From 1fc6817a30ea583a02bffb77215f213429d5b5bc Mon Sep 17 00:00:00 2001 From: NielsRogge <48327001+NielsRogge@users.noreply.github.com> Date: Tue, 29 Jun 2021 09:07:46 +0200 Subject: [PATCH] Rename detr targets to labels (#12280) * Rename target to labels in DetrFeatureExtractor * Update DetrFeatureExtractor tests accordingly * Improve docs of DetrFeatureExtractor * Improve docs * Make style --- .../models/detr/configuration_detr.py | 6 --- .../models/detr/feature_extraction_detr.py | 6 ++- src/transformers/models/detr/modeling_detr.py | 5 +-- tests/test_feature_extraction_detr.py | 37 +++++++++---------- 4 files changed, 24 insertions(+), 30 deletions(-) diff --git a/src/transformers/models/detr/configuration_detr.py b/src/transformers/models/detr/configuration_detr.py index 52625b1494..a8d9b4d6a2 100644 --- a/src/transformers/models/detr/configuration_detr.py +++ b/src/transformers/models/detr/configuration_detr.py @@ -64,11 +64,6 @@ class DetrConfig(PretrainedConfig): The dropout ratio for the attention probabilities. activation_dropout (:obj:`float`, `optional`, defaults to 0.0): The dropout ratio for activations inside the fully connected layer. - classifier_dropout (:obj:`float`, `optional`, defaults to 0.0): - The dropout ratio for classifier. - max_position_embeddings (:obj:`int`, `optional`, defaults to 1024): - The maximum sequence length that this model might ever be used with. Typically set this to something large - just in case (e.g., 512 or 1024 or 2048). init_std (:obj:`float`, `optional`, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. init_xavier_std (:obj:`float`, `optional`, defaults to 1): @@ -178,7 +173,6 @@ class DetrConfig(PretrainedConfig): self.init_xavier_std = init_xavier_std self.encoder_layerdrop = encoder_layerdrop self.decoder_layerdrop = decoder_layerdrop - self.classifier_dropout = classifier_dropout self.num_hidden_layers = encoder_layers self.scale_embedding = scale_embedding # scale factor will be sqrt(d_model) if True self.auxiliary_loss = auxiliary_loss diff --git a/src/transformers/models/detr/feature_extraction_detr.py b/src/transformers/models/detr/feature_extraction_detr.py index 94a848f340..238ac6a0d1 100644 --- a/src/transformers/models/detr/feature_extraction_detr.py +++ b/src/transformers/models/detr/feature_extraction_detr.py @@ -440,7 +440,8 @@ class DetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin): annotations. return_segmentation_masks (:obj:`Dict`, :obj:`List[Dict]`, `optional`, defaults to :obj:`False`): - Whether to also return instance segmentation masks in case :obj:`format = "coco_detection"`. + Whether to also include instance segmentation masks as part of the labels in case :obj:`format = + "coco_detection"`. masks_path (:obj:`pathlib.Path`, `optional`): Path to the directory containing the PNG files that store the class-agnostic image segmentations. Only @@ -465,6 +466,7 @@ class DetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin): - **pixel_values** -- Pixel values to be fed to a model. - **pixel_mask** -- Pixel mask to be fed to a model (when :obj:`pad_and_return_pixel_mask=True` or if `"pixel_mask"` is in :obj:`self.model_input_names`). + - **labels** -- Optional labels to be fed to a model (when :obj:`annotations` are provided) """ # Input type checking for clearer error @@ -613,7 +615,7 @@ class DetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin): if not is_torch_available(): raise ImportError("Unable to convert output to PyTorch tensors format, PyTorch is not installed.") - encoded_inputs["target"] = [ + encoded_inputs["labels"] = [ {k: torch.from_numpy(v) for k, v in target.items()} for target in annotations ] diff --git a/src/transformers/models/detr/modeling_detr.py b/src/transformers/models/detr/modeling_detr.py index 0e4721e2b3..9043e8cc0a 100644 --- a/src/transformers/models/detr/modeling_detr.py +++ b/src/transformers/models/detr/modeling_detr.py @@ -828,8 +828,8 @@ DETR_INPUTS_DOCSTRING = r""" pixel_values (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, num_channels, height, width)`): Pixel values. Padding will be ignored by default should you provide it. - Pixel values can be obtained using :class:`~transformers.DetrTokenizer`. See - :meth:`transformers.DetrTokenizer.__call__` for details. + Pixel values can be obtained using :class:`~transformers.DetrFeatureExtractor`. See + :meth:`transformers.DetrFeatureExtractor.__call__` for details. pixel_mask (:obj:`torch.LongTensor` of shape :obj:`(batch_size, height, width)`, `optional`): Mask to avoid performing attention on padding pixel values. Mask values selected in ``[0, 1]``: @@ -990,7 +990,6 @@ class DetrDecoder(DetrPreTrainedModel): super().__init__(config) self.dropout = config.dropout self.layerdrop = config.decoder_layerdrop - self.max_target_positions = config.max_position_embeddings self.layers = nn.ModuleList([DetrDecoderLayer(config) for _ in range(config.decoder_layers)]) # in DETR, the decoder uses layernorm after the last decoder layer output diff --git a/tests/test_feature_extraction_detr.py b/tests/test_feature_extraction_detr.py index 8f36ad418f..4207d88fe0 100644 --- a/tests/test_feature_extraction_detr.py +++ b/tests/test_feature_extraction_detr.py @@ -253,8 +253,7 @@ class DetrFeatureExtractionTest(FeatureExtractionSavingTestMixin, unittest.TestC target = {"image_id": 39769, "annotations": target} # encode them - # TODO replace by facebook/detr-resnet-50 - feature_extractor = DetrFeatureExtractor.from_pretrained("nielsr/detr-resnet-50") + feature_extractor = DetrFeatureExtractor.from_pretrained("facebook/detr-resnet-50") encoding = feature_extractor(images=image, annotations=target, return_tensors="pt") # verify pixel values @@ -266,27 +265,27 @@ class DetrFeatureExtractionTest(FeatureExtractionSavingTestMixin, unittest.TestC # verify area expected_area = torch.tensor([5887.9600, 11250.2061, 489353.8438, 837122.7500, 147967.5156, 165732.3438]) - assert torch.allclose(encoding["target"][0]["area"], expected_area) + assert torch.allclose(encoding["labels"][0]["area"], expected_area) # verify boxes expected_boxes_shape = torch.Size([6, 4]) - self.assertEqual(encoding["target"][0]["boxes"].shape, expected_boxes_shape) + self.assertEqual(encoding["labels"][0]["boxes"].shape, expected_boxes_shape) expected_boxes_slice = torch.tensor([0.5503, 0.2765, 0.0604, 0.2215]) - assert torch.allclose(encoding["target"][0]["boxes"][0], expected_boxes_slice, atol=1e-3) + assert torch.allclose(encoding["labels"][0]["boxes"][0], expected_boxes_slice, atol=1e-3) # verify image_id expected_image_id = torch.tensor([39769]) - assert torch.allclose(encoding["target"][0]["image_id"], expected_image_id) + assert torch.allclose(encoding["labels"][0]["image_id"], expected_image_id) # verify is_crowd expected_is_crowd = torch.tensor([0, 0, 0, 0, 0, 0]) - assert torch.allclose(encoding["target"][0]["iscrowd"], expected_is_crowd) + assert torch.allclose(encoding["labels"][0]["iscrowd"], expected_is_crowd) # verify class_labels expected_class_labels = torch.tensor([75, 75, 63, 65, 17, 17]) - assert torch.allclose(encoding["target"][0]["class_labels"], expected_class_labels) + assert torch.allclose(encoding["labels"][0]["class_labels"], expected_class_labels) # verify orig_size expected_orig_size = torch.tensor([480, 640]) - assert torch.allclose(encoding["target"][0]["orig_size"], expected_orig_size) + assert torch.allclose(encoding["labels"][0]["orig_size"], expected_orig_size) # verify size expected_size = torch.tensor([800, 1066]) - assert torch.allclose(encoding["target"][0]["size"], expected_size) + assert torch.allclose(encoding["labels"][0]["size"], expected_size) @slow def test_call_pytorch_with_coco_panoptic_annotations(self): @@ -313,27 +312,27 @@ class DetrFeatureExtractionTest(FeatureExtractionSavingTestMixin, unittest.TestC # verify area expected_area = torch.tensor([147979.6875, 165527.0469, 484638.5938, 11292.9375, 5879.6562, 7634.1147]) - assert torch.allclose(encoding["target"][0]["area"], expected_area) + assert torch.allclose(encoding["labels"][0]["area"], expected_area) # verify boxes expected_boxes_shape = torch.Size([6, 4]) - self.assertEqual(encoding["target"][0]["boxes"].shape, expected_boxes_shape) + self.assertEqual(encoding["labels"][0]["boxes"].shape, expected_boxes_shape) expected_boxes_slice = torch.tensor([0.2625, 0.5437, 0.4688, 0.8625]) - assert torch.allclose(encoding["target"][0]["boxes"][0], expected_boxes_slice, atol=1e-3) + assert torch.allclose(encoding["labels"][0]["boxes"][0], expected_boxes_slice, atol=1e-3) # verify image_id expected_image_id = torch.tensor([39769]) - assert torch.allclose(encoding["target"][0]["image_id"], expected_image_id) + assert torch.allclose(encoding["labels"][0]["image_id"], expected_image_id) # verify is_crowd expected_is_crowd = torch.tensor([0, 0, 0, 0, 0, 0]) - assert torch.allclose(encoding["target"][0]["iscrowd"], expected_is_crowd) + assert torch.allclose(encoding["labels"][0]["iscrowd"], expected_is_crowd) # verify class_labels expected_class_labels = torch.tensor([17, 17, 63, 75, 75, 93]) - assert torch.allclose(encoding["target"][0]["class_labels"], expected_class_labels) + assert torch.allclose(encoding["labels"][0]["class_labels"], expected_class_labels) # verify masks expected_masks_sum = 822338 - self.assertEqual(encoding["target"][0]["masks"].sum().item(), expected_masks_sum) + self.assertEqual(encoding["labels"][0]["masks"].sum().item(), expected_masks_sum) # verify orig_size expected_orig_size = torch.tensor([480, 640]) - assert torch.allclose(encoding["target"][0]["orig_size"], expected_orig_size) + assert torch.allclose(encoding["labels"][0]["orig_size"], expected_orig_size) # verify size expected_size = torch.tensor([800, 1066]) - assert torch.allclose(encoding["target"][0]["size"], expected_size) + assert torch.allclose(encoding["labels"][0]["size"], expected_size)