# Copyright 2026 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 DEIMv2 model.""" import copy import inspect import math import tempfile import unittest from functools import cached_property from parameterized import parameterized from transformers import ( AutoImageProcessor, Deimv2Config, DINOv3ViTConfig, HGNetV2Config, is_torch_available, ) from transformers.testing_utils import ( require_torch, require_torch_accelerator, require_vision, slow, torch_device, ) if is_torch_available(): import torch from transformers import Deimv2ForObjectDetection, Deimv2Model from ...test_configuration_common import ConfigTester from ...test_modeling_common import ( TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION, ModelTesterMixin, _test_eager_matches_sdpa_inference, floats_tensor, ) from ...test_pipeline_mixin import PipelineTesterMixin from ...test_processing_common import url_to_local_path # TODO: Replace with the official Transformers ckpt once uploaded. CHECKPOINT = "harshaljanjani/DEIMv2_HGNetv2_N_COCO_Transformers" CHECKPOINT_LITE = "harshaljanjani/DEIMv2_HGNetv2_PICO_COCO_Transformers" CHECKPOINT_DINOV3 = "harshaljanjani/DEIMv2_DINOv3_S_COCO_Transformers" class Deimv2ModelTester: def __init__( self, parent, batch_size=3, is_training=True, use_labels=True, n_targets=3, num_labels=10, initializer_range=0.02, layer_norm_eps=1e-5, batch_norm_eps=1e-5, backbone_config=None, encoder_hidden_dim=32, encoder_in_channels=[128, 256, 512], feat_strides=[8, 16, 32], encoder_layers=1, encoder_ffn_dim=64, encoder_attention_heads=2, dropout=0.0, activation_dropout=0.0, encode_proj_layers=[2], positional_encoding_temperature=10000, encoder_activation_function="gelu", activation_function="silu", eval_size=None, normalize_before=False, d_model=32, num_queries=30, decoder_in_channels=[32, 32, 32], decoder_ffn_dim=64, num_feature_levels=3, decoder_n_points=[3, 6, 3], decoder_n_levels=3, decoder_layers=2, decoder_attention_heads=2, decoder_activation_function="relu", attention_dropout=0.0, num_denoising=0, label_noise_ratio=0.5, box_noise_scale=1.0, learn_initial_query=False, anchor_image_size=None, image_size=64, disable_custom_kernels=True, with_box_refine=True, decoder_offset_scale=0.5, eval_idx=-1, layer_scale=1, reg_max=32, reg_scale=4.0, depth_mult=0.34, hidden_expansion=0.5, ): self.parent = parent self.batch_size = batch_size self.num_channels = 3 self.is_training = is_training self.use_labels = use_labels self.n_targets = n_targets self.num_labels = num_labels self.initializer_range = initializer_range self.layer_norm_eps = layer_norm_eps self.batch_norm_eps = batch_norm_eps self.backbone_config = backbone_config self.encoder_hidden_dim = encoder_hidden_dim self.encoder_in_channels = encoder_in_channels self.feat_strides = feat_strides self.encoder_layers = encoder_layers self.encoder_ffn_dim = encoder_ffn_dim self.encoder_attention_heads = encoder_attention_heads self.dropout = dropout self.activation_dropout = activation_dropout self.encode_proj_layers = encode_proj_layers self.positional_encoding_temperature = positional_encoding_temperature self.encoder_activation_function = encoder_activation_function self.activation_function = activation_function self.eval_size = eval_size self.normalize_before = normalize_before self.d_model = d_model self.num_queries = num_queries self.decoder_in_channels = decoder_in_channels self.decoder_ffn_dim = decoder_ffn_dim self.num_feature_levels = num_feature_levels self.decoder_n_points = decoder_n_points self.decoder_n_levels = decoder_n_levels self.decoder_layers = decoder_layers self.decoder_attention_heads = decoder_attention_heads self.decoder_activation_function = decoder_activation_function self.attention_dropout = attention_dropout self.decoder_offset_scale = decoder_offset_scale self.eval_idx = eval_idx self.layer_scale = layer_scale self.reg_max = reg_max self.reg_scale = reg_scale self.depth_mult = depth_mult self.num_denoising = num_denoising self.label_noise_ratio = label_noise_ratio self.box_noise_scale = box_noise_scale self.learn_initial_query = learn_initial_query self.anchor_image_size = anchor_image_size self.image_size = image_size self.disable_custom_kernels = disable_custom_kernels self.with_box_refine = with_box_refine self.hidden_expansion = hidden_expansion self.encoder_seq_length = math.ceil(self.image_size / 32) * math.ceil(self.image_size / 32) def prepare_config_and_inputs(self): pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size]) pixel_mask = torch.ones([self.batch_size, self.image_size, self.image_size], device=torch_device) labels = None if self.use_labels: labels = [] for i in range(self.batch_size): target = {} target["class_labels"] = torch.randint( high=self.num_labels, size=(self.n_targets,), device=torch_device ) target["boxes"] = torch.rand(self.n_targets, 4, device=torch_device) labels.append(target) config = self.get_config() config.num_labels = self.num_labels return config, pixel_values, pixel_mask, labels def get_config(self): hidden_sizes = [64, 128, 256, 512] backbone_config = HGNetV2Config( stage_in_channels=[16, 64, 128, 256], stage_mid_channels=[16, 32, 64, 128], stage_out_channels=[64, 128, 256, 512], stage_num_blocks=[1, 1, 2, 1], stage_downsample=[False, True, True, True], stage_light_block=[False, False, True, True], stage_kernel_size=[3, 3, 5, 5], stage_numb_of_layers=[3, 3, 3, 3], embeddings_size=10, hidden_sizes=hidden_sizes, depths=[1, 1, 2, 1], out_features=["stage2", "stage3", "stage4"], out_indices=[2, 3, 4], stem_channels=[3, 16, 16], use_lab=True, ) return Deimv2Config( backbone_config=backbone_config, encoder_hidden_dim=self.encoder_hidden_dim, encoder_in_channels=self.encoder_in_channels, feat_strides=self.feat_strides, encoder_layers=self.encoder_layers, encoder_ffn_dim=self.encoder_ffn_dim, encoder_attention_heads=self.encoder_attention_heads, dropout=self.dropout, activation_dropout=self.activation_dropout, encode_proj_layers=self.encode_proj_layers, positional_encoding_temperature=self.positional_encoding_temperature, encoder_activation_function=self.encoder_activation_function, activation_function=self.activation_function, eval_size=self.eval_size, normalize_before=self.normalize_before, d_model=self.d_model, num_queries=self.num_queries, decoder_in_channels=self.decoder_in_channels, decoder_ffn_dim=self.decoder_ffn_dim, num_feature_levels=self.num_feature_levels, decoder_n_points=self.decoder_n_points, decoder_n_levels=self.decoder_n_levels, decoder_layers=self.decoder_layers, decoder_attention_heads=self.decoder_attention_heads, decoder_activation_function=self.decoder_activation_function, decoder_offset_scale=self.decoder_offset_scale, eval_idx=self.eval_idx, layer_scale=self.layer_scale, reg_max=self.reg_max, reg_scale=self.reg_scale, depth_mult=self.depth_mult, attention_dropout=self.attention_dropout, num_denoising=self.num_denoising, label_noise_ratio=self.label_noise_ratio, box_noise_scale=self.box_noise_scale, learn_initial_query=self.learn_initial_query, anchor_image_size=self.anchor_image_size, image_size=self.image_size, disable_custom_kernels=self.disable_custom_kernels, with_box_refine=self.with_box_refine, ) def prepare_config_and_inputs_for_common(self): config, pixel_values, pixel_mask, labels = self.prepare_config_and_inputs() inputs_dict = {"pixel_values": pixel_values} return config, inputs_dict def create_and_check_deimv2_model(self, config, pixel_values, pixel_mask, labels): model = Deimv2Model(config=config) model.to(torch_device) model.eval() result = model(pixel_values=pixel_values, pixel_mask=pixel_mask) result = model(pixel_values) self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.num_queries, self.d_model)) def create_and_check_deimv2_object_detection_head_model(self, config, pixel_values, pixel_mask, labels): model = Deimv2ForObjectDetection(config=config) model.to(torch_device) model.eval() result = model(pixel_values=pixel_values, pixel_mask=pixel_mask) result = model(pixel_values) self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_queries, self.num_labels)) self.parent.assertEqual(result.pred_boxes.shape, (self.batch_size, self.num_queries, 4)) result = model(pixel_values=pixel_values, pixel_mask=pixel_mask, labels=labels) self.parent.assertEqual(result.loss.shape, ()) self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_queries, self.num_labels)) self.parent.assertEqual(result.pred_boxes.shape, (self.batch_size, self.num_queries, 4)) @require_torch class Deimv2ModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase): all_model_classes = (Deimv2Model, Deimv2ForObjectDetection) if is_torch_available() else () pipeline_model_mapping = ( {"image-feature-extraction": Deimv2Model, "object-detection": Deimv2ForObjectDetection} if is_torch_available() else {} ) is_encoder_decoder = True test_resize_embeddings = False test_missing_keys = False def _prepare_for_class(self, inputs_dict, model_class, return_labels=False): inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels) if return_labels: if model_class.__name__ == "Deimv2ForObjectDetection": labels = [] for i in range(self.model_tester.batch_size): target = {} target["class_labels"] = torch.ones( size=(self.model_tester.n_targets,), device=torch_device, dtype=torch.long ) target["boxes"] = torch.ones( self.model_tester.n_targets, 4, device=torch_device, dtype=torch.float ) labels.append(target) inputs_dict["labels"] = labels return inputs_dict def setUp(self): self.model_tester = Deimv2ModelTester(self) self.config_tester = ConfigTester( self, config_class=Deimv2Config, has_text_modality=False, common_properties=["hidden_size", "num_attention_heads"], ) def test_config(self): self.config_tester.run_common_tests() def test_deimv2_model(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_deimv2_model(*config_and_inputs) def test_deimv2_object_detection_head_model(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_deimv2_object_detection_head_model(*config_and_inputs) @unittest.skip(reason="Multi-scale deformable attention is incompatible with nn.DataParallel") def test_multi_gpu_data_parallel_forward(self): pass @unittest.skip( reason="Deimv2 is a vision model but inputs_embeds is in the forward signature (inherited from D-FINE)" ) def test_inputs_embeds(self): pass @unittest.skip(reason="Forward signature has inputs_embeds but no input_ids") def test_inputs_embeds_matches_input_ids(self): pass @unittest.skip(reason="Base test asserts get_input_embeddings() returns nn.Embedding which vision models lack") def test_model_get_set_embeddings(self): pass @unittest.skip(reason="Decoder heads are shared via reference assignment so untied saving is not applicable") def test_load_save_without_tied_weights(self): pass # Override: Multi-scale deformable attention outputs have different shapes than standard self-attention def test_attention_outputs(self): config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config.return_dict = True 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._from_config(config, attn_implementation="eager") config = model.config model.to(torch_device) model.eval() with torch.no_grad(): outputs = model(**self._prepare_for_class(inputs_dict, model_class)) attentions = outputs.encoder_attentions self.assertEqual(len(attentions), self.model_tester.encoder_layers) 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.encoder_attentions self.assertEqual(len(attentions), self.model_tester.encoder_layers) self.assertListEqual( list(attentions[0].shape[-3:]), [ self.model_tester.encoder_attention_heads, self.model_tester.encoder_seq_length, self.model_tester.encoder_seq_length, ], ) out_len = len(outputs) correct_outlen = 15 if "labels" in inputs_dict: correct_outlen += 1 if model_class.__name__ == "Deimv2ForObjectDetection": correct_outlen += 2 self.assertEqual(out_len, correct_outlen) decoder_attentions = outputs.decoder_attentions self.assertIsInstance(decoder_attentions, (list, tuple)) self.assertEqual(len(decoder_attentions), self.model_tester.decoder_layers) self.assertListEqual( list(decoder_attentions[0].shape[-3:]), [ self.model_tester.decoder_attention_heads, self.model_tester.num_queries, self.model_tester.num_queries, ], ) cross_attentions = outputs.cross_attentions self.assertIsInstance(cross_attentions, (list, tuple)) self.assertEqual(len(cross_attentions), self.model_tester.decoder_layers) self.assertListEqual( list(cross_attentions[0].shape[-3:]), [ self.model_tester.num_queries, self.model_tester.decoder_attention_heads, self.model_tester.decoder_n_levels * self.model_tester.decoder_n_points if isinstance(self.model_tester.decoder_n_points, int) else sum(self.model_tester.decoder_n_points), ], ) 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)) if hasattr(self.model_tester, "num_hidden_states_types"): added_hidden_states = self.model_tester.num_hidden_states_types else: added_hidden_states = 2 self.assertEqual(out_len + added_hidden_states, len(outputs)) self_attentions = outputs.encoder_attentions self.assertEqual(len(self_attentions), self.model_tester.encoder_layers) self.assertListEqual( list(self_attentions[0].shape[-3:]), [ self.model_tester.encoder_attention_heads, self.model_tester.encoder_seq_length, self.model_tester.encoder_seq_length, ], ) 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", len(self.model_tester.encoder_in_channels) - 1 ) self.assertEqual(len(hidden_states), expected_num_layers) self.assertListEqual( list(hidden_states[1].shape[-2:]), [ self.model_tester.image_size // self.model_tester.feat_strides[-1], self.model_tester.image_size // self.model_tester.feat_strides[-1], ], ) if config.is_encoder_decoder: hidden_states = outputs.decoder_hidden_states expected_num_layers = getattr( self.model_tester, "expected_num_hidden_layers", self.model_tester.decoder_layers + 1 ) self.assertIsInstance(hidden_states, (list, tuple)) self.assertEqual(len(hidden_states), expected_num_layers) self.assertListEqual( list(hidden_states[0].shape[-2:]), [self.model_tester.num_queries, self.model_tester.d_model], ) 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) del inputs_dict["output_hidden_states"] config.output_hidden_states = True check_hidden_states_output(inputs_dict, config, model_class) # Override: Custom gradient retention check for multi-scale deformable attention outputs def test_retain_grad_hidden_states_attentions(self): config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config.output_hidden_states = True config.output_attentions = True model_class = self.all_model_classes[0] model = model_class(config) model.to(torch_device) inputs = self._prepare_for_class(inputs_dict, model_class) outputs = model(**inputs) output = outputs[0] encoder_hidden_states = outputs.encoder_hidden_states[0] encoder_attentions = outputs.encoder_attentions[0] encoder_hidden_states.retain_grad() encoder_attentions.retain_grad() decoder_attentions = outputs.decoder_attentions[0] decoder_attentions.retain_grad() cross_attentions = outputs.cross_attentions[0] cross_attentions.retain_grad() output.flatten()[0].backward(retain_graph=True) self.assertIsNotNone(encoder_hidden_states.grad) self.assertIsNotNone(encoder_attentions.grad) self.assertIsNotNone(decoder_attentions.grad) self.assertIsNotNone(cross_attentions.grad) # Override: Deimv2 uses pixel_values as main input, not input_ids 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) arg_names = [*signature.parameters.keys()] expected_arg_names = ["pixel_values"] self.assertListEqual(arg_names[:1], expected_arg_names) def test_backbone_selection(self): config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() def _validate_backbone_init(config): for model_class in self.all_model_classes: model = model_class(copy.deepcopy(config)) model.to(torch_device) model.eval() with torch.no_grad(): outputs = model(**self._prepare_for_class(inputs_dict, model_class)) if model_class.__name__ == "Deimv2ForObjectDetection": expected_shape = ( self.model_tester.batch_size, self.model_tester.num_queries, self.model_tester.num_labels, ) self.assertEqual(outputs.logits.shape, expected_shape) self.assertEqual(len(model.model.conv_encoder.intermediate_channel_sizes), 3) else: self.assertEqual(len(model.conv_encoder.intermediate_channel_sizes), 3) self.assertTrue(outputs) config_dict = config.to_dict() config_dict["encoder_in_channels"] = [24, 40, 432] config_dict["backbone"] = "tf_mobilenetv3_small_075" config_dict["backbone_config"] = None config_dict["use_timm_backbone"] = True config_dict["backbone_kwargs"] = {"out_indices": [2, 3, 4]} config = config.__class__(**config_dict) _validate_backbone_init(config) config_dict = config.to_dict() config_dict["backbone"] = "microsoft/resnet-18" config_dict["backbone_config"] = None config_dict["use_timm_backbone"] = False config_dict["use_pretrained_backbone"] = True config_dict["backbone_kwargs"] = {"out_indices": [2, 3, 4]} config = config.__class__(**config_dict) _validate_backbone_init(config) @parameterized.expand(["float32", "float16", "bfloat16"]) @require_torch_accelerator @slow def test_inference_with_different_dtypes(self, dtype_str): dtype = { "float32": torch.float32, "float16": torch.float16, "bfloat16": torch.bfloat16, }[dtype_str] 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).to(dtype) model.eval() for key, tensor in inputs_dict.items(): if tensor.dtype == torch.float32: inputs_dict[key] = tensor.to(dtype) with torch.no_grad(): _ = model(**self._prepare_for_class(inputs_dict, model_class)) @parameterized.expand(["float32", "float16", "bfloat16"]) @require_torch_accelerator @slow def test_inference_equivalence_for_static_and_dynamic_anchors(self, dtype_str): dtype = { "float32": torch.float32, "float16": torch.float16, "bfloat16": torch.bfloat16, }[dtype_str] config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() h, w = inputs_dict["pixel_values"].shape[-2:] for key, tensor in inputs_dict.items(): if tensor.dtype == torch.float32: inputs_dict[key] = tensor.to(dtype) for model_class in self.all_model_classes: with tempfile.TemporaryDirectory() as tmpdirname: model_class(config).save_pretrained(tmpdirname) model_static = model_class.from_pretrained( tmpdirname, anchor_image_size=[h, w], device_map=torch_device, dtype=dtype ).eval() model_dynamic = model_class.from_pretrained( tmpdirname, anchor_image_size=None, device_map=torch_device, dtype=dtype ).eval() self.assertIsNotNone(model_static.config.anchor_image_size) self.assertIsNone(model_dynamic.config.anchor_image_size) with torch.no_grad(): outputs_static = model_static(**self._prepare_for_class(inputs_dict, model_class)) outputs_dynamic = model_dynamic(**self._prepare_for_class(inputs_dict, model_class)) torch.testing.assert_close( outputs_static.last_hidden_state, outputs_dynamic.last_hidden_state, rtol=1e-4, atol=1e-4 ) class Deimv2LiteEncoderModelTester: def __init__( self, parent, batch_size=3, is_training=True, use_labels=True, n_targets=3, num_labels=10, initializer_range=0.02, layer_norm_eps=1e-5, batch_norm_eps=1e-5, encoder_hidden_dim=32, encoder_in_channels=[256], feat_strides=[16, 32], dropout=0.0, activation_dropout=0.0, positional_encoding_temperature=10000, encoder_activation_function="gelu", activation_function="silu", eval_size=None, normalize_before=False, d_model=32, num_queries=10, decoder_in_channels=[32, 32], decoder_ffn_dim=64, num_feature_levels=2, decoder_n_points=[4, 2], decoder_n_levels=2, decoder_layers=2, decoder_attention_heads=2, decoder_activation_function="relu", attention_dropout=0.0, num_denoising=0, label_noise_ratio=0.5, box_noise_scale=1.0, learn_initial_query=False, anchor_image_size=None, image_size=64, disable_custom_kernels=True, with_box_refine=True, decoder_offset_scale=0.5, eval_idx=-1, layer_scale=1, reg_max=32, reg_scale=4.0, depth_mult=0.34, hidden_expansion=0.5, ): self.parent = parent self.batch_size = batch_size self.num_channels = 3 self.is_training = is_training self.use_labels = use_labels self.n_targets = n_targets self.num_labels = num_labels self.initializer_range = initializer_range self.layer_norm_eps = layer_norm_eps self.batch_norm_eps = batch_norm_eps self.encoder_hidden_dim = encoder_hidden_dim self.encoder_in_channels = encoder_in_channels self.feat_strides = feat_strides self.encoder_layers = 0 self.encoder_ffn_dim = 64 self.encoder_attention_heads = 2 self.dropout = dropout self.activation_dropout = activation_dropout self.encode_proj_layers = [] self.positional_encoding_temperature = positional_encoding_temperature self.encoder_activation_function = encoder_activation_function self.activation_function = activation_function self.eval_size = eval_size self.normalize_before = normalize_before self.d_model = d_model self.num_queries = num_queries self.decoder_in_channels = decoder_in_channels self.decoder_ffn_dim = decoder_ffn_dim self.num_feature_levels = num_feature_levels self.decoder_n_points = decoder_n_points self.decoder_n_levels = decoder_n_levels self.decoder_layers = decoder_layers self.decoder_attention_heads = decoder_attention_heads self.decoder_activation_function = decoder_activation_function self.attention_dropout = attention_dropout self.decoder_offset_scale = decoder_offset_scale self.eval_idx = eval_idx self.layer_scale = layer_scale self.reg_max = reg_max self.reg_scale = reg_scale self.depth_mult = depth_mult self.num_denoising = num_denoising self.label_noise_ratio = label_noise_ratio self.box_noise_scale = box_noise_scale self.learn_initial_query = learn_initial_query self.anchor_image_size = anchor_image_size self.image_size = image_size self.disable_custom_kernels = disable_custom_kernels self.with_box_refine = with_box_refine self.hidden_expansion = hidden_expansion def prepare_config_and_inputs(self): pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size]) pixel_mask = torch.ones([self.batch_size, self.image_size, self.image_size], device=torch_device) labels = None if self.use_labels: labels = [] for i in range(self.batch_size): target = {} target["class_labels"] = torch.randint( high=self.num_labels, size=(self.n_targets,), device=torch_device ) target["boxes"] = torch.rand(self.n_targets, 4, device=torch_device) labels.append(target) config = self.get_config() config.num_labels = self.num_labels return config, pixel_values, pixel_mask, labels def get_config(self): backbone_config = HGNetV2Config( stage_in_channels=[16, 64, 128], stage_mid_channels=[16, 32, 64], stage_out_channels=[64, 128, 256], stage_num_blocks=[1, 1, 1], stage_downsample=[False, True, True], stage_light_block=[False, False, True], stage_kernel_size=[3, 3, 3], stage_numb_of_layers=[3, 3, 3], embeddings_size=10, hidden_sizes=[64, 128, 256], depths=[1, 1, 1], out_features=["stage3"], out_indices=[3], stem_channels=[3, 16, 16], use_lab=True, ) return Deimv2Config( backbone_config=backbone_config, encoder_hidden_dim=self.encoder_hidden_dim, encoder_in_channels=self.encoder_in_channels, feat_strides=self.feat_strides, encoder_layers=self.encoder_layers, encoder_ffn_dim=self.encoder_ffn_dim, encoder_attention_heads=self.encoder_attention_heads, dropout=self.dropout, activation_dropout=self.activation_dropout, encode_proj_layers=self.encode_proj_layers, positional_encoding_temperature=self.positional_encoding_temperature, encoder_activation_function=self.encoder_activation_function, activation_function=self.activation_function, eval_size=self.eval_size, normalize_before=self.normalize_before, d_model=self.d_model, num_queries=self.num_queries, decoder_in_channels=self.decoder_in_channels, decoder_ffn_dim=self.decoder_ffn_dim, num_feature_levels=self.num_feature_levels, decoder_n_points=self.decoder_n_points, decoder_n_levels=self.decoder_n_levels, decoder_layers=self.decoder_layers, decoder_attention_heads=self.decoder_attention_heads, decoder_activation_function=self.decoder_activation_function, decoder_offset_scale=self.decoder_offset_scale, eval_idx=self.eval_idx, layer_scale=self.layer_scale, reg_max=self.reg_max, reg_scale=self.reg_scale, depth_mult=self.depth_mult, attention_dropout=self.attention_dropout, num_denoising=self.num_denoising, label_noise_ratio=self.label_noise_ratio, box_noise_scale=self.box_noise_scale, learn_initial_query=self.learn_initial_query, anchor_image_size=self.anchor_image_size, image_size=self.image_size, disable_custom_kernels=self.disable_custom_kernels, with_box_refine=self.with_box_refine, encoder_type="lite", use_gateway=False, share_bbox_head=False, ) def prepare_config_and_inputs_for_common(self): config, pixel_values, pixel_mask, labels = self.prepare_config_and_inputs() inputs_dict = {"pixel_values": pixel_values} return config, inputs_dict def create_and_check_deimv2_model(self, config, pixel_values, pixel_mask, labels): model = Deimv2Model(config=config) model.to(torch_device) model.eval() result = model(pixel_values=pixel_values, pixel_mask=pixel_mask) result = model(pixel_values) self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.num_queries, self.d_model)) def create_and_check_deimv2_object_detection_head_model(self, config, pixel_values, pixel_mask, labels): model = Deimv2ForObjectDetection(config=config) model.to(torch_device) model.eval() result = model(pixel_values=pixel_values, pixel_mask=pixel_mask) result = model(pixel_values) self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_queries, self.num_labels)) self.parent.assertEqual(result.pred_boxes.shape, (self.batch_size, self.num_queries, 4)) result = model(pixel_values=pixel_values, pixel_mask=pixel_mask, labels=labels) self.parent.assertEqual(result.loss.shape, ()) self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_queries, self.num_labels)) self.parent.assertEqual(result.pred_boxes.shape, (self.batch_size, self.num_queries, 4)) @require_torch class Deimv2LiteEncoderModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase): all_model_classes = (Deimv2Model, Deimv2ForObjectDetection) if is_torch_available() else () pipeline_model_mapping = ( {"image-feature-extraction": Deimv2Model, "object-detection": Deimv2ForObjectDetection} if is_torch_available() else {} ) is_encoder_decoder = True test_resize_embeddings = False has_attentions = False test_missing_keys = False def _prepare_for_class(self, inputs_dict, model_class, return_labels=False): inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels) if return_labels: if model_class.__name__ == "Deimv2ForObjectDetection": labels = [] for i in range(self.model_tester.batch_size): target = {} target["class_labels"] = torch.ones( size=(self.model_tester.n_targets,), device=torch_device, dtype=torch.long ) target["boxes"] = torch.ones( self.model_tester.n_targets, 4, device=torch_device, dtype=torch.float ) labels.append(target) inputs_dict["labels"] = labels return inputs_dict def setUp(self): self.model_tester = Deimv2LiteEncoderModelTester(self) self.config_tester = ConfigTester( self, config_class=Deimv2Config, has_text_modality=False, common_properties=["hidden_size", "num_attention_heads"], ) def test_config(self): self.config_tester.run_common_tests() def test_deimv2_lite_encoder_model(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_deimv2_model(*config_and_inputs) def test_deimv2_lite_encoder_object_detection_head_model(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_deimv2_object_detection_head_model(*config_and_inputs) @unittest.skip(reason="Multi-scale deformable attention is incompatible with nn.DataParallel") def test_multi_gpu_data_parallel_forward(self): pass @unittest.skip( reason="Deimv2 is a vision model but inputs_embeds is in the forward signature (inherited from D-FINE)" ) def test_inputs_embeds(self): pass @unittest.skip(reason="Forward signature has inputs_embeds but no input_ids") def test_inputs_embeds_matches_input_ids(self): pass @unittest.skip(reason="Base test asserts get_input_embeddings() returns nn.Embedding which vision models lack") def test_model_get_set_embeddings(self): pass @unittest.skip(reason="Decoder heads are shared via reference assignment so untied saving is not applicable") def test_load_save_without_tied_weights(self): pass @unittest.skip( reason="LiteEncoder has no encoder_hidden_states so the base test fails accessing encoder_hidden_states[0]" ) def test_retain_grad_hidden_states_attentions(self): pass # Override: LiteEncoder has no encoder hidden states, only decoder hidden states 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)) if config.is_encoder_decoder: hidden_states = outputs.decoder_hidden_states expected_num_layers = self.model_tester.decoder_layers + 1 self.assertIsInstance(hidden_states, (list, tuple)) self.assertEqual(len(hidden_states), expected_num_layers) self.assertListEqual( list(hidden_states[0].shape[-2:]), [self.model_tester.num_queries, self.model_tester.d_model], ) 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) del inputs_dict["output_hidden_states"] config.output_hidden_states = True check_hidden_states_output(inputs_dict, config, model_class) # Override: Deimv2 uses pixel_values as main input, not input_ids 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) arg_names = [*signature.parameters.keys()] expected_arg_names = ["pixel_values"] self.assertListEqual(arg_names[:1], expected_arg_names) class Deimv2DINOv3ModelTester: def __init__( self, parent, batch_size=3, is_training=True, use_labels=True, n_targets=3, num_labels=10, initializer_range=0.02, layer_norm_eps=1e-5, batch_norm_eps=1e-5, encoder_hidden_dim=32, encoder_in_channels=[32, 32, 32], feat_strides=[8, 16, 32], encoder_layers=1, encoder_ffn_dim=64, encoder_attention_heads=2, dropout=0.0, activation_dropout=0.0, encode_proj_layers=[2], positional_encoding_temperature=10000, encoder_activation_function="gelu", activation_function="silu", eval_size=None, normalize_before=False, d_model=32, num_queries=30, decoder_in_channels=[32, 32, 32], decoder_ffn_dim=64, num_feature_levels=3, decoder_n_points=4, decoder_n_levels=3, decoder_layers=2, decoder_attention_heads=2, decoder_activation_function="relu", attention_dropout=0.0, num_denoising=0, label_noise_ratio=0.5, box_noise_scale=1.0, learn_initial_query=False, anchor_image_size=None, image_size=64, disable_custom_kernels=True, with_box_refine=True, decoder_offset_scale=0.5, eval_idx=-1, layer_scale=1, reg_max=32, reg_scale=4.0, depth_mult=0.34, hidden_expansion=0.5, sta_inplanes=8, ): self.parent = parent self.batch_size = batch_size self.num_channels = 3 self.is_training = is_training self.use_labels = use_labels self.n_targets = n_targets self.num_labels = num_labels self.initializer_range = initializer_range self.layer_norm_eps = layer_norm_eps self.batch_norm_eps = batch_norm_eps self.encoder_hidden_dim = encoder_hidden_dim self.encoder_in_channels = encoder_in_channels self.feat_strides = feat_strides self.encoder_layers = encoder_layers self.encoder_ffn_dim = encoder_ffn_dim self.encoder_attention_heads = encoder_attention_heads self.dropout = dropout self.activation_dropout = activation_dropout self.encode_proj_layers = encode_proj_layers self.positional_encoding_temperature = positional_encoding_temperature self.encoder_activation_function = encoder_activation_function self.activation_function = activation_function self.eval_size = eval_size self.normalize_before = normalize_before self.d_model = d_model self.num_queries = num_queries self.decoder_in_channels = decoder_in_channels self.decoder_ffn_dim = decoder_ffn_dim self.num_feature_levels = num_feature_levels self.decoder_n_points = decoder_n_points self.decoder_n_levels = decoder_n_levels self.decoder_layers = decoder_layers self.decoder_attention_heads = decoder_attention_heads self.decoder_activation_function = decoder_activation_function self.attention_dropout = attention_dropout self.decoder_offset_scale = decoder_offset_scale self.eval_idx = eval_idx self.layer_scale = layer_scale self.reg_max = reg_max self.reg_scale = reg_scale self.depth_mult = depth_mult self.num_denoising = num_denoising self.label_noise_ratio = label_noise_ratio self.box_noise_scale = box_noise_scale self.learn_initial_query = learn_initial_query self.anchor_image_size = anchor_image_size self.image_size = image_size self.disable_custom_kernels = disable_custom_kernels self.with_box_refine = with_box_refine self.hidden_expansion = hidden_expansion self.sta_inplanes = sta_inplanes self.encoder_seq_length = math.ceil(self.image_size / 32) * math.ceil(self.image_size / 32) def prepare_config_and_inputs(self): pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size]) pixel_mask = torch.ones([self.batch_size, self.image_size, self.image_size], device=torch_device) labels = None if self.use_labels: labels = [] for i in range(self.batch_size): target = {} target["class_labels"] = torch.randint( high=self.num_labels, size=(self.n_targets,), device=torch_device ) target["boxes"] = torch.rand(self.n_targets, 4, device=torch_device) labels.append(target) config = self.get_config() config.num_labels = self.num_labels return config, pixel_values, pixel_mask, labels def get_config(self): backbone_config = DINOv3ViTConfig( hidden_size=32, num_attention_heads=2, num_hidden_layers=4, intermediate_size=64, num_register_tokens=1, layerscale_value=1.0, use_gated_mlp=False, rope_theta=100.0, patch_size=16, image_size=self.image_size, out_indices=[2, 3, 4], apply_layernorm=False, reshape_hidden_states=True, ) return Deimv2Config( backbone_config=backbone_config, encoder_hidden_dim=self.encoder_hidden_dim, encoder_in_channels=self.encoder_in_channels, feat_strides=self.feat_strides, encoder_layers=self.encoder_layers, encoder_ffn_dim=self.encoder_ffn_dim, encoder_attention_heads=self.encoder_attention_heads, dropout=self.dropout, activation_dropout=self.activation_dropout, encode_proj_layers=self.encode_proj_layers, positional_encoding_temperature=self.positional_encoding_temperature, encoder_activation_function=self.encoder_activation_function, activation_function=self.activation_function, eval_size=self.eval_size, normalize_before=self.normalize_before, d_model=self.d_model, num_queries=self.num_queries, decoder_in_channels=self.decoder_in_channels, decoder_ffn_dim=self.decoder_ffn_dim, num_feature_levels=self.num_feature_levels, decoder_n_points=self.decoder_n_points, decoder_n_levels=self.decoder_n_levels, decoder_layers=self.decoder_layers, decoder_attention_heads=self.decoder_attention_heads, decoder_activation_function=self.decoder_activation_function, decoder_offset_scale=self.decoder_offset_scale, eval_idx=self.eval_idx, layer_scale=self.layer_scale, reg_max=self.reg_max, reg_scale=self.reg_scale, depth_mult=self.depth_mult, attention_dropout=self.attention_dropout, num_denoising=self.num_denoising, label_noise_ratio=self.label_noise_ratio, box_noise_scale=self.box_noise_scale, learn_initial_query=self.learn_initial_query, anchor_image_size=self.anchor_image_size, image_size=self.image_size, disable_custom_kernels=self.disable_custom_kernels, with_box_refine=self.with_box_refine, sta_inplanes=self.sta_inplanes, encoder_has_trailing_conv=False, ) def prepare_config_and_inputs_for_common(self): config, pixel_values, pixel_mask, labels = self.prepare_config_and_inputs() inputs_dict = {"pixel_values": pixel_values} return config, inputs_dict def create_and_check_deimv2_model(self, config, pixel_values, pixel_mask, labels): model = Deimv2Model(config=config) model.to(torch_device) model.eval() result = model(pixel_values=pixel_values, pixel_mask=pixel_mask) result = model(pixel_values) self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.num_queries, self.d_model)) def create_and_check_deimv2_object_detection_head_model(self, config, pixel_values, pixel_mask, labels): model = Deimv2ForObjectDetection(config=config) model.to(torch_device) model.eval() result = model(pixel_values=pixel_values, pixel_mask=pixel_mask) result = model(pixel_values) self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_queries, self.num_labels)) self.parent.assertEqual(result.pred_boxes.shape, (self.batch_size, self.num_queries, 4)) result = model(pixel_values=pixel_values, pixel_mask=pixel_mask, labels=labels) self.parent.assertEqual(result.loss.shape, ()) self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_queries, self.num_labels)) self.parent.assertEqual(result.pred_boxes.shape, (self.batch_size, self.num_queries, 4)) @require_torch class Deimv2DINOv3ModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase): all_model_classes = (Deimv2Model, Deimv2ForObjectDetection) if is_torch_available() else () pipeline_model_mapping = ( {"image-feature-extraction": Deimv2Model, "object-detection": Deimv2ForObjectDetection} if is_torch_available() else {} ) is_encoder_decoder = True test_resize_embeddings = False test_missing_keys = False def _prepare_for_class(self, inputs_dict, model_class, return_labels=False): inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels) if return_labels: if model_class.__name__ == "Deimv2ForObjectDetection": labels = [] for i in range(self.model_tester.batch_size): target = {} target["class_labels"] = torch.ones( size=(self.model_tester.n_targets,), device=torch_device, dtype=torch.long ) target["boxes"] = torch.ones( self.model_tester.n_targets, 4, device=torch_device, dtype=torch.float ) labels.append(target) inputs_dict["labels"] = labels return inputs_dict def setUp(self): self.model_tester = Deimv2DINOv3ModelTester(self) self.config_tester = ConfigTester( self, config_class=Deimv2Config, has_text_modality=False, common_properties=["hidden_size", "num_attention_heads"], ) def test_config(self): self.config_tester.run_common_tests() def test_deimv2_dinov3_model(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_deimv2_model(*config_and_inputs) def test_deimv2_dinov3_object_detection_head_model(self): config_and_inputs = self.model_tester.prepare_config_and_inputs() self.model_tester.create_and_check_deimv2_object_detection_head_model(*config_and_inputs) @unittest.skip(reason="Multi-scale deformable attention is incompatible with nn.DataParallel") def test_multi_gpu_data_parallel_forward(self): pass @unittest.skip( reason="Deimv2 is a vision model but inputs_embeds is in the forward signature (inherited from D-FINE)" ) def test_inputs_embeds(self): pass @unittest.skip(reason="Forward signature has inputs_embeds but no input_ids") def test_inputs_embeds_matches_input_ids(self): pass @unittest.skip(reason="Base test asserts get_input_embeddings() returns nn.Embedding which vision models lack") def test_model_get_set_embeddings(self): pass @unittest.skip(reason="Decoder heads are shared via reference assignment so untied saving is not applicable") def test_load_save_without_tied_weights(self): pass @unittest.skip(reason="DINOv3 RoPE with dynamic interpolation causes torch.compile inductor overflow") def test_sdpa_can_compile_dynamic(self): pass # Override: DINOv3 backbone requires wider tolerances for SDPA vs eager comparison @parameterized.expand(TEST_EAGER_MATCHES_SDPA_INFERENCE_PARAMETERIZATION) def test_eager_matches_sdpa_inference( self, name, torch_dtype, padding_side, use_attention_mask, output_attentions, enable_kernels ): atols = { ("cpu", False, torch.float32): 1e-4, ("cpu", False, torch.float16): 5e-3, ("cpu", False, torch.bfloat16): 1e-2, ("cpu", True, torch.float32): 1e-4, ("cpu", True, torch.float16): 5e-3, ("cpu", True, torch.bfloat16): 1e-2, ("cuda", False, torch.float32): 1e-4, ("cuda", False, torch.bfloat16): 1e-2, ("cuda", False, torch.float16): 5e-3, ("cuda", True, torch.float32): 1e-4, ("cuda", True, torch.bfloat16): 1e-2, ("cuda", True, torch.float16): 5e-3, } rtols = { ("cpu", False, torch.float32): 1e-3, ("cpu", False, torch.float16): 5e-3, ("cpu", False, torch.bfloat16): 1e-2, ("cpu", True, torch.float32): 1e-3, ("cpu", True, torch.float16): 5e-3, ("cpu", True, torch.bfloat16): 1e-2, ("cuda", False, torch.float32): 1e-3, ("cuda", False, torch.bfloat16): 1e-2, ("cuda", False, torch.float16): 5e-3, ("cuda", True, torch.float32): 1e-3, ("cuda", True, torch.bfloat16): 3e-2, ("cuda", True, torch.float16): 5e-3, } _test_eager_matches_sdpa_inference( self, name, torch_dtype, padding_side, use_attention_mask, output_attentions, enable_kernels, atols=atols, rtols=rtols, ) # Override: DINOv3 backbone numerical precision requires wider tolerances def test_batching_equivalence(self): super().test_batching_equivalence(atol=1e-4, rtol=1e-4) @unittest.skip(reason="Flex attention test requires decoder_input_ids which detection models don't have") def test_flex_attention_with_grads(self): pass # Override: Multi-scale deformable attention outputs have different shapes than standard self-attention def test_attention_outputs(self): config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config.return_dict = True 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._from_config(config, attn_implementation="eager") config = model.config model.to(torch_device) model.eval() with torch.no_grad(): outputs = model(**self._prepare_for_class(inputs_dict, model_class)) attentions = outputs.encoder_attentions self.assertEqual(len(attentions), self.model_tester.encoder_layers) 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.encoder_attentions self.assertEqual(len(attentions), self.model_tester.encoder_layers) self.assertListEqual( list(attentions[0].shape[-3:]), [ self.model_tester.encoder_attention_heads, self.model_tester.encoder_seq_length, self.model_tester.encoder_seq_length, ], ) out_len = len(outputs) correct_outlen = 15 if "labels" in inputs_dict: correct_outlen += 1 if model_class.__name__ == "Deimv2ForObjectDetection": correct_outlen += 2 self.assertEqual(out_len, correct_outlen) decoder_attentions = outputs.decoder_attentions self.assertIsInstance(decoder_attentions, (list, tuple)) self.assertEqual(len(decoder_attentions), self.model_tester.decoder_layers) self.assertListEqual( list(decoder_attentions[0].shape[-3:]), [ self.model_tester.decoder_attention_heads, self.model_tester.num_queries, self.model_tester.num_queries, ], ) cross_attentions = outputs.cross_attentions self.assertIsInstance(cross_attentions, (list, tuple)) self.assertEqual(len(cross_attentions), self.model_tester.decoder_layers) self.assertListEqual( list(cross_attentions[0].shape[-3:]), [ self.model_tester.num_queries, self.model_tester.decoder_attention_heads, self.model_tester.decoder_n_levels * self.model_tester.decoder_n_points if isinstance(self.model_tester.decoder_n_points, int) else sum(self.model_tester.decoder_n_points), ], ) 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)) if hasattr(self.model_tester, "num_hidden_states_types"): added_hidden_states = self.model_tester.num_hidden_states_types else: added_hidden_states = 2 self.assertEqual(out_len + added_hidden_states, len(outputs)) self_attentions = outputs.encoder_attentions self.assertEqual(len(self_attentions), self.model_tester.encoder_layers) self.assertListEqual( list(self_attentions[0].shape[-3:]), [ self.model_tester.encoder_attention_heads, self.model_tester.encoder_seq_length, self.model_tester.encoder_seq_length, ], ) # Override: Encoder hidden states are multi-scale feature maps, not a standard sequence of layer outputs 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", len(self.model_tester.encoder_in_channels) - 1 ) self.assertEqual(len(hidden_states), expected_num_layers) self.assertListEqual( list(hidden_states[1].shape[-2:]), [ self.model_tester.image_size // self.model_tester.feat_strides[-1], self.model_tester.image_size // self.model_tester.feat_strides[-1], ], ) if config.is_encoder_decoder: hidden_states = outputs.decoder_hidden_states expected_num_layers = getattr( self.model_tester, "expected_num_hidden_layers", self.model_tester.decoder_layers + 1 ) self.assertIsInstance(hidden_states, (list, tuple)) self.assertEqual(len(hidden_states), expected_num_layers) self.assertListEqual( list(hidden_states[0].shape[-2:]), [self.model_tester.num_queries, self.model_tester.d_model], ) 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) del inputs_dict["output_hidden_states"] config.output_hidden_states = True check_hidden_states_output(inputs_dict, config, model_class) # Override: Custom gradient retention check for multi-scale deformable attention outputs def test_retain_grad_hidden_states_attentions(self): config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() config.output_hidden_states = True config.output_attentions = True model_class = self.all_model_classes[0] model = model_class(config) model.to(torch_device) inputs = self._prepare_for_class(inputs_dict, model_class) outputs = model(**inputs) output = outputs[0] encoder_hidden_states = outputs.encoder_hidden_states[0] encoder_attentions = outputs.encoder_attentions[0] encoder_hidden_states.retain_grad() encoder_attentions.retain_grad() decoder_attentions = outputs.decoder_attentions[0] decoder_attentions.retain_grad() cross_attentions = outputs.cross_attentions[0] cross_attentions.retain_grad() output.flatten()[0].backward(retain_graph=True) self.assertIsNotNone(encoder_hidden_states.grad) self.assertIsNotNone(encoder_attentions.grad) self.assertIsNotNone(decoder_attentions.grad) self.assertIsNotNone(cross_attentions.grad) # Override: Deimv2 uses pixel_values as main input, not input_ids 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) arg_names = [*signature.parameters.keys()] expected_arg_names = ["pixel_values"] self.assertListEqual(arg_names[:1], expected_arg_names) @parameterized.expand(["float32", "float16", "bfloat16"]) @require_torch_accelerator @slow def test_inference_with_different_dtypes(self, dtype_str): dtype = { "float32": torch.float32, "float16": torch.float16, "bfloat16": torch.bfloat16, }[dtype_str] 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).to(dtype) model.eval() for key, tensor in inputs_dict.items(): if tensor.dtype == torch.float32: inputs_dict[key] = tensor.to(dtype) with torch.no_grad(): _ = model(**self._prepare_for_class(inputs_dict, model_class)) @parameterized.expand(["float32", "float16", "bfloat16"]) @require_torch_accelerator @slow def test_inference_equivalence_for_static_and_dynamic_anchors(self, dtype_str): dtype = { "float32": torch.float32, "float16": torch.float16, "bfloat16": torch.bfloat16, }[dtype_str] config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() h, w = inputs_dict["pixel_values"].shape[-2:] for key, tensor in inputs_dict.items(): if tensor.dtype == torch.float32: inputs_dict[key] = tensor.to(dtype) for model_class in self.all_model_classes: with tempfile.TemporaryDirectory() as tmpdirname: model_class(config).save_pretrained(tmpdirname) model_static = model_class.from_pretrained( tmpdirname, anchor_image_size=[h, w], device_map=torch_device, dtype=dtype ).eval() model_dynamic = model_class.from_pretrained( tmpdirname, anchor_image_size=None, device_map=torch_device, dtype=dtype ).eval() self.assertIsNotNone(model_static.config.anchor_image_size) self.assertIsNone(model_dynamic.config.anchor_image_size) with torch.no_grad(): outputs_static = model_static(**self._prepare_for_class(inputs_dict, model_class)) outputs_dynamic = model_dynamic(**self._prepare_for_class(inputs_dict, model_class)) torch.testing.assert_close( outputs_static.last_hidden_state, outputs_dynamic.last_hidden_state, rtol=5e-3, atol=5e-3 ) def prepare_img(): from transformers.image_utils import load_image url = url_to_local_path("http://images.cocodataset.org/val2017/000000039769.jpg") return load_image(url) @require_torch @require_vision @slow class Deimv2ModelIntegrationTest(unittest.TestCase): @cached_property def default_image_processor(self): return AutoImageProcessor.from_pretrained(CHECKPOINT, use_fast=False) def test_inference_object_detection_head(self): model = Deimv2ForObjectDetection.from_pretrained(CHECKPOINT).to(torch_device) image_processor = self.default_image_processor image = prepare_img() inputs = image_processor(images=image, return_tensors="pt").to(torch_device) with torch.no_grad(): outputs = model(**inputs) expected_shape_logits = torch.Size((1, 300, model.config.num_labels)) self.assertEqual(outputs.logits.shape, expected_shape_logits) expected_logits = torch.tensor( [[-4.0859, -6.9373, -5.4723], [-5.5887, -6.0078, -6.4360], [-6.1448, -6.8509, -6.8703]] ).to(torch_device) expected_boxes = torch.tensor( [[0.1886, 0.1662, 0.2875], [0.0690, 0.1814, 0.9368], [0.2510, 0.2141, 0.9115]] ).to(torch_device) torch.testing.assert_close(outputs.logits[0, :3, :3], expected_logits, atol=2e-4, rtol=2e-4) expected_shape_boxes = torch.Size((1, 300, 4)) self.assertEqual(outputs.pred_boxes.shape, expected_shape_boxes) torch.testing.assert_close(outputs.pred_boxes[0, :3, :3], expected_boxes, atol=2e-4, rtol=2e-4) results = image_processor.post_process_object_detection( outputs, threshold=0.0, target_sizes=[image.size[::-1]] )[0] expected_scores = torch.tensor([0.7606, 0.3165, 0.2726, 0.2488], device=torch_device) expected_labels = [65, 65, 15, 59] expected_slice_boxes = torch.tensor( [ [4.0781e01, 6.8216e01, 1.7560e02, 1.1085e02], [4.8195e01, 7.5405e01, 2.1123e02, 9.1451e01], [1.1296e01, 6.8089e01, 6.1285e02, 4.0393e02], [1.9821e01, -9.0347e01, 7.0787e02, 3.7968e02], ], device=torch_device, ) torch.testing.assert_close(results["scores"][:4], expected_scores, atol=1e-3, rtol=1e-4) self.assertSequenceEqual(results["labels"][:4].tolist(), expected_labels) torch.testing.assert_close(results["boxes"][:4], expected_slice_boxes[:4], atol=5e-3, rtol=5e-4) @require_torch @require_vision @slow class Deimv2LiteEncoderIntegrationTest(unittest.TestCase): @cached_property def default_image_processor(self): return AutoImageProcessor.from_pretrained(CHECKPOINT_LITE, use_fast=False) def test_inference_object_detection_head(self): model = Deimv2ForObjectDetection.from_pretrained(CHECKPOINT_LITE).to(torch_device) image_processor = self.default_image_processor image = prepare_img() inputs = image_processor(images=image, return_tensors="pt").to(torch_device) with torch.no_grad(): outputs = model(**inputs) expected_shape_logits = torch.Size((1, model.config.num_queries, model.config.num_labels)) self.assertEqual(outputs.logits.shape, expected_shape_logits) expected_logits = torch.tensor( [[-2.6151, -6.4701, -6.3505], [-3.8592, -6.2610, -7.2720], [-2.3801, -4.3216, -3.5101]] ).to(torch_device) expected_boxes = torch.tensor( [[0.7994, 0.2984, 0.3822], [0.5536, 0.5362, 0.0392], [0.3501, 0.4577, 0.7440]] ).to(torch_device) torch.testing.assert_close(outputs.logits[0, :3, :3], expected_logits, atol=2e-4, rtol=2e-4) expected_shape_boxes = torch.Size((1, model.config.num_queries, 4)) self.assertEqual(outputs.pred_boxes.shape, expected_shape_boxes) torch.testing.assert_close(outputs.pred_boxes[0, :3, :3], expected_boxes, atol=2e-4, rtol=2e-4) @require_torch @require_vision @slow class Deimv2DINOv3IntegrationTest(unittest.TestCase): @cached_property def default_image_processor(self): return AutoImageProcessor.from_pretrained(CHECKPOINT_DINOV3, use_fast=False) def test_inference_object_detection_head(self): model = Deimv2ForObjectDetection.from_pretrained(CHECKPOINT_DINOV3).to(torch_device) image_processor = self.default_image_processor image = prepare_img() inputs = image_processor(images=image, return_tensors="pt").to(torch_device) with torch.no_grad(): outputs = model(**inputs) expected_shape_logits = torch.Size((1, 300, model.config.num_labels)) self.assertEqual(outputs.logits.shape, expected_shape_logits) expected_logits = torch.tensor( [[-2.1404, -2.8207, -3.2710], [-2.3058, -2.7178, -3.2924], [-3.2780, -4.0269, -4.6266]] ).to(torch_device) expected_boxes = torch.tensor( [[0.5258, 0.7694, 0.7997], [0.3734, 0.1949, 0.7989], [0.5082, 0.5847, 0.8590]] ).to(torch_device) torch.testing.assert_close(outputs.logits[0, :3, :3], expected_logits, atol=2e-4, rtol=2e-4) expected_shape_boxes = torch.Size((1, 300, 4)) self.assertEqual(outputs.pred_boxes.shape, expected_shape_boxes) torch.testing.assert_close(outputs.pred_boxes[0, :3, :3], expected_boxes, atol=2e-4, rtol=2e-4)