# Copyright 2025 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. import unittest import numpy as np from transformers.image_utils import IMAGENET_STANDARD_MEAN, IMAGENET_STANDARD_STD from transformers.testing_utils import require_torch, require_vision from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs if is_torch_available(): import torch if is_vision_available(): from PIL import Image if is_torchvision_available(): from transformers import PerceptionLMImageProcessorFast class PerceptionLMImageProcessingTester: def __init__( self, parent, batch_size=7, num_channels=3, image_size=18, min_resolution=30, max_resolution=400, do_resize=True, tile_size=16, do_normalize=True, image_mean=IMAGENET_STANDARD_MEAN, image_std=IMAGENET_STANDARD_STD, do_convert_rgb=True, max_num_tiles=4, vision_input_type="thumb+tile", resample=Image.Resampling.BICUBIC, # dummy value size={"shortest_edge": 20}, # dummy value ): super().__init__() self.parent = parent self.batch_size = batch_size self.num_channels = num_channels self.image_size = image_size self.min_resolution = min_resolution self.max_resolution = max_resolution self.do_resize = do_resize self.tile_size = tile_size self.do_normalize = do_normalize self.image_mean = image_mean self.image_std = image_std self.do_convert_rgb = do_convert_rgb self.max_num_tiles = max_num_tiles self.vision_input_type = vision_input_type self.resample = resample self.size = size def prepare_image_processor_dict(self): return { "do_resize": self.do_resize, "tile_size": self.tile_size, "do_normalize": self.do_normalize, "image_mean": self.image_mean, "image_std": self.image_std, "do_convert_rgb": self.do_convert_rgb, "max_num_tiles": self.max_num_tiles, "vision_input_type": self.vision_input_type, "resample": self.resample, "size": self.size, } def expected_output_image_shape(self, images): return self.num_channels, self.crop_size["height"], self.crop_size["width"] # Copied from tests.models.clip.test_image_processing_clip.CLIPImageProcessingTester.prepare_image_inputs def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False): return prepare_image_inputs( batch_size=self.batch_size, num_channels=self.num_channels, min_resolution=self.min_resolution, max_resolution=self.max_resolution, equal_resolution=equal_resolution, numpify=numpify, torchify=torchify, ) @require_torch @require_vision class PerceptionLMImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase): fast_image_processing_class = PerceptionLMImageProcessorFast if is_torchvision_available() else None test_slow_image_processor = False def setUp(self): super().setUp() self.image_processor_tester = PerceptionLMImageProcessingTester(self) @property # Copied from tests.models.clip.test_image_processing_clip.CLIPImageProcessingTest.image_processor_dict def image_processor_dict(self): return self.image_processor_tester.prepare_image_processor_dict() def test_image_processor_properties(self): for image_processing_class in self.image_processor_list: image_processing = image_processing_class(**self.image_processor_dict) self.assertTrue(hasattr(image_processing, "do_resize")) self.assertTrue(hasattr(image_processing, "tile_size")) self.assertTrue(hasattr(image_processing, "do_normalize")) self.assertTrue(hasattr(image_processing, "image_mean")) self.assertTrue(hasattr(image_processing, "image_std")) self.assertTrue(hasattr(image_processing, "do_convert_rgb")) self.assertTrue(hasattr(image_processing, "max_num_tiles")) self.assertTrue(hasattr(image_processing, "vision_input_type")) def test_image_processor_from_dict_with_kwargs(self): for image_processing_class in self.image_processor_list: image_processor = image_processing_class.from_dict(self.image_processor_dict) self.assertEqual(image_processor.tile_size, 16) self.assertEqual(image_processor.max_num_tiles, 4) self.assertEqual(image_processor.vision_input_type, "thumb+tile") image_processor = image_processing_class.from_dict( self.image_processor_dict, tile_size=42, max_num_tiles=9 ) self.assertEqual(image_processor.tile_size, 42) self.assertEqual(image_processor.max_num_tiles, 9) self.assertEqual(image_processor.vision_input_type, "thumb+tile") def test_call_pil(self): for image_processing_class in self.image_processor_list: # Initialize image_processing image_processing = image_processing_class(**self.image_processor_dict) # create random PIL images image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True) for image in image_inputs: self.assertIsInstance(image, Image.Image) # Test not batched input encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values expected_output_image_shape = (1, 5, 3, 16, 16) self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape) # Test batched encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values expected_output_image_shape = (7, 5, 3, 16, 16) self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape) def test_call_numpy(self): for image_processing_class in self.image_processor_list: # Initialize image_processing image_processing = image_processing_class(**self.image_processor_dict) # create random numpy tensors image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True, numpify=True) for image in image_inputs: self.assertIsInstance(image, np.ndarray) # Test not batched input encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values expected_output_image_shape = (1, 5, 3, 16, 16) self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape) # Test batched encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values expected_output_image_shape = (7, 5, 3, 16, 16) self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape) def test_call_pytorch(self): for image_processing_class in self.image_processor_list: # Initialize image_processing image_processing = image_processing_class(**self.image_processor_dict) # create random PyTorch tensors image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True, torchify=True) for image in image_inputs: self.assertIsInstance(image, torch.Tensor) # Test not batched input encoded_images = image_processing(image_inputs[0], return_tensors="pt").pixel_values expected_output_image_shape = (1, 5, 3, 16, 16) self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape) # Test batched encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values expected_output_image_shape = (7, 5, 3, 16, 16) self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape) @unittest.skip(reason="PerceptionLMImageProcessor doesn't treat 4 channel PIL and numpy consistently yet") def test_call_numpy_4_channels(self): pass def test_nested_input(self): for image_processing_class in self.image_processor_list: image_processing = image_processing_class(**self.image_processor_dict) image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=True) # Test batched as a list of images encoded_images = image_processing(image_inputs, return_tensors="pt").pixel_values expected_output_image_shape = (7, 5, 3, 16, 16) self.assertEqual(tuple(encoded_images.shape), expected_output_image_shape) # Test batched as a nested list of images, where each sublist is one batch image_inputs_nested = [image_inputs[:3], image_inputs[3:]] encoded_images_nested = image_processing(image_inputs_nested, return_tensors="pt").pixel_values expected_output_image_shape = (7, 5, 3, 16, 16) self.assertEqual(tuple(encoded_images_nested.shape), expected_output_image_shape) # Image processor should return same pixel values, independently of ipnut format self.assertTrue((encoded_images_nested == encoded_images).all())