first commit

tests/models/colpali/test_modeling_colpali.py

@@ -0,0 +1,286 @@
+# Copyright 2024 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 ColPali model."""
+
+import gc
+import unittest
+from typing import ClassVar
+
+import pytest
+import torch
+from datasets import load_dataset
+
+from tests.test_configuration_common import ConfigTester
+from tests.test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor
+from transformers import (
+    is_torch_available,
+)
+from transformers.models.colpali.configuration_colpali import ColPaliConfig
+from transformers.models.colpali.modeling_colpali import ColPaliForRetrieval, ColPaliForRetrievalOutput
+from transformers.models.colpali.processing_colpali import ColPaliProcessor
+from transformers.testing_utils import (
+    backend_empty_cache,
+    require_torch,
+    require_vision,
+    slow,
+    torch_device,
+)
+
+
+if is_torch_available():
+    import torch
+
+
+class ColPaliForRetrievalModelTester:
+    def __init__(
+        self,
+        parent,
+        ignore_index=-100,
+        image_token_index=0,
+        projector_hidden_act="gelu",
+        seq_length=25,
+        vision_feature_select_strategy="default",
+        vision_feature_layer=-1,
+        projection_dim=32,
+        text_config={
+            "model_type": "gemma",
+            "seq_length": 128,
+            "is_training": True,
+            "use_token_type_ids": False,
+            "use_labels": True,
+            "vocab_size": 99,
+            "hidden_size": 32,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 4,
+            "num_key_value_heads": 1,
+            "head_dim": 8,
+            "intermediate_size": 37,
+            "hidden_activation": "gelu_pytorch_tanh",
+            "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,
+            "pad_token_id": 1,
+        },
+        is_training=False,
+        vision_config={
+            "use_labels": True,
+            "image_size": 20,
+            "patch_size": 5,
+            "num_image_tokens": 4,
+            "num_channels": 3,
+            "is_training": True,
+            "hidden_size": 32,
+            "projection_dim": 32,
+            "num_key_value_heads": 1,
+            "num_hidden_layers": 2,
+            "num_attention_heads": 4,
+            "intermediate_size": 37,
+            "dropout": 0.1,
+            "attention_dropout": 0.1,
+            "initializer_range": 0.02,
+        },
+        use_cache=False,
+        embedding_dim=128,
+    ):
+        self.parent = parent
+        self.ignore_index = ignore_index
+        # `image_token_index` is set to 0 to pass "resize_embeddings" test, do not modify
+        self.image_token_index = image_token_index
+        self.projector_hidden_act = projector_hidden_act
+        self.vision_feature_select_strategy = vision_feature_select_strategy
+        self.vision_feature_layer = vision_feature_layer
+        self.text_config = text_config
+        self.vision_config = vision_config
+        self.seq_length = seq_length
+        self.projection_dim = projection_dim
+        self.pad_token_id = text_config["pad_token_id"]
+
+        self.num_hidden_layers = text_config["num_hidden_layers"]
+        self.vocab_size = text_config["vocab_size"]
+        self.hidden_size = text_config["hidden_size"]
+        self.num_attention_heads = text_config["num_attention_heads"]
+        self.is_training = is_training
+
+        self.batch_size = 3
+        self.num_channels = vision_config["num_channels"]
+        self.image_size = vision_config["image_size"]
+        self.encoder_seq_length = seq_length
+        self.use_cache = use_cache
+
+        self.embedding_dim = embedding_dim
+        self.vlm_config = {
+            "model_type": "paligemma",
+            "text_config": self.text_config,
+            "vision_config": self.vision_config,
+            "ignore_index": self.ignore_index,
+            "image_token_index": self.image_token_index,
+            "projector_hidden_act": self.projector_hidden_act,
+            "projection_dim": self.projection_dim,
+            "vision_feature_select_strategy": self.vision_feature_select_strategy,
+            "vision_feature_layer": self.vision_feature_layer,
+        }
+
+    def get_config(self):
+        return ColPaliConfig(
+            vlm_config=self.vlm_config,
+            embedding_dim=self.embedding_dim,
+        )
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor(
+            [
+                self.batch_size,
+                self.vision_config["num_channels"],
+                self.vision_config["image_size"],
+                self.vision_config["image_size"],
+            ]
+        )
+        config = self.get_config()
+
+        return config, pixel_values
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values = config_and_inputs
+        input_ids = ids_tensor([self.batch_size, self.seq_length], config.vlm_config.text_config.vocab_size - 1) + 1
+        attention_mask = input_ids.ne(1).to(torch_device)
+        # set the 16 first tokens to be image, and ensure that no other tokens are image tokens
+        # do not change this unless you modified image size or patch size
+        input_ids[input_ids == config.vlm_config.image_token_index] = self.pad_token_id
+        input_ids[:, :16] = config.vlm_config.image_token_index
+        inputs_dict = {
+            "pixel_values": pixel_values,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "labels": input_ids,
+            "token_type_ids": torch.zeros_like(input_ids),
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class ColPaliForRetrievalModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Model tester for `ColPaliForRetrieval`.
+    """
+
+    all_model_classes = (ColPaliForRetrieval,) if is_torch_available() else ()
+    test_resize_embeddings = True
+    test_missing_keys = False
+    additional_model_inputs = ["token_type_ids"]
+
+    def setUp(self):
+        self.model_tester = ColPaliForRetrievalModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=ColPaliConfig, has_text_modality=False)
+
+    @slow
+    @require_vision
+    def test_colpali_forward_inputs(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            inputs = self._prepare_for_class(inputs_dict, model_class)
+
+            with torch.no_grad():
+                outputs = model(**inputs, return_dict=True)
+
+            self.assertIsInstance(outputs, ColPaliForRetrievalOutput)
+
+    @unittest.skip(
+        reason="From PaliGemma: Some undefined behavior encountered with test versions of this model. Skip for now."
+    )
+    def test_model_parallelism(self):
+        pass
+
+    # TODO extend valid outputs to include this test @Molbap
+    @unittest.skip(reason="PaliGemma has currently one output format.")
+    def test_model_outputs_equivalence(self):
+        pass
+
+    @unittest.skip(reason="Pass because ColPali requires `attention_mask is not None`")
+    def test_sdpa_can_dispatch_on_flash(self):
+        pass
+
+    @unittest.skip(reason="Pass because ColPali requires `attention_mask is not None`")
+    @pytest.mark.torch_compile_test
+    def test_sdpa_can_compile_dynamic(self):
+        pass
+
+
+@require_torch
+class ColPaliModelIntegrationTest(unittest.TestCase):
+    model_name: ClassVar[str] = "vidore/colpali-v1.2-hf"
+
+    def setUp(self):
+        self.processor = ColPaliProcessor.from_pretrained(self.model_name)
+
+    def tearDown(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+
+    @slow
+    def test_model_integration_test(self):
+        """
+        Test if the model is able to retrieve the correct pages for a small and easy dataset.
+        """
+        model = ColPaliForRetrieval.from_pretrained(
+            self.model_name,
+            dtype=torch.bfloat16,
+            device_map=torch_device,
+        ).eval()
+
+        # Load the test dataset
+        ds = load_dataset("hf-internal-testing/document-visual-retrieval-test", split="test")
+
+        # Preprocess the examples
+        batch_images = self.processor(images=ds["image"][:]).to(torch_device)
+        batch_queries = self.processor(text=ds["query"][:]).to(torch_device)
+
+        # Run inference
+        with torch.inference_mode():
+            image_embeddings = model(**batch_images).embeddings
+            query_embeddings = model(**batch_queries).embeddings
+
+        # Compute retrieval scores
+        scores = self.processor.score_retrieval(
+            query_embeddings=query_embeddings,
+            passage_embeddings=image_embeddings,
+        )  # (num_queries, num_passages)
+
+        assert scores.ndim == 2, f"Expected 2D tensor, got {scores.ndim}"
+        assert scores.shape == (len(ds), len(ds)), f"Expected shape {(len(ds), len(ds))}, got {scores.shape}"
+
+        # Check if the maximum scores per row are in the diagonal of the matrix score
+        self.assertTrue((scores.argmax(axis=1) == torch.arange(len(ds), device=scores.device)).all())
+
+        # Further validation: fine-grained check, with a hardcoded score from the original implementation
+        expected_scores = torch.tensor(
+            [
+                [15.5625, 6.5938, 14.4375],
+                [12.2500, 16.2500, 11.0000],
+                [15.0625, 11.7500, 21.0000],
+            ],
+            dtype=scores.dtype,
+        )
+
+        assert torch.allclose(scores, expected_scores, atol=1), f"Expected scores {expected_scores}, got {scores}"

tests/models/colpali/test_processing_colpali.py

@@ -0,0 +1,294 @@
+# Copyright 2024 HuggingFace Inc.
+#
+# 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 ColPali processor."""
+
+import unittest
+
+import torch
+
+from transformers.models.colpali.processing_colpali import ColPaliProcessor
+from transformers.testing_utils import get_tests_dir, require_torch, require_vision
+from transformers.utils import is_vision_available
+
+from ...test_processing_common import ProcessorTesterMixin
+
+
+if is_vision_available():
+    from transformers import ColPaliProcessor, GemmaTokenizer
+
+SAMPLE_VOCAB = get_tests_dir("fixtures/test_sentencepiece.model")
+
+
+@require_vision
+class ColPaliProcessorTest(ProcessorTesterMixin, unittest.TestCase):
+    processor_class = ColPaliProcessor
+
+    @classmethod
+    def _setup_tokenizer(cls):
+        return GemmaTokenizer.from_pretrained(SAMPLE_VOCAB, keep_accents=True)
+
+    @classmethod
+    def _setup_image_processor(cls):
+        image_processor_class = cls._get_component_class_from_processor("image_processor")
+        image_processor = image_processor_class.from_pretrained("google/siglip-so400m-patch14-384")
+        image_processor.image_seq_length = 0
+        return image_processor
+
+    @unittest.skip("ColpaliProcessor can only process one of text or images at a time")
+    def test_processor_with_multiple_inputs(self):
+        pass
+
+    @unittest.skip("ColpaliProcessor adds a prefix and suffix to the text")
+    def test_tokenizer_defaults(self):
+        pass
+
+    def test_get_num_vision_tokens(self):
+        "Tests general functionality of the helper used internally in vLLM"
+
+        processor = self.get_processor()
+
+        output = processor._get_num_multimodal_tokens(image_sizes=[(100, 100), (300, 100), (500, 30)])
+        self.assertTrue("num_image_tokens" in output)
+        self.assertEqual(len(output["num_image_tokens"]), 3)
+
+        self.assertTrue("num_image_patches" in output)
+        self.assertEqual(len(output["num_image_patches"]), 3)
+
+    @require_torch
+    @require_vision
+    def test_process_images(self):
+        # Processor configuration
+        image_input = self.prepare_image_inputs()
+        image_processor = self.get_component("image_processor")
+        tokenizer = self.get_component("tokenizer", max_length=112, padding="max_length")
+        image_processor.image_seq_length = 14
+
+        # Get the processor
+        processor = self.processor_class(
+            tokenizer=tokenizer,
+            image_processor=image_processor,
+        )
+
+        # Process the image
+        batch_feature = processor.process_images(images=image_input, return_tensors="pt")
+
+        # Assertions
+        self.assertIn("pixel_values", batch_feature)
+        self.assertEqual(batch_feature["pixel_values"].shape, torch.Size([1, 3, 384, 384]))
+
+    @require_torch
+    @require_vision
+    def test_process_queries(self):
+        # Inputs
+        queries = [
+            "Is attention really all you need?",
+            "Are Benjamin, Antoine, Merve, and Jo best friends?",
+        ]
+
+        # Processor configuration
+        image_processor = self.get_component("image_processor")
+        tokenizer = self.get_component("tokenizer", max_length=112, padding="max_length")
+        image_processor.image_seq_length = 14
+
+        # Get the processor
+        processor = self.processor_class(
+            tokenizer=tokenizer,
+            image_processor=image_processor,
+        )
+
+        # Process the image
+        batch_feature = processor.process_queries(text=queries, return_tensors="pt")
+
+        # Assertions
+        self.assertIn("input_ids", batch_feature)
+        self.assertIsInstance(batch_feature["input_ids"], torch.Tensor)
+        self.assertEqual(batch_feature["input_ids"].shape[0], len(queries))
+
+    # The following tests override the parent tests because ColPaliProcessor can only take one of images or text as input at a time.
+
+    def test_tokenizer_defaults_preserved_by_kwargs(self):
+        if "image_processor" not in self.processor_class.get_attributes():
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        processor_components = self.prepare_components()
+        processor_components["tokenizer"] = self.get_component("tokenizer", max_length=117, padding="max_length")
+
+        processor = self.processor_class(**processor_components)
+        self.skip_processor_without_typed_kwargs(processor)
+        input_str = self.prepare_text_inputs()
+        inputs = processor(text=input_str, return_tensors="pt")
+        self.assertEqual(inputs[self.text_input_name].shape[-1], 117)
+
+    def test_image_processor_defaults_preserved_by_image_kwargs(self):
+        """
+        We use do_rescale=True, rescale_factor=-1.0 to ensure that image_processor kwargs are preserved in the processor.
+        We then check that the mean of the pixel_values is less than or equal to 0 after processing.
+        Since the original pixel_values are in [0, 255], this is a good indicator that the rescale_factor is indeed applied.
+        """
+        if "image_processor" not in self.processor_class.get_attributes():
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        processor_components = self.prepare_components()
+        processor_components["image_processor"] = self.get_component(
+            "image_processor", do_rescale=True, rescale_factor=-1.0
+        )
+        processor_components["tokenizer"] = self.get_component("tokenizer", max_length=117, padding="max_length")
+
+        processor = self.processor_class(**processor_components)
+        self.skip_processor_without_typed_kwargs(processor)
+
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(images=image_input, return_tensors="pt")
+        self.assertLessEqual(inputs[self.images_input_name][0][0].mean(), 0)
+
+    def test_kwargs_overrides_default_tokenizer_kwargs(self):
+        if "image_processor" not in self.processor_class.get_attributes():
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        processor_components = self.prepare_components()
+        processor_components["tokenizer"] = self.get_component("tokenizer", padding="longest")
+
+        processor = self.processor_class(**processor_components)
+        self.skip_processor_without_typed_kwargs(processor)
+        input_str = self.prepare_text_inputs()
+        inputs = processor(text=input_str, return_tensors="pt", max_length=112, padding="max_length")
+        self.assertEqual(inputs[self.text_input_name].shape[-1], 112)
+
+    def test_kwargs_overrides_default_image_processor_kwargs(self):
+        if "image_processor" not in self.processor_class.get_attributes():
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        processor_components = self.prepare_components()
+        processor_components["image_processor"] = self.get_component(
+            "image_processor", do_rescale=True, rescale_factor=1
+        )
+        processor_components["tokenizer"] = self.get_component("tokenizer", max_length=117, padding="max_length")
+
+        processor = self.processor_class(**processor_components)
+        self.skip_processor_without_typed_kwargs(processor)
+
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(images=image_input, do_rescale=True, rescale_factor=-1.0, return_tensors="pt")
+        self.assertLessEqual(inputs[self.images_input_name][0][0].mean(), 0)
+
+    def test_unstructured_kwargs(self):
+        if "image_processor" not in self.processor_class.get_attributes():
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        processor_components = self.prepare_components()
+        processor = self.processor_class(**processor_components)
+        self.skip_processor_without_typed_kwargs(processor)
+
+        input_str = self.prepare_text_inputs()
+        inputs = processor(
+            text=input_str,
+            return_tensors="pt",
+            do_rescale=True,
+            rescale_factor=-1.0,
+            padding="max_length",
+            max_length=76,
+        )
+
+        self.assertEqual(inputs[self.text_input_name].shape[-1], 76)
+
+    def test_unstructured_kwargs_batched(self):
+        if "image_processor" not in self.processor_class.get_attributes():
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        processor_components = self.prepare_components()
+        processor = self.processor_class(**processor_components)
+        self.skip_processor_without_typed_kwargs(processor)
+
+        image_input = self.prepare_image_inputs(batch_size=2)
+        inputs = processor(
+            images=image_input,
+            return_tensors="pt",
+            do_rescale=True,
+            rescale_factor=-1.0,
+            padding="longest",
+            max_length=76,
+        )
+
+        self.assertLessEqual(inputs[self.images_input_name][0][0].mean(), 0)
+
+    def test_doubly_passed_kwargs(self):
+        if "image_processor" not in self.processor_class.get_attributes():
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        processor_components = self.prepare_components()
+        processor = self.processor_class(**processor_components)
+        self.skip_processor_without_typed_kwargs(processor)
+
+        image_input = self.prepare_image_inputs()
+        with self.assertRaises(ValueError):
+            _ = processor(
+                images=image_input,
+                images_kwargs={"do_rescale": True, "rescale_factor": -1.0},
+                do_rescale=True,
+                return_tensors="pt",
+            )
+
+    def test_structured_kwargs_nested(self):
+        if "image_processor" not in self.processor_class.get_attributes():
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        processor_components = self.prepare_components()
+        processor = self.processor_class(**processor_components)
+        self.skip_processor_without_typed_kwargs(processor)
+
+        input_str = self.prepare_text_inputs()
+
+        # Define the kwargs for each modality
+        all_kwargs = {
+            "common_kwargs": {"return_tensors": "pt"},
+            "images_kwargs": {"do_rescale": True, "rescale_factor": -1.0},
+            "text_kwargs": {"padding": "max_length", "max_length": 76},
+        }
+
+        inputs = processor(text=input_str, **all_kwargs)
+        self.skip_processor_without_typed_kwargs(processor)
+
+        self.assertEqual(inputs[self.text_input_name].shape[-1], 76)
+
+    def test_structured_kwargs_nested_from_dict(self):
+        if "image_processor" not in self.processor_class.get_attributes():
+            self.skipTest(f"image_processor attribute not present in {self.processor_class}")
+        processor_components = self.prepare_components()
+        processor = self.processor_class(**processor_components)
+        self.skip_processor_without_typed_kwargs(processor)
+        image_input = self.prepare_image_inputs()
+
+        # Define the kwargs for each modality
+        all_kwargs = {
+            "common_kwargs": {"return_tensors": "pt"},
+            "images_kwargs": {"do_rescale": True, "rescale_factor": -1.0},
+            "text_kwargs": {"padding": "max_length", "max_length": 76},
+        }
+
+        inputs = processor(images=image_input, **all_kwargs)
+        self.assertEqual(inputs[self.text_input_name].shape[-1], 76)
+
+    # Can process only text or images at a time
+    def test_model_input_names(self):
+        processor = self.get_processor()
+        image_input = self.prepare_image_inputs()
+        inputs = processor(images=image_input)
+
+        self.assertSetEqual(set(inputs.keys()), set(processor.model_input_names))
+
+    @unittest.skip("ColPali can't process text+image inputs at the same time")
+    def test_processor_text_has_no_visual(self):
+        pass
+
+    @unittest.skip("ColPaliProcessor can't process text+image inputs at the same time")
+    def test_get_num_multimodal_tokens_matches_processor_call(self):
+        pass
+
+    @unittest.skip("ColPaliProcessor can't process text+image inputs at the same time")
+    def test_flat_kwarg_applied_when_modality_dict_lacks_it(self):
+        pass