test_modeling_clip.py

# coding=utf-8
# Copyright 2021 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 CLIP model. """


import inspect
import os
import tempfile
import unittest

import requests
from transformers.file_utils import is_torch_available, is_vision_available
from transformers.testing_utils import require_torch, require_vision, slow, torch_device

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 transformers import CLIPConfig, CLIPModel, CLIPTextConfig, CLIPTextModel, CLIPVisionConfig, CLIPVisionModel
    from transformers.models.clip.modeling_clip import CLIP_PRETRAINED_MODEL_ARCHIVE_LIST


if is_vision_available():
    from PIL import Image

    from transformers import CLIPProcessor


class CLIPVisionModelTester:
    def __init__(
        self,
        parent,
        batch_size=12,
        image_size=30,
        patch_size=2,
        num_channels=3,
        is_training=True,
        hidden_size=32,
        num_hidden_layers=5,
        num_attention_heads=4,
        intermediate_size=37,
        dropout=0.1,
        attention_dropout=0.1,
        initializer_range=0.02,
        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.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

    def prepare_config_and_inputs(self):
        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
        config = CLIPVisionConfig(
            image_size=self.image_size,
            patch_size=self.patch_size,
            num_channels=self.num_channels,
            hidden_size=self.hidden_size,
            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,
        )

        return config, pixel_values

    def create_and_check_model(self, config, pixel_values):
        model = CLIPVisionModel(config=config)
        model.to(torch_device)
        model.eval()
        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 CLIPVisionModelTest(ModelTesterMixin, unittest.TestCase):
    """
    Here we also overwrite some of the tests of test_modeling_common.py, as CLIP does not use input_ids, inputs_embeds,
    attention_mask and seq_length.
    """

    all_model_classes = (CLIPVisionModel,) if is_torch_available() else ()

    test_pruning = False
    test_torchscript = False
    test_resize_embeddings = False
    test_head_masking = False

    def setUp(self):
        self.model_tester = CLIPVisionModelTester(self)
        self.config_tester = ConfigTester(self, config_class=CLIPVisionConfig, has_text_modality=False, hidden_size=37)

    def test_config(self):
        self.config_tester.run_common_tests()

    def test_inputs_embeds(self):
        # CLIP does not use inputs_embeds
        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(), (torch.nn.Module))
            x = model.get_output_embeddings()
            self.assertTrue(x is None or isinstance(x, torch.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_attention_outputs(self):
        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
        config.return_dict = True

        # in CLIP, the seq_len equals the number of patches + 1 (we add 1 for the [CLS] token)
        image_size = (self.model_tester.image_size, self.model_tester.image_size)
        patch_size = (self.model_tester.patch_size, self.model_tester.patch_size)
        num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
        seq_len = num_patches + 1

        for model_class in self.all_model_classes:
            inputs_dict["output_attentions"] = True
            inputs_dict["output_hidden_states"] = False
            config.return_dict = True
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs.attentions
            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)

            # check that output_attentions also work using config
            del inputs_dict["output_attentions"]
            config.output_attentions = True
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
            attentions = outputs.attentions
            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)

            out_len = len(outputs)

            # Check attention is always last and order is fine
            inputs_dict["output_attentions"] = True
            inputs_dict["output_hidden_states"] = True
            model = model_class(config)
            model.to(torch_device)
            model.eval()
            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))

            added_hidden_states = 1
            self.assertEqual(out_len + added_hidden_states, len(outputs))

            self_attentions = outputs.attentions

            self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)

            self.assertListEqual(
                list(self_attentions[0].shape[-3:]),
                [self.model_tester.num_attention_heads, seq_len, seq_len],
            )

    def test_hidden_states_output(self):
        def check_hidden_states_output(inputs_dict, config, model_class):
            model = model_class(config)
            model.to(torch_device)
            model.eval()

            with torch.no_grad():
                outputs = model(**self._prepare_for_class(inputs_dict, model_class))

            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states

            expected_num_layers = getattr(
                self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
            )
            self.assertEqual(len(hidden_states), expected_num_layers)

            # CLIP has a different seq_length
            image_size = (self.model_tester.image_size, self.model_tester.image_size)
            patch_size = (self.model_tester.patch_size, self.model_tester.patch_size)
            num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0])
            seq_length = num_patches + 1

            self.assertListEqual(
                list(hidden_states[0].shape[-2:]),
                [seq_length, self.model_tester.hidden_size],
            )

        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()

        for model_class in self.all_model_classes:
            inputs_dict["output_hidden_states"] = True
            check_hidden_states_output(inputs_dict, config, model_class)

            # check that output_hidden_states also work using config
            del inputs_dict["output_hidden_states"]
            config.output_hidden_states = True

            check_hidden_states_output(inputs_dict, config, model_class)

    def test_training(self):
        pass

    def test_training_gradient_checkpointing(self):
        pass

    # skip this test as CLIPVisionModel has no base class and is
    # not available in MODEL_MAPPING
    def test_save_load_fast_init_from_base(self):
        pass

    # skip this test as CLIPVisionModel 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 CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
            model = CLIPVisionModel.from_pretrained(model_name)
            self.assertIsNotNone(model)


class CLIPTextModelTester:
    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,
        num_hidden_layers=5,
        num_attention_heads=4,
        intermediate_size=37,
        dropout=0.1,
        attention_dropout=0.1,
        max_position_embeddings=512,
        initializer_range=0.02,
        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.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

    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])

        config = CLIPTextConfig(
            vocab_size=self.vocab_size,
            hidden_size=self.hidden_size,
            num_hidden_layers=self.num_hidden_layers,
            num_attention_heads=self.num_attention_heads,
            intermediate_size=self.intermediate_size,
            dropout=self.dropout,
            attention_dropout=self.attention_dropout,
            max_position_embeddings=self.max_position_embeddings,
            initializer_range=self.initializer_range,
        )

        return config, input_ids, input_mask

    def create_and_check_model(self, config, input_ids, input_mask):
        model = CLIPTextModel(config=config)
        model.to(torch_device)
        model.eval()
        result = model(input_ids, attention_mask=input_mask)
        result = model(input_ids)
        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, 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, input_ids, input_mask = config_and_inputs
        inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
        return config, inputs_dict


@require_torch
class CLIPTextModelTest(ModelTesterMixin, unittest.TestCase):

    all_model_classes = (CLIPTextModel,) if is_torch_available() else ()
    test_pruning = False
    test_head_masking = False

    def setUp(self):
        self.model_tester = CLIPTextModelTester(self)
        self.config_tester = ConfigTester(self, config_class=CLIPTextConfig, hidden_size=37)

    def test_config(self):
        self.config_tester.run_common_tests()

    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

    def test_inputs_embeds(self):
        # CLIP does not use inputs_embeds
        pass

    # skip this test as CLIPTextModel has no base class and is
    # not available in MODEL_MAPPING
    def test_save_load_fast_init_from_base(self):
        pass

    # skip this test as CLIPTextModel 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 CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
            model = CLIPTextModel.from_pretrained(model_name)
            self.assertIsNotNone(model)


class CLIPModelTester:
    def __init__(self, parent, is_training=True):
        self.parent = parent
        self.text_model_tester = CLIPTextModelTester(parent)
        self.vision_model_tester = CLIPVisionModelTester(parent)
        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 = CLIPConfig.from_text_vision_configs(text_config, vision_config, projection_dim=64)

        return config, input_ids, attention_mask, pixel_values

    def create_and_check_model(self, config, input_ids, attention_mask, pixel_values):
        model = CLIPModel(config).to(torch_device).eval()
        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_loss": True,
        }
        return config, inputs_dict


@require_torch
class CLIPModelTest(ModelTesterMixin, unittest.TestCase):
    all_model_classes = (CLIPModel,) if is_torch_available() else ()
    test_head_masking = False
    test_pruning = False
    test_resize_embeddings = False
    test_attention_outputs = False

    def setUp(self):
        self.model_tester = CLIPModelTester(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)

    # hidden_states are tested in individual model tests
    def test_hidden_states_output(self):
        pass

    # input_embeds are tested in individual model tests
    def test_inputs_embeds(self):
        pass

    # tested in individual model tests
    def test_retain_grad_hidden_states_attentions(self):
        pass

    # CLIPModel does not have input/output embeddings
    def test_model_common_attributes(self):
        pass

    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"]  # CLIP 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)

    @slow
    def test_model_from_pretrained(self):
        for model_name in CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
            model = CLIPModel.from_pretrained(model_name)
            self.assertIsNotNone(model)


# We will verify our results on an image of cute cats
def prepare_img():
    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
    im = Image.open(requests.get(url, stream=True).raw)
    return im


@require_vision
class CLIPModelIntegrationTest(unittest.TestCase):
    @slow
    def test_inference(self):
        model_name = "openai/clip-vit-base-patch32"
        model = CLIPModel.from_pretrained(model_name).to(torch_device)
        processor = CLIPProcessor.from_pretrained(model_name)

        image = prepare_img()
        inputs = processor(
            text=["a photo of a cat", "a photo of a dog"], images=image, padding=True, return_tensors="pt"
        ).to(torch_device)

        # forward pass
        with torch.no_grad():
            outputs = model(**inputs)

        # verify the logits
        self.assertEqual(
            outputs.logits_per_image.shape,
            torch.Size((inputs.pixel_values.shape[0], inputs.input_ids.shape[0])),
        )
        self.assertEqual(
            outputs.logits_per_text.shape,
            torch.Size((inputs.input_ids.shape[0], inputs.pixel_values.shape[0])),
        )

        expected_logits = torch.tensor([[24.5701, 19.3049]], device=torch_device)

        self.assertTrue(torch.allclose(outputs.logits_per_image, expected_logits, atol=1e-3))