuestionAnswering(BlipPreTrainedModel):
         self.text_decoder = BlipTextLMHeadModel(config.text_config)
 
         self.decoder_pad_token_id = config.text_config.pad_token_id
-        self.decoder_bos_token_id = config.text_config.bos_token_id
+        self.decoder_start_token_id = config.text_config.bos_token_id
 
         # Initialize weights and apply final processing
         self.post_init()
@@ -1133,6 +1134,19 @@ class BlipForQuestionAnswering(BlipPreTrainedModel):
     def get_input_embeddings(self) -> nn.Module:
         return self.vision_model.embeddings.patch_embedding
 
+    # Adapted from transformers.models.t5.modeling_t5.T5PreTrainedModel._shift_right
+    def _shift_right(self, input_ids):
+        pad_token_id = self.decoder_pad_token_id
+
+        shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+        shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
+        shifted_input_ids[..., 0] = self.decoder_start_token_id
+
+        # replace possible -100 values in labels by `pad_token_id`
+        shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+        return shifted_input_ids
+
     @add_start_docstrings_to_model_forward(BLIP_VISION_INPUTS_DOCSTRING)
     @replace_return_docstrings(output_type=BlipTextVisionModelOutput, config_class=BlipVisionConfig)
     def forward(
@@ -1168,8 +1182,14 @@ class BlipForQuestionAnswering(BlipPreTrainedModel):
 
         >>> outputs = model(**inputs)
         ```"""
+        if labels is None and decoder_input_ids is None:
+            raise ValueError(
+                "Either `decoder_input_ids` or `labels` should be passed when calling `forward` with"
+                " `BlipForQuestionAnswering`. if you are training the model make sure that `labels` is passed, if you"
+                " are using the model for inference make sure that `decoder_input_ids` is passed."
+            )
+
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-        batch_size = input_ids.shape[0]
 
         vision_outputs = self.vision_model(
             pixel_values=pixel_values,
@@ -1191,11 +1211,11 @@ class BlipForQuestionAnswering(BlipPreTrainedModel):
 
         question_embeds = question_embeds[0] if not return_dict else question_embeds.last_hidden_state
 
-        if decoder_input_ids is None:
-            decoder_input_ids = torch.LongTensor([self.decoder_bos_token_id]).repeat((batch_size, 1))
-
-        if labels is None:
-            labels = decoder_input_ids.masked_fill(decoder_input_ids == self.decoder_pad_token_id, -100)
+        if labels is not None and decoder_input_ids is None:
+            # get decoder inputs from shifting lm labels to the right - this is used in training mode
+            decoder_input_ids = self._shift_right(labels)
+            # replace possible -100 values in labels by `pad_token_id`
+            labels = labels.masked_fill(labels == self.decoder_pad_token_id, -100)
 
         answer_output = self.text_decoder(
             input_ids=decoder_input_ids,
@@ -1204,10 +1224,13 @@ class BlipForQuestionAnswering(BlipPreTrainedModel):
             encoder_attention_mask=attention_mask,
             labels=labels,
             return_dict=return_dict,
-            reduction="none",
+            reduction="mean",
         )
 
+        if labels is not None:
             decoder_loss = answer_output.loss.mean() if return_dict else answer_output[0].mean()
+        else:
+            decoder_loss = None
 
         if not return_dict:
             outputs = (decoder_loss, image_embeds, vision_outputs[0]) + vision_outputs[2:]
@@ -1288,7 +1311,7 @@ class BlipForQuestionAnswering(BlipPreTrainedModel):
         question_attention_mask = torch.ones(question_embeds.size()[:-1], dtype=torch.long).to(question_embeds.device)
 
         bos_ids = torch.full(
-            (question_embeds.size(0), 1), fill_value=self.decoder_bos_token_id, device=question_embeds.device
+            (question_embeds.size(0), 1), fill_value=self.decoder_start_token_id, device=question_embeds.device
         )
 
         outputs = self.text_decoder.generate(
@@ -1330,8 +1353,16 @@ class BlipForImageTextRetrieval(BlipPreTrainedModel):
         # image text matching head
         self.itm_head = nn.Linear(config.text_config.hidden_size, 2)
 
-        self.decoder_pad_token_id = config.text_config.pad_token_id
-        self.decoder_bos_token_id = config.text_config.bos_token_id
+        self.decoder_pad_token_id = (
+            config.text_config.pad_token_id
+            if not hasattr(config, "decoder_pad_token_id")
+            else config.decoder_pad_token_id
+        )
+        self.decoder_start_token_id = (
+            config.text_config.bos_token_id
+            if not hasattr(config, "decoder_start_token_id")
+            else config.decoder_start_token_id
+        )
 
         # Initialize weights and apply final processing
         self.post_init()

src/transformers/models/blip/modeling_blip_text.py

@@ -731,7 +731,7 @@ class BlipTextModel(BlipTextPreTrainedModel):
         past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
 
         if attention_mask is None:
-            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)))
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length))).to(device)
 
         # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
         # ourselves in which case we just need to make it broadcastable to all heads.

tests/models/blip/test_modeling_blip.py

@@ -521,6 +521,54 @@ class BlipModelTest(ModelTesterMixin, unittest.TestCase):
             self.assertIsNotNone(model)
 
 
+class BlipTextRetrievalModelTester:
+    def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
+
+        if text_kwargs is None:
+            text_kwargs = {}
+        if vision_kwargs is None:
+            vision_kwargs = {}
+
+        self.parent = parent
+        self.text_model_tester = BlipTextModelTester(parent, **text_kwargs)
+        self.vision_model_tester = BlipVisionModelTester(parent, **vision_kwargs)
+        self.is_training = is_training
+
+    def prepare_config_and_inputs(self):
+        text_config, input_ids, attention_mask = self.text_model_tester.prepare_config_and_inputs()
+        vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
+
+        config = self.get_config()
+
+        return config, input_ids, attention_mask, pixel_values
+
+    def get_config(self):
+        return BlipConfig.from_text_vision_configs(
+            self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
+        )
+
+    def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
+        model = BlipModel(config).to(torch_device).eval()
+        with torch.no_grad():
+            result = model(input_ids, pixel_values, attention_mask)
+        self.parent.assertEqual(
+            result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
+        )
+        self.parent.assertEqual(
+            result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_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 = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "pixel_values": pixel_values,
+        }
+        return config, inputs_dict
+
+
 class BlipTextImageModelsModelTester:
     def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
 
@@ -563,19 +611,277 @@ class BlipTextImageModelsModelTester:
         config, input_ids, attention_mask, pixel_values = config_and_inputs
         inputs_dict = {
             "input_ids": input_ids,
+            "labels": input_ids,
             "attention_mask": attention_mask,
             "pixel_values": pixel_values,
         }
         return config, inputs_dict
 
 
+@require_torch
+@require_vision
+class BlipVQAModelTest(unittest.TestCase):
+    all_model_classes = (BlipForQuestionAnswering,) if is_torch_available() else ()
+
+    def setUp(self):
+        self.model_tester = BlipModelTester(self)
+
+    def _prepare_inputs_for_vqa(self):
+        _, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        inputs_dict["labels"] = inputs_dict["input_ids"]
+        inputs_dict.pop("return_loss")
+        return inputs_dict
+
+    def test_class_name_consistency(self):
+        """
+        Tests that all VQA models have a class name that ends with "ForQuestionAnswering"
+        """
+        for model_class in self.all_model_classes:
+            model = model_class(self.model_tester.get_config())
+            self.assertTrue(
+                model.__class__.__name__.endswith("ForQuestionAnswering"),
+                f"Class name should end with 'ForVisualQuestionAnswering' got {model.__class__.__name__}",
+            )
+
+    def test_training(self):
+        """
+        Tests that all VQA models can be trained on a single batch
+        """
+        for model_class in self.all_model_classes:
+            model = model_class(self.model_tester.get_config()).to(torch_device)
+            model.train()
+            loss = model(**self._prepare_inputs_for_vqa()).loss
+            loss.backward()
+
+            # verify the gradients are not None
+            for name, param in model.named_parameters():
+                self.assertIsNotNone(param.grad, f"Gradients should not be None - got {param.grad} for {name}")
+
+    def test_forward_signature(self):
+        """
+        Test if the forward function has the expected arguments.
+        """
+        for model_class in self.all_model_classes:
+            model = model_class(self.model_tester.get_config())
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so args are the first n entries
+            args = list(signature.parameters.keys())
+            expected_args = [
+                "input_ids",
+                "attention_mask",
+                "labels",
+                "decoder_input_ids",
+                "decoder_attention_mask",
+            ]
+            for arg in expected_args:
+                self.assertTrue(
+                    arg in args,
+                    f"Argument {arg} of forward function signature should include {arg}. Found {args}.",
+                )
+
+
+@require_torch
+class BlipTextRetrievalModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (BlipForImageTextRetrieval,) 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 = BlipTextRetrievalModelTester(self)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*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="BlipModel does not have input/output embeddings")
+    def test_model_common_attributes(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()]
+
+            if model.config.is_encoder_decoder:
+                expected_arg_names = [
+                    "input_ids",
+                    "attention_mask",
+                    "decoder_input_ids",
+                    "decoder_attention_mask",
+                ]
+                expected_arg_names.extend(
+                    ["head_mask", "decoder_head_mask", "cross_attn_head_mask", "encoder_outputs"]
+                    if "head_mask" and "decoder_head_mask" and "cross_attn_head_mask" in arg_names
+                    else ["encoder_outputs"]
+                )
+                self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
+            else:
+                expected_arg_names = ["input_ids"] if model_class != BlipForConditionalGeneration else ["pixel_values"]
+                self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    def test_training(self):
+        if not self.model_tester.is_training:
+            return
+
+        for model_class in self.all_model_classes[:-1]:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            config.return_dict = True
+
+            model = model_class(config)
+            model.to(torch_device)
+            model.train()
+            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+
+            # hardcode labels to be the same as input_ids
+            inputs["labels"] = inputs["input_ids"]
+
+            loss = model(**inputs).loss
+            loss.backward()
+
+    def test_training_gradient_checkpointing(self):
+        if not self.model_tester.is_training:
+            return
+
+        for model_class in self.all_model_classes[:-1]:
+            config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+            config.use_cache = False
+            config.return_dict = True
+
+            model = model_class(config)
+            model.to(torch_device)
+            model.gradient_checkpointing_enable()
+            model.train()
+            inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+
+            # hardcode labels to be the same as input_ids
+            inputs["labels"] = inputs["input_ids"]
+
+            loss = model(**inputs).loss
+            loss.backward()
+
+    # override as the `logit_scale` parameter initilization is different for Blip
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        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:
+                    # check if `logit_scale` is initilized as per the original implementation
+                    if name == "logit_scale":
+                        self.assertAlmostEqual(
+                            param.data.item(),
+                            np.log(1 / 0.07),
+                            delta=1e-3,
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+                    else:
+                        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",
+                        )
+
+    def _create_and_check_torchscript(self, config, inputs_dict):
+        if not self.test_torchscript:
+            return
+
+        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
+        configs_no_init.torchscript = True
+        configs_no_init.return_dict = False
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            model.to(torch_device)
+            model.eval()
+
+            try:
+                input_ids = inputs_dict["input_ids"]
+                pixel_values = inputs_dict["pixel_values"]  # Blip needs pixel_values
+                traced_model = torch.jit.trace(model, (input_ids, pixel_values))
+            except RuntimeError:
+                self.fail("Couldn't trace module.")
+
+            with tempfile.TemporaryDirectory() as tmp_dir_name:
+                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
+
+                try:
+                    torch.jit.save(traced_model, pt_file_name)
+                except Exception:
+                    self.fail("Couldn't save module.")
+
+                try:
+                    loaded_model = torch.jit.load(pt_file_name)
+                except Exception:
+                    self.fail("Couldn't load module.")
+
+            model.to(torch_device)
+            model.eval()
+
+            loaded_model.to(torch_device)
+            loaded_model.eval()
+
+            model_state_dict = model.state_dict()
+            loaded_model_state_dict = loaded_model.state_dict()
+
+            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
+
+            models_equal = True
+            for layer_name, p1 in model_state_dict.items():
+                p2 = loaded_model_state_dict[layer_name]
+                if p1.data.ne(p2.data).sum() > 0:
+                    models_equal = False
+
+            self.assertTrue(models_equal)
+
+    def test_load_vision_text_config(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # Save BlipConfig and check if we can load BlipVisionConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            vision_config = BlipVisionConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
+
+        # Save BlipConfig and check if we can load BlipTextConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            text_config = BlipTextConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in BLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = BlipModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
 @require_torch
 class BlipTextImageModelTest(ModelTesterMixin, unittest.TestCase):
     all_model_classes = (
         (
             BlipForConditionalGeneration,
             BlipForQuestionAnswering,
-            BlipForImageTextRetrieval,
         )
         if is_torch_available()
         else ()
@@ -648,6 +954,10 @@ class BlipTextImageModelTest(ModelTesterMixin, unittest.TestCase):
             model.to(torch_device)
             model.train()
             inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+
+            # hardcode labels to be the same as input_ids
+            inputs["labels"] = inputs["input_ids"]
+
             loss = model(**inputs).loss
             loss.backward()
 
@@ -665,6 +975,10 @@ class BlipTextImageModelTest(ModelTesterMixin, unittest.TestCase):
             model.gradient_checkpointing_enable()
             model.train()
             inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
+
+            # hardcode labels to be the same as input_ids
+            inputs["labels"] = inputs["input_ids"]
+
             loss = model(**inputs).loss
             loss.backward()