# 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 numpy as np import requests import transformers from transformers import CLIPConfig, CLIPTextConfig, CLIPVisionConfig from transformers.testing_utils import ( is_flax_available, is_pt_flax_cross_test, 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, ) from ...test_pipeline_mixin import PipelineTesterMixin if is_torch_available(): import torch from torch import nn from transformers import ( CLIPModel, CLIPTextModel, CLIPTextModelWithProjection, CLIPVisionModel, CLIPVisionModelWithProjection, ) from transformers.models.clip.modeling_clip import CLIP_PRETRAINED_MODEL_ARCHIVE_LIST if is_vision_available(): from PIL import Image from transformers import CLIPProcessor if is_flax_available(): import jax.numpy as jnp from transformers.modeling_flax_pytorch_utils import ( convert_pytorch_state_dict_to_flax, load_flax_weights_in_pytorch_model, ) class CLIPVisionModelTester: 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=2, 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.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 CLIPVisionConfig( 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 = CLIPVisionModel(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 create_and_check_model_with_projection(self, config, pixel_values): model = CLIPVisionModelWithProjection(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.image_embeds.shape, (self.batch_size, self.projection_dim)) 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, CLIPVisionModelWithProjection) if is_torch_available() else () fx_compatible = True test_pruning = 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() @unittest.skip(reason="CLIP 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_model_with_projection(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_model_with_projection(*config_and_inputs) def test_training(self): pass def test_training_gradient_checkpointing(self): pass @unittest.skip(reason="CLIPVisionModel has no base class and is not available in MODEL_MAPPING") def test_save_load_fast_init_from_base(self): pass @unittest.skip(reason="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) @slow def test_model_with_projection_from_pretrained(self): for model_name in CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]: model = CLIPVisionModelWithProjection.from_pretrained(model_name) self.assertIsNotNone(model) self.assertTrue(hasattr(model, "visual_projection")) 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, projection_dim=32, num_hidden_layers=2, 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.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 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 CLIPTextConfig( 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, ) def create_and_check_model(self, config, input_ids, input_mask): model = CLIPTextModel(config=config) model.to(torch_device) model.eval() with torch.no_grad(): 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 create_and_check_model_with_projection(self, config, input_ids, input_mask): model = CLIPTextModelWithProjection(config=config) model.to(torch_device) model.eval() with torch.no_grad(): 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.text_embeds.shape, (self.batch_size, self.projection_dim)) 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, CLIPTextModelWithProjection) if is_torch_available() else () fx_compatible = True test_pruning = False test_head_masking = False model_split_percents = [0.5, 0.8, 0.9] 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_model_with_projection(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_model_with_projection(*config_and_inputs) def test_training(self): pass def test_training_gradient_checkpointing(self): pass @unittest.skip(reason="CLIP does not use inputs_embeds") def test_inputs_embeds(self): pass @unittest.skip(reason="CLIPTextModel has no base class and is not available in MODEL_MAPPING") def test_save_load_fast_init_from_base(self): pass @unittest.skip(reason="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) @slow def test_model_with_projection_from_pretrained(self): for model_name in CLIP_PRETRAINED_MODEL_ARCHIVE_LIST[:1]: model = CLIPTextModelWithProjection.from_pretrained(model_name) self.assertIsNotNone(model) self.assertTrue(hasattr(model, "text_projection")) class CLIPModelTester: 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 = CLIPTextModelTester(parent, **text_kwargs) self.vision_model_tester = CLIPVisionModelTester(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 CLIPConfig.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 = CLIPModel(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_loss": True, } return config, inputs_dict @require_torch class CLIPModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase): all_model_classes = (CLIPModel,) if is_torch_available() else () pipeline_model_mapping = {"feature-extraction": CLIPModel} if is_torch_available() else {} fx_compatible = True 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) @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="CLIPModel does not have input/output embeddings") def test_model_common_attributes(self): pass # override as the `logit_scale` parameter initilization is different for CLIP 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"] # 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() non_persistent_buffers = {} for key in loaded_model_state_dict.keys(): if key not in model_state_dict.keys(): non_persistent_buffers[key] = loaded_model_state_dict[key] loaded_model_state_dict = { key: value for key, value in loaded_model_state_dict.items() if key not in non_persistent_buffers } self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys())) model_buffers = list(model.buffers()) for non_persistent_buffer in non_persistent_buffers.values(): found_buffer = False for i, model_buffer in enumerate(model_buffers): if torch.equal(non_persistent_buffer, model_buffer): found_buffer = True break self.assertTrue(found_buffer) model_buffers.pop(i) 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 CLIPConfig and check if we can load CLIPVisionConfig from it with tempfile.TemporaryDirectory() as tmp_dir_name: config.save_pretrained(tmp_dir_name) vision_config = CLIPVisionConfig.from_pretrained(tmp_dir_name) self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict()) # Save CLIPConfig and check if we can load CLIPTextConfig from it with tempfile.TemporaryDirectory() as tmp_dir_name: config.save_pretrained(tmp_dir_name) text_config = CLIPTextConfig.from_pretrained(tmp_dir_name) self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict()) # overwrite from common since FlaxCLIPModel returns nested output # which is not supported in the common test @is_pt_flax_cross_test def test_equivalence_pt_to_flax(self): config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() for model_class in self.all_model_classes: with self.subTest(model_class.__name__): # load PyTorch class pt_model = model_class(config).eval() # Flax models don't use the `use_cache` option and cache is not returned as a default. # So we disable `use_cache` here for PyTorch model. pt_model.config.use_cache = False fx_model_class_name = "Flax" + model_class.__name__ if not hasattr(transformers, fx_model_class_name): return fx_model_class = getattr(transformers, fx_model_class_name) # load Flax class fx_model = fx_model_class(config, dtype=jnp.float32) # make sure only flax inputs are forward that actually exist in function args fx_input_keys = inspect.signature(fx_model.__call__).parameters.keys() # prepare inputs pt_inputs = self._prepare_for_class(inputs_dict, model_class) # remove function args that don't exist in Flax pt_inputs = {k: v for k, v in pt_inputs.items() if k in fx_input_keys} fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model) fx_model.params = fx_state with torch.no_grad(): pt_outputs = pt_model(**pt_inputs).to_tuple() # convert inputs to Flax fx_inputs = {k: np.array(v.to("cpu")) for k, v in pt_inputs.items() if torch.is_tensor(v)} fx_outputs = fx_model(**fx_inputs).to_tuple() self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch") for fx_output, pt_output in zip(fx_outputs[:4], pt_outputs[:4]): self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2) with tempfile.TemporaryDirectory() as tmpdirname: pt_model.save_pretrained(tmpdirname) fx_model_loaded = fx_model_class.from_pretrained(tmpdirname, from_pt=True) fx_outputs_loaded = fx_model_loaded(**fx_inputs).to_tuple() self.assertEqual( len(fx_outputs_loaded), len(pt_outputs), "Output lengths differ between Flax and PyTorch" ) for fx_output_loaded, pt_output in zip(fx_outputs_loaded[:4], pt_outputs[:4]): self.assert_almost_equals(fx_output_loaded, pt_output.numpy(), 4e-2) # overwrite from common since FlaxCLIPModel returns nested output # which is not supported in the common test @is_pt_flax_cross_test def test_equivalence_flax_to_pt(self): config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() for model_class in self.all_model_classes: with self.subTest(model_class.__name__): # load corresponding PyTorch class pt_model = model_class(config).eval() # So we disable `use_cache` here for PyTorch model. pt_model.config.use_cache = False fx_model_class_name = "Flax" + model_class.__name__ if not hasattr(transformers, fx_model_class_name): # no flax model exists for this class return fx_model_class = getattr(transformers, fx_model_class_name) # load Flax class fx_model = fx_model_class(config, dtype=jnp.float32) # make sure only flax inputs are forward that actually exist in function args fx_input_keys = inspect.signature(fx_model.__call__).parameters.keys() pt_model = load_flax_weights_in_pytorch_model(pt_model, fx_model.params) # make sure weights are tied in PyTorch pt_model.tie_weights() # prepare inputs pt_inputs = self._prepare_for_class(inputs_dict, model_class) # remove function args that don't exist in Flax pt_inputs = {k: v for k, v in pt_inputs.items() if k in fx_input_keys} with torch.no_grad(): pt_outputs = pt_model(**pt_inputs).to_tuple() fx_inputs = {k: np.array(v.to("cpu")) for k, v in pt_inputs.items() if torch.is_tensor(v)} fx_outputs = fx_model(**fx_inputs).to_tuple() self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch") for fx_output, pt_output in zip(fx_outputs[:4], pt_outputs[:4]): self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2) with tempfile.TemporaryDirectory() as tmpdirname: fx_model.save_pretrained(tmpdirname) pt_model_loaded = model_class.from_pretrained(tmpdirname, from_flax=True) with torch.no_grad(): pt_outputs_loaded = pt_model_loaded(**pt_inputs).to_tuple() self.assertEqual( len(fx_outputs), len(pt_outputs_loaded), "Output lengths differ between Flax and PyTorch" ) for fx_output, pt_output in zip(fx_outputs[:4], pt_outputs_loaded[:4]): self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2) @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 @require_torch 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))