DETR

DETR

Overview

The DETR model was proposed in End-to-End Object Detection with Transformers by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov and Sergey Zagoruyko. DETR consists of a convolutional backbone followed by an encoder-decoder Transformer which can be trained end-to-end for object detection. It greatly simplifies a lot of the complexity of models like Faster-R-CNN and Mask-R-CNN, which use things like region proposals, non-maximum suppression procedure and anchor generation. Moreover, DETR can also be naturally extended to perform panoptic segmentation, by simply adding a mask head on top of the decoder outputs.

The abstract from the paper is the following:

We present a new method that views object detection as a direct set prediction problem. Our approach streamlines the detection pipeline, effectively removing the need for many hand-designed components like a non-maximum suppression procedure or anchor generation that explicitly encode our prior knowledge about the task. The main ingredients of the new framework, called DEtection TRansformer or DETR, are a set-based global loss that forces unique predictions via bipartite matching, and a transformer encoder-decoder architecture. Given a fixed small set of learned object queries, DETR reasons about the relations of the objects and the global image context to directly output the final set of predictions in parallel. The new model is conceptually simple and does not require a specialized library, unlike many other modern detectors. DETR demonstrates accuracy and run-time performance on par with the well-established and highly-optimized Faster RCNN baseline on the challenging COCO object detection dataset. Moreover, DETR can be easily generalized to produce panoptic segmentation in a unified manner. We show that it significantly outperforms competitive baselines.

This model was contributed by nielsr. The original code can be found here.

Here’s a TLDR explaining how DetrForObjectDetection works:

First, an image is sent through a pre-trained convolutional backbone (in the paper, the authors use ResNet-50/ResNet-101). Let’s assume we also add a batch dimension. This means that the input to the backbone is a tensor of shape (batch_size, 3, height, width), assuming the image has 3 color channels (RGB). The CNN backbone outputs a new lower-resolution feature map, typically of shape (batch_size, 2048, height/32, width/32). This is then projected to match the hidden dimension of the Transformer of DETR, which is 256 by default, using a nn.Conv2D layer. So now, we have a tensor of shape (batch_size, 256, height/32, width/32). Next, the feature map is flattened and transposed to obtain a tensor of shape (batch_size, seq_len, d_model) = (batch_size, width/32*height/32, 256). So a difference with NLP models is that the sequence length is actually longer than usual, but with a smaller d_model (which in NLP is typically 768 or higher).

Next, this is sent through the encoder, outputting encoder_hidden_states of the same shape (you can consider these as image features). Next, so-called object queries are sent through the decoder. This is a tensor of shape (batch_size, num_queries, d_model), with num_queries typically set to 100 and initialized with zeros. These input embeddings are learnt positional encodings that the authors refer to as object queries, and similarly to the encoder, they are added to the input of each attention layer. Each object query will look for a particular object in the image. The decoder updates these embeddings through multiple self-attention and encoder-decoder attention layers to output decoder_hidden_states of the same shape: (batch_size, num_queries, d_model). Next, two heads are added on top for object detection: a linear layer for classifying each object query into one of the objects or “no object”, and a MLP to predict bounding boxes for each query.

The model is trained using a bipartite matching loss: so what we actually do is compare the predicted classes + bounding boxes of each of the N = 100 object queries to the ground truth annotations, padded up to the same length N (so if an image only contains 4 objects, 96 annotations will just have a “no object” as class and “no bounding box” as bounding box). The Hungarian matching algorithm is used to find an optimal one-to-one mapping of each of the N queries to each of the N annotations. Next, standard cross-entropy (for the classes) and a linear combination of the L1 and generalized IoU loss (for the bounding boxes) are used to optimize the parameters of the model.

DETR can be naturally extended to perform panoptic segmentation (which unifies semantic segmentation and instance segmentation). DetrForSegmentation adds a segmentation mask head on top of DetrForObjectDetection. The mask head can be trained either jointly, or in a two steps process, where one first trains a DetrForObjectDetection model to detect bounding boxes around both “things” (instances) and “stuff” (background things like trees, roads, sky), then freeze all the weights and train only the mask head for 25 epochs. Experimentally, these two approaches give similar results. Note that predicting boxes is required for the training to be possible, since the Hungarian matching is computed using distances between boxes.

Tips:

  • DETR uses so-called object queries to detect objects in an image. The number of queries determines the maximum number of objects that can be detected in a single image, and is set to 100 by default (see parameter num_queries of DetrConfig). Note that it’s good to have some slack (in COCO, the authors used 100, while the maximum number of objects in a COCO image is ~70).

  • The decoder of DETR updates the query embeddings in parallel. This is different from language models like GPT-2, which use autoregressive decoding instead of parallel. Hence, no causal attention mask is used.

  • DETR adds position embeddings to the hidden states at each self-attention and cross-attention layer before projecting to queries and keys. For the position embeddings of the image, one can choose between fixed sinusoidal or learned absolute position embeddings. By default, the parameter position_embedding_type of DetrConfig is set to "sine".

  • During training, the authors of DETR did find it helpful to use auxiliary losses in the decoder, especially to help the model output the correct number of objects of each class. If you set the parameter auxiliary_loss of DetrConfig to True, then prediction feedforward neural networks and Hungarian losses are added after each decoder layer (with the FFNs sharing parameters).

  • If you want to train the model in a distributed environment across multiple nodes, then one should update the num_boxes variable in the DetrLoss class of modeling_detr.py. When training on multiple nodes, this should be set to the average number of target boxes across all nodes, as can be seen in the original implementation here.

  • DetrForObjectDetection and DetrForSegmentation can be initialized with any convolutional backbone available in the timm library. Initializing with a MobileNet backbone for example can be done by setting the backbone attribute of DetrConfig to "tf_mobilenetv3_small_075", and then initializing the model with that config.

  • DETR resizes the input images such that the shortest side is at least a certain amount of pixels while the longest is at most 1333 pixels. At training time, scale augmentation is used such that the shortest side is randomly set to at least 480 and at most 800 pixels. At inference time, the shortest side is set to 800. One can use DetrImageProcessor to prepare images (and optional annotations in COCO format) for the model. Due to this resizing, images in a batch can have different sizes. DETR solves this by padding images up to the largest size in a batch, and by creating a pixel mask that indicates which pixels are real/which are padding. Alternatively, one can also define a custom collate_fn in order to batch images together, using ~transformers.DetrImageProcessor.pad_and_create_pixel_mask.

  • The size of the images will determine the amount of memory being used, and will thus determine the batch_size. It is advised to use a batch size of 2 per GPU. See this Github thread for more info.

There are three ways to instantiate a DETR model (depending on what you prefer):

Option 1: Instantiate DETR with pre-trained weights for entire model

Copied

>>> from transformers import DetrForObjectDetection

>>> model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50")

Option 2: Instantiate DETR with randomly initialized weights for Transformer, but pre-trained weights for backbone

Copied

>>> from transformers import DetrConfig, DetrForObjectDetection

>>> config = DetrConfig()
>>> model = DetrForObjectDetection(config)

Option 3: Instantiate DETR with randomly initialized weights for backbone + Transformer

Copied

>>> config = DetrConfig(use_pretrained_backbone=False)
>>> model = DetrForObjectDetection(config)

As a summary, consider the following table:

TaskObject detectionInstance segmentationPanoptic segmentation

Description

Predicting bounding boxes and class labels around objects in an image

Predicting masks around objects (i.e. instances) in an image

Predicting masks around both objects (i.e. instances) as well as “stuff” (i.e. background things like trees and roads) in an image

Model

Example dataset

COCO detection

COCO detection, COCO panoptic

COCO panoptic

Format of annotations to provide to DetrImageProcessor

{‘image_id’: int, ‘annotations’: List[Dict]} each Dict being a COCO object annotation

{‘image_id’: int, ‘annotations’: List[Dict]} (in case of COCO detection) or {‘file_name’: str, ‘image_id’: int, ‘segments_info’: List[Dict]} (in case of COCO panoptic)

{‘file_name’: str, ‘image_id’: int, ‘segments_info’: List[Dict]} and masks_path (path to directory containing PNG files of the masks)

Postprocessing (i.e. converting the output of the model to COCO API)

post_process()

post_process_segmentation()

post_process_segmentation(), post_process_panoptic()

evaluators

CocoEvaluator with iou_types="bbox"

CocoEvaluator with iou_types="bbox" or "segm"

CocoEvaluator with iou_tupes="bbox" or "segm", PanopticEvaluator

In short, one should prepare the data either in COCO detection or COCO panoptic format, then use DetrImageProcessor to create pixel_values, pixel_mask and optional labels, which can then be used to train (or fine-tune) a model. For evaluation, one should first convert the outputs of the model using one of the postprocessing methods of DetrImageProcessor. These can be be provided to either CocoEvaluator or PanopticEvaluator, which allow you to calculate metrics like mean Average Precision (mAP) and Panoptic Quality (PQ). The latter objects are implemented in the original repository. See the example notebooks for more info regarding evaluation.

Resources

A list of official BOINC AI and community (indicated by 🌎) resources to help you get started with DETR.

Object Detection

If you’re interested in submitting a resource to be included here, please feel free to open a Pull Request and we’ll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.

DETR specific outputs

class transformers.models.detr.modeling_detr.DetrModelOutput

<source>

( last_hidden_state: FloatTensor = Nonepast_key_values: typing.Optional[typing.Tuple[typing.Tuple[torch.FloatTensor]]] = Nonedecoder_hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = Nonedecoder_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = Nonecross_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = Noneencoder_last_hidden_state: typing.Optional[torch.FloatTensor] = Noneencoder_hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = Noneencoder_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = Noneintermediate_hidden_states: typing.Optional[torch.FloatTensor] = None )

Parameters

  • last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the decoder of the model.

  • decoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.

  • decoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads.

  • cross_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the decoder’s cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads.

  • encoder_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the encoder of the model.

  • encoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.

  • encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads.

  • intermediate_hidden_states (torch.FloatTensor of shape (config.decoder_layers, batch_size, sequence_length, hidden_size), optional, returned when config.auxiliary_loss=True) — Intermediate decoder activations, i.e. the output of each decoder layer, each of them gone through a layernorm.

Base class for outputs of the DETR encoder-decoder model. This class adds one attribute to Seq2SeqModelOutput, namely an optional stack of intermediate decoder activations, i.e. the output of each decoder layer, each of them gone through a layernorm. This is useful when training the model with auxiliary decoding losses.

class transformers.models.detr.modeling_detr.DetrObjectDetectionOutput

<source>

( loss: typing.Optional[torch.FloatTensor] = Noneloss_dict: typing.Optional[typing.Dict] = Nonelogits: FloatTensor = Nonepred_boxes: FloatTensor = Noneauxiliary_outputs: typing.Optional[typing.List[typing.Dict]] = Nonelast_hidden_state: typing.Optional[torch.FloatTensor] = Nonedecoder_hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = Nonedecoder_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = Nonecross_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = Noneencoder_last_hidden_state: typing.Optional[torch.FloatTensor] = Noneencoder_hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = Noneencoder_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None )

Parameters

  • loss (torch.FloatTensor of shape (1,), optional, returned when labels are provided)) — Total loss as a linear combination of a negative log-likehood (cross-entropy) for class prediction and a bounding box loss. The latter is defined as a linear combination of the L1 loss and the generalized scale-invariant IoU loss.

  • loss_dict (Dict, optional) — A dictionary containing the individual losses. Useful for logging.

  • logits (torch.FloatTensor of shape (batch_size, num_queries, num_classes + 1)) — Classification logits (including no-object) for all queries.

  • pred_boxes (torch.FloatTensor of shape (batch_size, num_queries, 4)) — Normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height). These values are normalized in [0, 1], relative to the size of each individual image in the batch (disregarding possible padding). You can use post_process_object_detection() to retrieve the unnormalized bounding boxes.

  • auxiliary_outputs (list[Dict], optional) — Optional, only returned when auxilary losses are activated (i.e. config.auxiliary_loss is set to True) and labels are provided. It is a list of dictionaries containing the two above keys (logits and pred_boxes) for each decoder layer.

  • last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the decoder of the model.

  • decoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.

  • decoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads.

  • cross_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the decoder’s cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads.

  • encoder_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the encoder of the model.

  • encoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.

  • encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads.

Output type of DetrForObjectDetection.

class transformers.models.detr.modeling_detr.DetrSegmentationOutput

<source>

( loss: typing.Optional[torch.FloatTensor] = Noneloss_dict: typing.Optional[typing.Dict] = Nonelogits: FloatTensor = Nonepred_boxes: FloatTensor = Nonepred_masks: FloatTensor = Noneauxiliary_outputs: typing.Optional[typing.List[typing.Dict]] = Nonelast_hidden_state: typing.Optional[torch.FloatTensor] = Nonedecoder_hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = Nonedecoder_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = Nonecross_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = Noneencoder_last_hidden_state: typing.Optional[torch.FloatTensor] = Noneencoder_hidden_states: typing.Optional[typing.Tuple[torch.FloatTensor]] = Noneencoder_attentions: typing.Optional[typing.Tuple[torch.FloatTensor]] = None )

Parameters

  • loss (torch.FloatTensor of shape (1,), optional, returned when labels are provided)) — Total loss as a linear combination of a negative log-likehood (cross-entropy) for class prediction and a bounding box loss. The latter is defined as a linear combination of the L1 loss and the generalized scale-invariant IoU loss.

  • loss_dict (Dict, optional) — A dictionary containing the individual losses. Useful for logging.

  • logits (torch.FloatTensor of shape (batch_size, num_queries, num_classes + 1)) — Classification logits (including no-object) for all queries.

  • pred_boxes (torch.FloatTensor of shape (batch_size, num_queries, 4)) — Normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height). These values are normalized in [0, 1], relative to the size of each individual image in the batch (disregarding possible padding). You can use post_process_object_detection() to retrieve the unnormalized bounding boxes.

  • pred_masks (torch.FloatTensor of shape (batch_size, num_queries, height/4, width/4)) — Segmentation masks logits for all queries. See also post_process_semantic_segmentation() or post_process_instance_segmentation() post_process_panoptic_segmentation() to evaluate semantic, instance and panoptic segmentation masks respectively.

  • auxiliary_outputs (list[Dict], optional) — Optional, only returned when auxiliary losses are activated (i.e. config.auxiliary_loss is set to True) and labels are provided. It is a list of dictionaries containing the two above keys (logits and pred_boxes) for each decoder layer.

  • last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the decoder of the model.

  • decoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.

  • decoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads.

  • cross_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the decoder’s cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads.

  • encoder_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the encoder of the model.

  • encoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.

  • encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads.

Output type of DetrForSegmentation.

DetrConfig

class transformers.DetrConfig

<source>

( use_timm_backbone = Truebackbone_config = Nonenum_channels = 3num_queries = 100encoder_layers = 6encoder_ffn_dim = 2048encoder_attention_heads = 8decoder_layers = 6decoder_ffn_dim = 2048decoder_attention_heads = 8encoder_layerdrop = 0.0decoder_layerdrop = 0.0is_encoder_decoder = Trueactivation_function = 'relu'd_model = 256dropout = 0.1attention_dropout = 0.0activation_dropout = 0.0init_std = 0.02init_xavier_std = 1.0auxiliary_loss = Falseposition_embedding_type = 'sine'backbone = 'resnet50'use_pretrained_backbone = Truedilation = Falseclass_cost = 1bbox_cost = 5giou_cost = 2mask_loss_coefficient = 1dice_loss_coefficient = 1bbox_loss_coefficient = 5giou_loss_coefficient = 2eos_coefficient = 0.1**kwargs )

Parameters

  • use_timm_backbone (bool, optional, defaults to True) — Whether or not to use the timm library for the backbone. If set to False, will use the AutoBackbone API.

  • backbone_config (PretrainedConfig or dict, optional) — The configuration of the backbone model. Only used in case use_timm_backbone is set to False in which case it will default to ResNetConfig().

  • num_channels (int, optional, defaults to 3) — The number of input channels.

  • num_queries (int, optional, defaults to 100) — Number of object queries, i.e. detection slots. This is the maximal number of objects DetrModel can detect in a single image. For COCO, we recommend 100 queries.

  • d_model (int, optional, defaults to 256) — Dimension of the layers.

  • encoder_layers (int, optional, defaults to 6) — Number of encoder layers.

  • decoder_layers (int, optional, defaults to 6) — Number of decoder layers.

  • encoder_attention_heads (int, optional, defaults to 8) — Number of attention heads for each attention layer in the Transformer encoder.

  • decoder_attention_heads (int, optional, defaults to 8) — Number of attention heads for each attention layer in the Transformer decoder.

  • decoder_ffn_dim (int, optional, defaults to 2048) — Dimension of the “intermediate” (often named feed-forward) layer in decoder.

  • encoder_ffn_dim (int, optional, defaults to 2048) — Dimension of the “intermediate” (often named feed-forward) layer in decoder.

  • activation_function (str or function, optional, defaults to "relu") — The non-linear activation function (function or string) in the encoder and pooler. If string, "gelu", "relu", "silu" and "gelu_new" are supported.

  • dropout (float, optional, defaults to 0.1) — The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.

  • attention_dropout (float, optional, defaults to 0.0) — The dropout ratio for the attention probabilities.

  • activation_dropout (float, optional, defaults to 0.0) — The dropout ratio for activations inside the fully connected layer.

  • init_std (float, optional, defaults to 0.02) — The standard deviation of the truncated_normal_initializer for initializing all weight matrices.

  • init_xavier_std (float, optional, defaults to 1) — The scaling factor used for the Xavier initialization gain in the HM Attention map module.

  • encoder_layerdrop (float, optional, defaults to 0.0) — The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556) for more details.

  • decoder_layerdrop (float, optional, defaults to 0.0) — The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556) for more details.

  • auxiliary_loss (bool, optional, defaults to False) — Whether auxiliary decoding losses (loss at each decoder layer) are to be used.

  • position_embedding_type (str, optional, defaults to "sine") — Type of position embeddings to be used on top of the image features. One of "sine" or "learned".

  • backbone (str, optional, defaults to "resnet50") — Name of convolutional backbone to use in case use_timm_backbone = True. Supports any convolutional backbone from the timm package. For a list of all available models, see this page.

  • use_pretrained_backbone (bool, optional, defaults to True) — Whether to use pretrained weights for the backbone. Only supported when use_timm_backbone = True.

  • dilation (bool, optional, defaults to False) — Whether to replace stride with dilation in the last convolutional block (DC5). Only supported when use_timm_backbone = True.

  • class_cost (float, optional, defaults to 1) — Relative weight of the classification error in the Hungarian matching cost.

  • bbox_cost (float, optional, defaults to 5) — Relative weight of the L1 error of the bounding box coordinates in the Hungarian matching cost.

  • giou_cost (float, optional, defaults to 2) — Relative weight of the generalized IoU loss of the bounding box in the Hungarian matching cost.

  • mask_loss_coefficient (float, optional, defaults to 1) — Relative weight of the Focal loss in the panoptic segmentation loss.

  • dice_loss_coefficient (float, optional, defaults to 1) — Relative weight of the DICE/F-1 loss in the panoptic segmentation loss.

  • bbox_loss_coefficient (float, optional, defaults to 5) — Relative weight of the L1 bounding box loss in the object detection loss.

  • giou_loss_coefficient (float, optional, defaults to 2) — Relative weight of the generalized IoU loss in the object detection loss.

  • eos_coefficient (float, optional, defaults to 0.1) — Relative classification weight of the ‘no-object’ class in the object detection loss.

This is the configuration class to store the configuration of a DetrModel. It is used to instantiate a DETR model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the DETR facebook/detr-resnet-50 architecture.

Configuration objects inherit from PretrainedConfig and can be used to control the model outputs. Read the documentation from PretrainedConfig for more information.

Examples:

Copied

>>> from transformers import DetrConfig, DetrModel

>>> # Initializing a DETR facebook/detr-resnet-50 style configuration
>>> configuration = DetrConfig()

>>> # Initializing a model (with random weights) from the facebook/detr-resnet-50 style configuration
>>> model = DetrModel(configuration)

>>> # Accessing the model configuration
>>> configuration = model.config

from_backbone_config

<source>

( backbone_config: PretrainedConfig**kwargs ) → DetrConfig

Parameters

Returns

DetrConfig

An instance of a configuration object

Instantiate a DetrConfig (or a derived class) from a pre-trained backbone model configuration.

DetrImageProcessor

class transformers.DetrImageProcessor

<source>

( format: typing.Union[str, transformers.models.detr.image_processing_detr.AnnotionFormat] = <AnnotionFormat.COCO_DETECTION: 'coco_detection'>do_resize: bool = Truesize: typing.Dict[str, int] = Noneresample: Resampling = <Resampling.BILINEAR: 2>do_rescale: bool = Truerescale_factor: typing.Union[int, float] = 0.00392156862745098do_normalize: bool = Trueimage_mean: typing.Union[float, typing.List[float]] = Noneimage_std: typing.Union[float, typing.List[float]] = Nonedo_pad: bool = True**kwargs )

Parameters

  • format (str, optional, defaults to "coco_detection") — Data format of the annotations. One of “coco_detection” or “coco_panoptic”.

  • do_resize (bool, optional, defaults to True) — Controls whether to resize the image’s (height, width) dimensions to the specified size. Can be overridden by the do_resize parameter in the preprocess method.

  • size (Dict[str, int] optional, defaults to {"shortest_edge" -- 800, "longest_edge": 1333}): Size of the image’s (height, width) dimensions after resizing. Can be overridden by the size parameter in the preprocess method.

  • resample (PILImageResampling, optional, defaults to PILImageResampling.BILINEAR) — Resampling filter to use if resizing the image.

  • do_rescale (bool, optional, defaults to True) — Controls whether to rescale the image by the specified scale rescale_factor. Can be overridden by the do_rescale parameter in the preprocess method.

  • rescale_factor (int or float, optional, defaults to 1/255) — Scale factor to use if rescaling the image. Can be overridden by the rescale_factor parameter in the preprocess method. do_normalize — Controls whether to normalize the image. Can be overridden by the do_normalize parameter in the preprocess method.

  • image_mean (float or List[float], optional, defaults to IMAGENET_DEFAULT_MEAN) — Mean values to use when normalizing the image. Can be a single value or a list of values, one for each channel. Can be overridden by the image_mean parameter in the preprocess method.

  • image_std (float or List[float], optional, defaults to IMAGENET_DEFAULT_STD) — Standard deviation values to use when normalizing the image. Can be a single value or a list of values, one for each channel. Can be overridden by the image_std parameter in the preprocess method.

  • do_pad (bool, optional, defaults to True) — Controls whether to pad the image to the largest image in a batch and create a pixel mask. Can be overridden by the do_pad parameter in the preprocess method.

Constructs a Detr image processor.

preprocess

<source>

( images: typing.Union[ForwardRef('PIL.Image.Image'), numpy.ndarray, ForwardRef('torch.Tensor'), typing.List[ForwardRef('PIL.Image.Image')], typing.List[numpy.ndarray], typing.List[ForwardRef('torch.Tensor')]]annotations: typing.Union[typing.Dict[str, typing.Union[int, str, typing.List[typing.Dict]]], typing.List[typing.Dict[str, typing.Union[int, str, typing.List[typing.Dict]]]], NoneType] = Nonereturn_segmentation_masks: bool = Nonemasks_path: typing.Union[str, pathlib.Path, NoneType] = Nonedo_resize: typing.Optional[bool] = Nonesize: typing.Union[typing.Dict[str, int], NoneType] = Noneresample = Nonedo_rescale: typing.Optional[bool] = Nonerescale_factor: typing.Union[int, float, NoneType] = Nonedo_normalize: typing.Optional[bool] = Noneimage_mean: typing.Union[float, typing.List[float], NoneType] = Noneimage_std: typing.Union[float, typing.List[float], NoneType] = Nonedo_pad: typing.Optional[bool] = Noneformat: typing.Union[str, transformers.models.detr.image_processing_detr.AnnotionFormat, NoneType] = Nonereturn_tensors: typing.Union[str, transformers.utils.generic.TensorType, NoneType] = Nonedata_format: typing.Union[str, transformers.image_utils.ChannelDimension] = <ChannelDimension.FIRST: 'channels_first'>input_data_format: typing.Union[str, transformers.image_utils.ChannelDimension, NoneType] = None**kwargs )

Parameters

  • images (ImageInput) — Image or batch of images to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If passing in images with pixel values between 0 and 1, set do_rescale=False.

  • annotations (AnnotationType or List[AnnotationType], optional) — List of annotations associated with the image or batch of images. If annotation is for object detection, the annotations should be a dictionary with the following keys:

    • “image_id” (int): The image id.

    • “annotations” (List[Dict]): List of annotations for an image. Each annotation should be a dictionary. An image can have no annotations, in which case the list should be empty. If annotation is for segmentation, the annotations should be a dictionary with the following keys:

    • “image_id” (int): The image id.

    • “segments_info” (List[Dict]): List of segments for an image. Each segment should be a dictionary. An image can have no segments, in which case the list should be empty.

    • “file_name” (str): The file name of the image.

  • return_segmentation_masks (bool, optional, defaults to self.return_segmentation_masks) — Whether to return segmentation masks.

  • masks_path (str or pathlib.Path, optional) — Path to the directory containing the segmentation masks.

  • do_resize (bool, optional, defaults to self.do_resize) — Whether to resize the image.

  • size (Dict[str, int], optional, defaults to self.size) — Size of the image after resizing.

  • resample (PILImageResampling, optional, defaults to self.resample) — Resampling filter to use when resizing the image.

  • do_rescale (bool, optional, defaults to self.do_rescale) — Whether to rescale the image.

  • rescale_factor (float, optional, defaults to self.rescale_factor) — Rescale factor to use when rescaling the image.

  • do_normalize (bool, optional, defaults to self.do_normalize) — Whether to normalize the image.

  • image_mean (float or List[float], optional, defaults to self.image_mean) — Mean to use when normalizing the image.

  • image_std (float or List[float], optional, defaults to self.image_std) — Standard deviation to use when normalizing the image.

  • do_pad (bool, optional, defaults to self.do_pad) — Whether to pad the image.

  • format (str or AnnotionFormat, optional, defaults to self.format) — Format of the annotations.

  • return_tensors (str or TensorType, optional, defaults to self.return_tensors) — Type of tensors to return. If None, will return the list of images.

  • data_format (ChannelDimension or str, optional, defaults to ChannelDimension.FIRST) — The channel dimension format for the output image. Can be one of:

    • "channels_first" or ChannelDimension.FIRST: image in (num_channels, height, width) format.

    • "channels_last" or ChannelDimension.LAST: image in (height, width, num_channels) format.

    • Unset: Use the channel dimension format of the input image.

  • input_data_format (ChannelDimension or str, optional) — The channel dimension format for the input image. If unset, the channel dimension format is inferred from the input image. Can be one of:

    • "channels_first" or ChannelDimension.FIRST: image in (num_channels, height, width) format.

    • "channels_last" or ChannelDimension.LAST: image in (height, width, num_channels) format.

    • "none" or ChannelDimension.NONE: image in (height, width) format.

Preprocess an image or a batch of images so that it can be used by the model.

post_process_object_detection

<source>

( outputsthreshold: float = 0.5target_sizes: typing.Union[transformers.utils.generic.TensorType, typing.List[typing.Tuple]] = None ) → List[Dict]

Parameters

  • outputs (DetrObjectDetectionOutput) — Raw outputs of the model.

  • threshold (float, optional) — Score threshold to keep object detection predictions.

  • target_sizes (torch.Tensor or List[Tuple[int, int]], optional) — Tensor of shape (batch_size, 2) or list of tuples (Tuple[int, int]) containing the target size (height, width) of each image in the batch. If unset, predictions will not be resized.

Returns

List[Dict]

A list of dictionaries, each dictionary containing the scores, labels and boxes for an image in the batch as predicted by the model.

Converts the raw output of DetrForObjectDetection into final bounding boxes in (top_left_x, top_left_y, bottom_right_x, bottom_right_y) format. Only supports PyTorch.

post_process_semantic_segmentation

<source>

( outputstarget_sizes: typing.List[typing.Tuple[int, int]] = None ) → List[torch.Tensor]

Parameters

  • outputs (DetrForSegmentation) — Raw outputs of the model.

  • target_sizes (List[Tuple[int, int]], optional) — A list of tuples (Tuple[int, int]) containing the target size (height, width) of each image in the batch. If unset, predictions will not be resized.

Returns

List[torch.Tensor]

A list of length batch_size, where each item is a semantic segmentation map of shape (height, width) corresponding to the target_sizes entry (if target_sizes is specified). Each entry of each torch.Tensor correspond to a semantic class id.

Converts the output of DetrForSegmentation into semantic segmentation maps. Only supports PyTorch.

post_process_instance_segmentation

<source>

( outputsthreshold: float = 0.5mask_threshold: float = 0.5overlap_mask_area_threshold: float = 0.8target_sizes: typing.Union[typing.List[typing.Tuple[int, int]], NoneType] = Nonereturn_coco_annotation: typing.Optional[bool] = False ) → List[Dict]

Parameters

  • outputs (DetrForSegmentation) — Raw outputs of the model.

  • threshold (float, optional, defaults to 0.5) — The probability score threshold to keep predicted instance masks.

  • mask_threshold (float, optional, defaults to 0.5) — Threshold to use when turning the predicted masks into binary values.

  • overlap_mask_area_threshold (float, optional, defaults to 0.8) — The overlap mask area threshold to merge or discard small disconnected parts within each binary instance mask.

  • target_sizes (List[Tuple], optional) — List of length (batch_size), where each list item (Tuple[int, int]]) corresponds to the requested final size (height, width) of each prediction. If unset, predictions will not be resized.

  • return_coco_annotation (bool, optional) — Defaults to False. If set to True, segmentation maps are returned in COCO run-length encoding (RLE) format.

Returns

List[Dict]

A list of dictionaries, one per image, each dictionary containing two keys:

  • segmentation — A tensor of shape (height, width) where each pixel represents a segment_id or List[List] run-length encoding (RLE) of the segmentation map if return_coco_annotation is set to True. Set to None if no mask if found above threshold.

  • segments_info — A dictionary that contains additional information on each segment.

    • id — An integer representing the segment_id.

    • label_id — An integer representing the label / semantic class id corresponding to segment_id.

    • score — Prediction score of segment with segment_id.

Converts the output of DetrForSegmentation into instance segmentation predictions. Only supports PyTorch.

post_process_panoptic_segmentation

<source>

( outputsthreshold: float = 0.5mask_threshold: float = 0.5overlap_mask_area_threshold: float = 0.8label_ids_to_fuse: typing.Optional[typing.Set[int]] = Nonetarget_sizes: typing.Union[typing.List[typing.Tuple[int, int]], NoneType] = None ) → List[Dict]

Parameters

  • outputs (DetrForSegmentation) — The outputs from DetrForSegmentation.

  • threshold (float, optional, defaults to 0.5) — The probability score threshold to keep predicted instance masks.

  • mask_threshold (float, optional, defaults to 0.5) — Threshold to use when turning the predicted masks into binary values.

  • overlap_mask_area_threshold (float, optional, defaults to 0.8) — The overlap mask area threshold to merge or discard small disconnected parts within each binary instance mask.

  • label_ids_to_fuse (Set[int], optional) — The labels in this state will have all their instances be fused together. For instance we could say there can only be one sky in an image, but several persons, so the label ID for sky would be in that set, but not the one for person.

  • target_sizes (List[Tuple], optional) — List of length (batch_size), where each list item (Tuple[int, int]]) corresponds to the requested final size (height, width) of each prediction in batch. If unset, predictions will not be resized.

Returns

List[Dict]

A list of dictionaries, one per image, each dictionary containing two keys:

  • segmentation — a tensor of shape (height, width) where each pixel represents a segment_id or None if no mask if found above threshold. If target_sizes is specified, segmentation is resized to the corresponding target_sizes entry.

  • segments_info — A dictionary that contains additional information on each segment.

    • id — an integer representing the segment_id.

    • label_id — An integer representing the label / semantic class id corresponding to segment_id.

    • was_fused — a boolean, True if label_id was in label_ids_to_fuse, False otherwise. Multiple instances of the same class / label were fused and assigned a single segment_id.

    • score — Prediction score of segment with segment_id.

Converts the output of DetrForSegmentation into image panoptic segmentation predictions. Only supports PyTorch.

DetrFeatureExtractor

class transformers.DetrFeatureExtractor

<source>

( *args**kwargs )

__call__

<source>

( images**kwargs )

Preprocess an image or a batch of images.

post_process_object_detection

<source>

( outputsthreshold: float = 0.5target_sizes: typing.Union[transformers.utils.generic.TensorType, typing.List[typing.Tuple]] = None ) → List[Dict]

Parameters

  • outputs (DetrObjectDetectionOutput) — Raw outputs of the model.

  • threshold (float, optional) — Score threshold to keep object detection predictions.

  • target_sizes (torch.Tensor or List[Tuple[int, int]], optional) — Tensor of shape (batch_size, 2) or list of tuples (Tuple[int, int]) containing the target size (height, width) of each image in the batch. If unset, predictions will not be resized.

Returns

List[Dict]

A list of dictionaries, each dictionary containing the scores, labels and boxes for an image in the batch as predicted by the model.

Converts the raw output of DetrForObjectDetection into final bounding boxes in (top_left_x, top_left_y, bottom_right_x, bottom_right_y) format. Only supports PyTorch.

post_process_semantic_segmentation

<source>

( outputstarget_sizes: typing.List[typing.Tuple[int, int]] = None ) → List[torch.Tensor]

Parameters

  • outputs (DetrForSegmentation) — Raw outputs of the model.

  • target_sizes (List[Tuple[int, int]], optional) — A list of tuples (Tuple[int, int]) containing the target size (height, width) of each image in the batch. If unset, predictions will not be resized.

Returns

List[torch.Tensor]

A list of length batch_size, where each item is a semantic segmentation map of shape (height, width) corresponding to the target_sizes entry (if target_sizes is specified). Each entry of each torch.Tensor correspond to a semantic class id.

Converts the output of DetrForSegmentation into semantic segmentation maps. Only supports PyTorch.

post_process_instance_segmentation

<source>

( outputsthreshold: float = 0.5mask_threshold: float = 0.5overlap_mask_area_threshold: float = 0.8target_sizes: typing.Union[typing.List[typing.Tuple[int, int]], NoneType] = Nonereturn_coco_annotation: typing.Optional[bool] = False ) → List[Dict]

Parameters

  • outputs (DetrForSegmentation) — Raw outputs of the model.

  • threshold (float, optional, defaults to 0.5) — The probability score threshold to keep predicted instance masks.

  • mask_threshold (float, optional, defaults to 0.5) — Threshold to use when turning the predicted masks into binary values.

  • overlap_mask_area_threshold (float, optional, defaults to 0.8) — The overlap mask area threshold to merge or discard small disconnected parts within each binary instance mask.

  • target_sizes (List[Tuple], optional) — List of length (batch_size), where each list item (Tuple[int, int]]) corresponds to the requested final size (height, width) of each prediction. If unset, predictions will not be resized.

  • return_coco_annotation (bool, optional) — Defaults to False. If set to True, segmentation maps are returned in COCO run-length encoding (RLE) format.

Returns

List[Dict]

A list of dictionaries, one per image, each dictionary containing two keys:

  • segmentation — A tensor of shape (height, width) where each pixel represents a segment_id or List[List] run-length encoding (RLE) of the segmentation map if return_coco_annotation is set to True. Set to None if no mask if found above threshold.

  • segments_info — A dictionary that contains additional information on each segment.

    • id — An integer representing the segment_id.

    • label_id — An integer representing the label / semantic class id corresponding to segment_id.

    • score — Prediction score of segment with segment_id.

Converts the output of DetrForSegmentation into instance segmentation predictions. Only supports PyTorch.

post_process_panoptic_segmentation

<source>

( outputsthreshold: float = 0.5mask_threshold: float = 0.5overlap_mask_area_threshold: float = 0.8label_ids_to_fuse: typing.Optional[typing.Set[int]] = Nonetarget_sizes: typing.Union[typing.List[typing.Tuple[int, int]], NoneType] = None ) → List[Dict]

Parameters

  • outputs (DetrForSegmentation) — The outputs from DetrForSegmentation.

  • threshold (float, optional, defaults to 0.5) — The probability score threshold to keep predicted instance masks.

  • mask_threshold (float, optional, defaults to 0.5) — Threshold to use when turning the predicted masks into binary values.

  • overlap_mask_area_threshold (float, optional, defaults to 0.8) — The overlap mask area threshold to merge or discard small disconnected parts within each binary instance mask.

  • label_ids_to_fuse (Set[int], optional) — The labels in this state will have all their instances be fused together. For instance we could say there can only be one sky in an image, but several persons, so the label ID for sky would be in that set, but not the one for person.

  • target_sizes (List[Tuple], optional) — List of length (batch_size), where each list item (Tuple[int, int]]) corresponds to the requested final size (height, width) of each prediction in batch. If unset, predictions will not be resized.

Returns

List[Dict]

A list of dictionaries, one per image, each dictionary containing two keys:

  • segmentation — a tensor of shape (height, width) where each pixel represents a segment_id or None if no mask if found above threshold. If target_sizes is specified, segmentation is resized to the corresponding target_sizes entry.

  • segments_info — A dictionary that contains additional information on each segment.

    • id — an integer representing the segment_id.

    • label_id — An integer representing the label / semantic class id corresponding to segment_id.

    • was_fused — a boolean, True if label_id was in label_ids_to_fuse, False otherwise. Multiple instances of the same class / label were fused and assigned a single segment_id.

    • score — Prediction score of segment with segment_id.

Converts the output of DetrForSegmentation into image panoptic segmentation predictions. Only supports PyTorch.

DetrModel

class transformers.DetrModel

<source>

( config: DetrConfig )

Parameters

  • config (DetrConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

The bare DETR Model (consisting of a backbone and encoder-decoder Transformer) outputting raw hidden-states without any specific head on top.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

forward

<source>

( pixel_values: FloatTensorpixel_mask: typing.Optional[torch.LongTensor] = Nonedecoder_attention_mask: typing.Optional[torch.FloatTensor] = Noneencoder_outputs: typing.Optional[torch.FloatTensor] = Noneinputs_embeds: typing.Optional[torch.FloatTensor] = Nonedecoder_inputs_embeds: typing.Optional[torch.FloatTensor] = Noneoutput_attentions: typing.Optional[bool] = Noneoutput_hidden_states: typing.Optional[bool] = Nonereturn_dict: typing.Optional[bool] = None ) → transformers.models.detr.modeling_detr.DetrModelOutput or tuple(torch.FloatTensor)

Parameters

  • pixel_values (torch.FloatTensor of shape (batch_size, num_channels, height, width)) — Pixel values. Padding will be ignored by default should you provide it.

    Pixel values can be obtained using AutoImageProcessor. See DetrImageProcessor.call() for details.

  • pixel_mask (torch.LongTensor of shape (batch_size, height, width), optional) — Mask to avoid performing attention on padding pixel values. Mask values selected in [0, 1]:

    • 1 for pixels that are real (i.e. not masked),

    • 0 for pixels that are padding (i.e. masked).

    What are attention masks?

  • decoder_attention_mask (torch.FloatTensor of shape (batch_size, num_queries), optional) — Not used by default. Can be used to mask object queries.

  • encoder_outputs (tuple(tuple(torch.FloatTensor), optional) — Tuple consists of (last_hidden_state, optional: hidden_states, optional: attentions) last_hidden_state of shape (batch_size, sequence_length, hidden_size), optional) is a sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.

  • inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passing the flattened feature map (output of the backbone + projection layer), you can choose to directly pass a flattened representation of an image.

  • decoder_inputs_embeds (torch.FloatTensor of shape (batch_size, num_queries, hidden_size), optional) — Optionally, instead of initializing the queries with a tensor of zeros, you can choose to directly pass an embedded representation.

  • output_attentions (bool, optional) — Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.

  • output_hidden_states (bool, optional) — Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.

  • return_dict (bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple.

Returns

transformers.models.detr.modeling_detr.DetrModelOutput or tuple(torch.FloatTensor)

A transformers.models.detr.modeling_detr.DetrModelOutput or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (DetrConfig) and inputs.

  • last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size)) — Sequence of hidden-states at the output of the last layer of the decoder of the model.

  • decoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.

  • decoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads.

  • cross_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the decoder’s cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads.

  • encoder_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the encoder of the model.

  • encoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.

  • encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads.

  • intermediate_hidden_states (torch.FloatTensor of shape (config.decoder_layers, batch_size, sequence_length, hidden_size), optional, returned when config.auxiliary_loss=True) — Intermediate decoder activations, i.e. the output of each decoder layer, each of them gone through a layernorm.

The DetrModel forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Examples:

Copied

>>> from transformers import AutoImageProcessor, DetrModel
>>> from PIL import Image
>>> import requests

>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)

>>> image_processor = AutoImageProcessor.from_pretrained("facebook/detr-resnet-50")
>>> model = DetrModel.from_pretrained("facebook/detr-resnet-50")

>>> # prepare image for the model
>>> inputs = image_processor(images=image, return_tensors="pt")

>>> # forward pass
>>> outputs = model(**inputs)

>>> # the last hidden states are the final query embeddings of the Transformer decoder
>>> # these are of shape (batch_size, num_queries, hidden_size)
>>> last_hidden_states = outputs.last_hidden_state
>>> list(last_hidden_states.shape)
[1, 100, 256]

DetrForObjectDetection

class transformers.DetrForObjectDetection

<source>

( config: DetrConfig )

Parameters

  • config (DetrConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

DETR Model (consisting of a backbone and encoder-decoder Transformer) with object detection heads on top, for tasks such as COCO detection.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

forward

<source>

( pixel_values: FloatTensorpixel_mask: typing.Optional[torch.LongTensor] = Nonedecoder_attention_mask: typing.Optional[torch.FloatTensor] = Noneencoder_outputs: typing.Optional[torch.FloatTensor] = Noneinputs_embeds: typing.Optional[torch.FloatTensor] = Nonedecoder_inputs_embeds: typing.Optional[torch.FloatTensor] = Nonelabels: typing.Optional[typing.List[dict]] = Noneoutput_attentions: typing.Optional[bool] = Noneoutput_hidden_states: typing.Optional[bool] = Nonereturn_dict: typing.Optional[bool] = None ) → transformers.models.detr.modeling_detr.DetrObjectDetectionOutput or tuple(torch.FloatTensor)

Parameters

  • pixel_values (torch.FloatTensor of shape (batch_size, num_channels, height, width)) — Pixel values. Padding will be ignored by default should you provide it.

    Pixel values can be obtained using AutoImageProcessor. See DetrImageProcessor.call() for details.

  • pixel_mask (torch.LongTensor of shape (batch_size, height, width), optional) — Mask to avoid performing attention on padding pixel values. Mask values selected in [0, 1]:

    • 1 for pixels that are real (i.e. not masked),

    • 0 for pixels that are padding (i.e. masked).

    What are attention masks?

  • decoder_attention_mask (torch.FloatTensor of shape (batch_size, num_queries), optional) — Not used by default. Can be used to mask object queries.

  • encoder_outputs (tuple(tuple(torch.FloatTensor), optional) — Tuple consists of (last_hidden_state, optional: hidden_states, optional: attentions) last_hidden_state of shape (batch_size, sequence_length, hidden_size), optional) is a sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.

  • inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passing the flattened feature map (output of the backbone + projection layer), you can choose to directly pass a flattened representation of an image.

  • decoder_inputs_embeds (torch.FloatTensor of shape (batch_size, num_queries, hidden_size), optional) — Optionally, instead of initializing the queries with a tensor of zeros, you can choose to directly pass an embedded representation.

  • output_attentions (bool, optional) — Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.

  • output_hidden_states (bool, optional) — Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.

  • return_dict (bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple.

  • labels (List[Dict] of len (batch_size,), optional) — Labels for computing the bipartite matching loss. List of dicts, each dictionary containing at least the following 2 keys: ‘class_labels’ and ‘boxes’ (the class labels and bounding boxes of an image in the batch respectively). The class labels themselves should be a torch.LongTensor of len (number of bounding boxes in the image,) and the boxes a torch.FloatTensor of shape (number of bounding boxes in the image, 4).

Returns

transformers.models.detr.modeling_detr.DetrObjectDetectionOutput or tuple(torch.FloatTensor)

A transformers.models.detr.modeling_detr.DetrObjectDetectionOutput or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (DetrConfig) and inputs.

  • loss (torch.FloatTensor of shape (1,), optional, returned when labels are provided)) — Total loss as a linear combination of a negative log-likehood (cross-entropy) for class prediction and a bounding box loss. The latter is defined as a linear combination of the L1 loss and the generalized scale-invariant IoU loss.

  • loss_dict (Dict, optional) — A dictionary containing the individual losses. Useful for logging.

  • logits (torch.FloatTensor of shape (batch_size, num_queries, num_classes + 1)) — Classification logits (including no-object) for all queries.

  • pred_boxes (torch.FloatTensor of shape (batch_size, num_queries, 4)) — Normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height). These values are normalized in [0, 1], relative to the size of each individual image in the batch (disregarding possible padding). You can use post_process_object_detection() to retrieve the unnormalized bounding boxes.

  • auxiliary_outputs (list[Dict], optional) — Optional, only returned when auxilary losses are activated (i.e. config.auxiliary_loss is set to True) and labels are provided. It is a list of dictionaries containing the two above keys (logits and pred_boxes) for each decoder layer.

  • last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the decoder of the model.

  • decoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.

  • decoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads.

  • cross_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the decoder’s cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads.

  • encoder_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the encoder of the model.

  • encoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.

  • encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads.

The DetrForObjectDetection forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Examples:

Copied

>>> from transformers import AutoImageProcessor, DetrForObjectDetection
>>> import torch
>>> from PIL import Image
>>> import requests

>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)

>>> image_processor = AutoImageProcessor.from_pretrained("facebook/detr-resnet-50")
>>> model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50")

>>> inputs = image_processor(images=image, return_tensors="pt")
>>> outputs = model(**inputs)

>>> # convert outputs (bounding boxes and class logits) to COCO API
>>> target_sizes = torch.tensor([image.size[::-1]])
>>> results = image_processor.post_process_object_detection(outputs, threshold=0.9, target_sizes=target_sizes)[
...     0
... ]

>>> for score, label, box in zip(results["scores"], results["labels"], results["boxes"]):
...     box = [round(i, 2) for i in box.tolist()]
...     print(
...         f"Detected {model.config.id2label[label.item()]} with confidence "
...         f"{round(score.item(), 3)} at location {box}"
...     )
Detected remote with confidence 0.998 at location [40.16, 70.81, 175.55, 117.98]
Detected remote with confidence 0.996 at location [333.24, 72.55, 368.33, 187.66]
Detected couch with confidence 0.995 at location [-0.02, 1.15, 639.73, 473.76]
Detected cat with confidence 0.999 at location [13.24, 52.05, 314.02, 470.93]
Detected cat with confidence 0.999 at location [345.4, 23.85, 640.37, 368.72]

DetrForSegmentation

class transformers.DetrForSegmentation

<source>

( config: DetrConfig )

Parameters

  • config (DetrConfig) — Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the from_pretrained() method to load the model weights.

DETR Model (consisting of a backbone and encoder-decoder Transformer) with a segmentation head on top, for tasks such as COCO panoptic.

This model inherits from PreTrainedModel. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.)

This model is also a PyTorch torch.nn.Module subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior.

forward

<source>

( pixel_values: FloatTensorpixel_mask: typing.Optional[torch.LongTensor] = Nonedecoder_attention_mask: typing.Optional[torch.FloatTensor] = Noneencoder_outputs: typing.Optional[torch.FloatTensor] = Noneinputs_embeds: typing.Optional[torch.FloatTensor] = Nonedecoder_inputs_embeds: typing.Optional[torch.FloatTensor] = Nonelabels: typing.Optional[typing.List[dict]] = Noneoutput_attentions: typing.Optional[bool] = Noneoutput_hidden_states: typing.Optional[bool] = Nonereturn_dict: typing.Optional[bool] = None ) → transformers.models.detr.modeling_detr.DetrSegmentationOutput or tuple(torch.FloatTensor)

Parameters

  • pixel_values (torch.FloatTensor of shape (batch_size, num_channels, height, width)) — Pixel values. Padding will be ignored by default should you provide it.

    Pixel values can be obtained using AutoImageProcessor. See DetrImageProcessor.call() for details.

  • pixel_mask (torch.LongTensor of shape (batch_size, height, width), optional) — Mask to avoid performing attention on padding pixel values. Mask values selected in [0, 1]:

    • 1 for pixels that are real (i.e. not masked),

    • 0 for pixels that are padding (i.e. masked).

    What are attention masks?

  • decoder_attention_mask (torch.FloatTensor of shape (batch_size, num_queries), optional) — Not used by default. Can be used to mask object queries.

  • encoder_outputs (tuple(tuple(torch.FloatTensor), optional) — Tuple consists of (last_hidden_state, optional: hidden_states, optional: attentions) last_hidden_state of shape (batch_size, sequence_length, hidden_size), optional) is a sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.

  • inputs_embeds (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Optionally, instead of passing the flattened feature map (output of the backbone + projection layer), you can choose to directly pass a flattened representation of an image.

  • decoder_inputs_embeds (torch.FloatTensor of shape (batch_size, num_queries, hidden_size), optional) — Optionally, instead of initializing the queries with a tensor of zeros, you can choose to directly pass an embedded representation.

  • output_attentions (bool, optional) — Whether or not to return the attentions tensors of all attention layers. See attentions under returned tensors for more detail.

  • output_hidden_states (bool, optional) — Whether or not to return the hidden states of all layers. See hidden_states under returned tensors for more detail.

  • return_dict (bool, optional) — Whether or not to return a ModelOutput instead of a plain tuple.

  • labels (List[Dict] of len (batch_size,), optional) — Labels for computing the bipartite matching loss, DICE/F-1 loss and Focal loss. List of dicts, each dictionary containing at least the following 3 keys: ‘class_labels’, ‘boxes’ and ‘masks’ (the class labels, bounding boxes and segmentation masks of an image in the batch respectively). The class labels themselves should be a torch.LongTensor of len (number of bounding boxes in the image,), the boxes a torch.FloatTensor of shape (number of bounding boxes in the image, 4) and the masks a torch.FloatTensor of shape (number of bounding boxes in the image, height, width).

Returns

transformers.models.detr.modeling_detr.DetrSegmentationOutput or tuple(torch.FloatTensor)

A transformers.models.detr.modeling_detr.DetrSegmentationOutput or a tuple of torch.FloatTensor (if return_dict=False is passed or when config.return_dict=False) comprising various elements depending on the configuration (DetrConfig) and inputs.

  • loss (torch.FloatTensor of shape (1,), optional, returned when labels are provided)) — Total loss as a linear combination of a negative log-likehood (cross-entropy) for class prediction and a bounding box loss. The latter is defined as a linear combination of the L1 loss and the generalized scale-invariant IoU loss.

  • loss_dict (Dict, optional) — A dictionary containing the individual losses. Useful for logging.

  • logits (torch.FloatTensor of shape (batch_size, num_queries, num_classes + 1)) — Classification logits (including no-object) for all queries.

  • pred_boxes (torch.FloatTensor of shape (batch_size, num_queries, 4)) — Normalized boxes coordinates for all queries, represented as (center_x, center_y, width, height). These values are normalized in [0, 1], relative to the size of each individual image in the batch (disregarding possible padding). You can use post_process_object_detection() to retrieve the unnormalized bounding boxes.

  • pred_masks (torch.FloatTensor of shape (batch_size, num_queries, height/4, width/4)) — Segmentation masks logits for all queries. See also post_process_semantic_segmentation() or post_process_instance_segmentation() post_process_panoptic_segmentation() to evaluate semantic, instance and panoptic segmentation masks respectively.

  • auxiliary_outputs (list[Dict], optional) — Optional, only returned when auxiliary losses are activated (i.e. config.auxiliary_loss is set to True) and labels are provided. It is a list of dictionaries containing the two above keys (logits and pred_boxes) for each decoder layer.

  • last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the decoder of the model.

  • decoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the decoder at the output of each layer plus the initial embedding outputs.

  • decoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads.

  • cross_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the decoder’s cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads.

  • encoder_last_hidden_state (torch.FloatTensor of shape (batch_size, sequence_length, hidden_size), optional) — Sequence of hidden-states at the output of the last layer of the encoder of the model.

  • encoder_hidden_states (tuple(torch.FloatTensor), optional, returned when output_hidden_states=True is passed or when config.output_hidden_states=True) — Tuple of torch.FloatTensor (one for the output of the embeddings + one for the output of each layer) of shape (batch_size, sequence_length, hidden_size). Hidden-states of the encoder at the output of each layer plus the initial embedding outputs.

  • encoder_attentions (tuple(torch.FloatTensor), optional, returned when output_attentions=True is passed or when config.output_attentions=True) — Tuple of torch.FloatTensor (one for each layer) of shape (batch_size, num_heads, sequence_length, sequence_length). Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads.

The DetrForSegmentation forward method, overrides the __call__ special method.

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the pre and post processing steps while the latter silently ignores them.

Examples:

Copied

>>> import io
>>> import requests
>>> from PIL import Image
>>> import torch
>>> import numpy

>>> from transformers import AutoImageProcessor, DetrForSegmentation
>>> from transformers.image_transforms import rgb_to_id

>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
>>> image = Image.open(requests.get(url, stream=True).raw)

>>> image_processor = AutoImageProcessor.from_pretrained("facebook/detr-resnet-50-panoptic")
>>> model = DetrForSegmentation.from_pretrained("facebook/detr-resnet-50-panoptic")

>>> # prepare image for the model
>>> inputs = image_processor(images=image, return_tensors="pt")

>>> # forward pass
>>> outputs = model(**inputs)

>>> # Use the `post_process_panoptic_segmentation` method of the `image_processor` to retrieve post-processed panoptic segmentation maps
>>> # Segmentation results are returned as a list of dictionaries
>>> result = image_processor.post_process_panoptic_segmentation(outputs, target_sizes=[(300, 500)])

>>> # A tensor of shape (height, width) where each value denotes a segment id, filled with -1 if no segment is found
>>> panoptic_seg = result[0]["segmentation"]
>>> # Get prediction score and segment_id to class_id mapping of each segment
>>> panoptic_segments_info = result[0]["segments_info"]

Last updated