# Copyright 2022 School of EIC, Huazhong University of Science & Technology and 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. """PyTorch YOLOS model.""" import collections.abc from collections.abc import Callable from dataclasses import dataclass import torch from torch import nn from ...activations import ACT2FN from ...modeling_layers import GradientCheckpointingLayer from ...modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling from ...modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel from ...processing_utils import Unpack from ...utils import ModelOutput, TransformersKwargs, auto_docstring, logging from ...utils.generic import can_return_tuple, merge_with_config_defaults from ...utils.output_capturing import capture_outputs from .configuration_yolos import YolosConfig logger = logging.get_logger(__name__) @dataclass @auto_docstring( custom_intro=""" Output type of [`YolosForObjectDetection`]. """ ) class YolosObjectDetectionOutput(ModelOutput): r""" 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 [`~YolosImageProcessor.post_process`] to retrieve the unnormalized bounding boxes. 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. """ loss: torch.FloatTensor | None = None loss_dict: dict | None = None logits: torch.FloatTensor | None = None pred_boxes: torch.FloatTensor | None = None auxiliary_outputs: list[dict] | None = None last_hidden_state: torch.FloatTensor | None = None hidden_states: tuple[torch.FloatTensor] | None = None attentions: tuple[torch.FloatTensor] | None = None class YolosEmbeddings(nn.Module): """ Construct the CLS token, detection tokens, position and patch embeddings. """ def __init__(self, config: YolosConfig) -> None: super().__init__() self.cls_token = nn.Parameter(torch.zeros(1, 1, config.hidden_size)) self.detection_tokens = nn.Parameter(torch.zeros(1, config.num_detection_tokens, config.hidden_size)) self.patch_embeddings = YolosPatchEmbeddings(config) num_patches = self.patch_embeddings.num_patches self.position_embeddings = nn.Parameter( torch.zeros(1, num_patches + config.num_detection_tokens + 1, config.hidden_size) ) self.dropout = nn.Dropout(config.hidden_dropout_prob) self.interpolation = InterpolateInitialPositionEmbeddings(config) self.config = config def forward(self, pixel_values: torch.Tensor) -> torch.Tensor: batch_size, num_channels, height, width = pixel_values.shape embeddings = self.patch_embeddings(pixel_values) batch_size, seq_len, _ = embeddings.size() # add the [CLS] and detection tokens to the embedded patch tokens cls_tokens = self.cls_token.expand(batch_size, -1, -1) detection_tokens = self.detection_tokens.expand(batch_size, -1, -1) embeddings = torch.cat((cls_tokens, embeddings, detection_tokens), dim=1) # add positional encoding to each token # this might require interpolation of the existing position embeddings position_embeddings = self.interpolation(self.position_embeddings, (height, width)) embeddings = embeddings + position_embeddings embeddings = self.dropout(embeddings) return embeddings class InterpolateInitialPositionEmbeddings(nn.Module): def __init__(self, config) -> None: super().__init__() self.config = config def forward(self, pos_embed, img_size=(800, 1344)) -> torch.Tensor: cls_pos_embed = pos_embed[:, 0, :] cls_pos_embed = cls_pos_embed[:, None] det_pos_embed = pos_embed[:, -self.config.num_detection_tokens :, :] patch_pos_embed = pos_embed[:, 1 : -self.config.num_detection_tokens, :] patch_pos_embed = patch_pos_embed.transpose(1, 2) batch_size, hidden_size, seq_len = patch_pos_embed.shape patch_height, patch_width = ( self.config.image_size[0] // self.config.patch_size, self.config.image_size[1] // self.config.patch_size, ) patch_pos_embed = patch_pos_embed.view(batch_size, hidden_size, patch_height, patch_width) height, width = img_size new_patch_height, new_patch_width = height // self.config.patch_size, width // self.config.patch_size patch_pos_embed = nn.functional.interpolate( patch_pos_embed, size=(new_patch_height, new_patch_width), mode="bicubic", align_corners=False ) patch_pos_embed = patch_pos_embed.flatten(2).transpose(1, 2) scale_pos_embed = torch.cat((cls_pos_embed, patch_pos_embed, det_pos_embed), dim=1) return scale_pos_embed class InterpolateMidPositionEmbeddings(nn.Module): def __init__(self, config) -> None: super().__init__() self.config = config def forward(self, pos_embed, img_size=(800, 1344)) -> torch.Tensor: cls_pos_embed = pos_embed[:, :, 0, :] cls_pos_embed = cls_pos_embed[:, None] det_pos_embed = pos_embed[:, :, -self.config.num_detection_tokens :, :] patch_pos_embed = pos_embed[:, :, 1 : -self.config.num_detection_tokens, :] patch_pos_embed = patch_pos_embed.transpose(2, 3) depth, batch_size, hidden_size, seq_len = patch_pos_embed.shape patch_height, patch_width = ( self.config.image_size[0] // self.config.patch_size, self.config.image_size[1] // self.config.patch_size, ) patch_pos_embed = patch_pos_embed.view(depth * batch_size, hidden_size, patch_height, patch_width) height, width = img_size new_patch_height, new_patch_width = height // self.config.patch_size, width // self.config.patch_size patch_pos_embed = nn.functional.interpolate( patch_pos_embed, size=(new_patch_height, new_patch_width), mode="bicubic", align_corners=False ) patch_pos_embed = ( patch_pos_embed.flatten(2) .transpose(1, 2) .contiguous() .view(depth, batch_size, new_patch_height * new_patch_width, hidden_size) ) scale_pos_embed = torch.cat((cls_pos_embed, patch_pos_embed, det_pos_embed), dim=2) return scale_pos_embed class YolosPatchEmbeddings(nn.Module): """ This class turns `pixel_values` of shape `(batch_size, num_channels, height, width)` into the initial `hidden_states` (patch embeddings) of shape `(batch_size, seq_length, hidden_size)` to be consumed by a Transformer. """ def __init__(self, config): super().__init__() image_size, patch_size = config.image_size, config.patch_size num_channels, hidden_size = config.num_channels, config.hidden_size image_size = image_size if isinstance(image_size, collections.abc.Iterable) else (image_size, image_size) patch_size = patch_size if isinstance(patch_size, collections.abc.Iterable) else (patch_size, patch_size) num_patches = (image_size[1] // patch_size[1]) * (image_size[0] // patch_size[0]) self.image_size = image_size self.patch_size = patch_size self.num_channels = num_channels self.num_patches = num_patches self.projection = nn.Conv2d(num_channels, hidden_size, kernel_size=patch_size, stride=patch_size) def forward(self, pixel_values: torch.Tensor) -> torch.Tensor: batch_size, num_channels, height, width = pixel_values.shape if num_channels != self.num_channels: raise ValueError( "Make sure that the channel dimension of the pixel values match with the one set in the configuration." ) embeddings = self.projection(pixel_values).flatten(2).transpose(1, 2) return embeddings # Copied from transformers.models.bert.modeling_bert.eager_attention_forward def eager_attention_forward( module: nn.Module, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, attention_mask: torch.Tensor | None, scaling: float | None = None, dropout: float = 0.0, **kwargs: Unpack[TransformersKwargs], ): if scaling is None: scaling = query.size(-1) ** -0.5 # Take the dot product between "query" and "key" to get the raw attention scores. attn_weights = torch.matmul(query, key.transpose(2, 3)) * scaling if attention_mask is not None: attn_weights = attn_weights + attention_mask attn_weights = nn.functional.softmax(attn_weights, dim=-1) attn_weights = nn.functional.dropout(attn_weights, p=dropout, training=module.training) attn_output = torch.matmul(attn_weights, value) attn_output = attn_output.transpose(1, 2).contiguous() return attn_output, attn_weights # Copied from transformers.models.vit.modeling_vit.ViTSelfAttention with ViT->Yolos class YolosSelfAttention(nn.Module): def __init__(self, config: YolosConfig): super().__init__() if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"): raise ValueError( f"The hidden size {config.hidden_size} is not a multiple of the number of attention " f"heads {config.num_attention_heads}." ) self.config = config self.num_attention_heads = config.num_attention_heads self.attention_head_size = int(config.hidden_size / config.num_attention_heads) self.all_head_size = self.num_attention_heads * self.attention_head_size self.dropout_prob = config.attention_probs_dropout_prob self.scaling = self.attention_head_size**-0.5 self.is_causal = False self.query = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias) self.key = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias) self.value = nn.Linear(config.hidden_size, self.all_head_size, bias=config.qkv_bias) def forward(self, hidden_states: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: batch_size = hidden_states.shape[0] new_shape = batch_size, -1, self.num_attention_heads, self.attention_head_size key_layer = self.key(hidden_states).view(*new_shape).transpose(1, 2) value_layer = self.value(hidden_states).view(*new_shape).transpose(1, 2) query_layer = self.query(hidden_states).view(*new_shape).transpose(1, 2) attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface( self.config._attn_implementation, eager_attention_forward ) context_layer, attention_probs = attention_interface( self, query_layer, key_layer, value_layer, None, is_causal=self.is_causal, scaling=self.scaling, dropout=0.0 if not self.training else self.dropout_prob, ) new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,) context_layer = context_layer.reshape(new_context_layer_shape) return context_layer, attention_probs # Copied from transformers.models.vit.modeling_vit.ViTSelfOutput with ViT->Yolos class YolosSelfOutput(nn.Module): """ The residual connection is defined in YolosLayer instead of here (as is the case with other models), due to the layernorm applied before each block. """ def __init__(self, config: YolosConfig): super().__init__() self.dense = nn.Linear(config.hidden_size, config.hidden_size) self.dropout = nn.Dropout(config.hidden_dropout_prob) def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor: hidden_states = self.dense(hidden_states) hidden_states = self.dropout(hidden_states) return hidden_states # Copied from transformers.models.vit.modeling_vit.ViTAttention with ViT->Yolos class YolosAttention(nn.Module): def __init__(self, config: YolosConfig): super().__init__() self.attention = YolosSelfAttention(config) self.output = YolosSelfOutput(config) def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: self_attn_output, _ = self.attention(hidden_states) output = self.output(self_attn_output, hidden_states) return output # Copied from transformers.models.vit.modeling_vit.ViTIntermediate with ViT->Yolos class YolosIntermediate(nn.Module): def __init__(self, config: YolosConfig): super().__init__() self.dense = nn.Linear(config.hidden_size, config.intermediate_size) if isinstance(config.hidden_act, str): self.intermediate_act_fn = ACT2FN[config.hidden_act] else: self.intermediate_act_fn = config.hidden_act def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: hidden_states = self.dense(hidden_states) hidden_states = self.intermediate_act_fn(hidden_states) return hidden_states # Copied from transformers.models.vit.modeling_vit.ViTOutput with ViT->Yolos class YolosOutput(nn.Module): def __init__(self, config: YolosConfig): super().__init__() self.dense = nn.Linear(config.intermediate_size, config.hidden_size) self.dropout = nn.Dropout(config.hidden_dropout_prob) def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor: hidden_states = self.dense(hidden_states) hidden_states = self.dropout(hidden_states) hidden_states = hidden_states + input_tensor return hidden_states # Copied from transformers.models.vit.modeling_vit.ViTLayer with ViT->Yolos,VIT->YOLOS class YolosLayer(GradientCheckpointingLayer): """This corresponds to the Block class in the timm implementation.""" def __init__(self, config: YolosConfig): super().__init__() self.chunk_size_feed_forward = config.chunk_size_feed_forward self.seq_len_dim = 1 self.attention = YolosAttention(config) self.intermediate = YolosIntermediate(config) self.output = YolosOutput(config) self.layernorm_before = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) self.layernorm_after = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: hidden_states_norm = self.layernorm_before(hidden_states) attention_output = self.attention(hidden_states_norm) # first residual connection hidden_states = attention_output + hidden_states # in Yolos, layernorm is also applied after self-attention layer_output = self.layernorm_after(hidden_states) layer_output = self.intermediate(layer_output) # second residual connection is done here layer_output = self.output(layer_output, hidden_states) return layer_output class YolosEncoder(nn.Module): def __init__(self, config: YolosConfig) -> None: super().__init__() self.config = config self.layer = nn.ModuleList([YolosLayer(config) for _ in range(config.num_hidden_layers)]) self.gradient_checkpointing = False seq_length = ( 1 + (config.image_size[0] * config.image_size[1] // config.patch_size**2) + config.num_detection_tokens ) self.mid_position_embeddings = ( nn.Parameter( torch.zeros( config.num_hidden_layers - 1, 1, seq_length, config.hidden_size, ) ) if config.use_mid_position_embeddings else None ) self.interpolation = InterpolateMidPositionEmbeddings(config) if config.use_mid_position_embeddings else None def forward( self, hidden_states: torch.Tensor, height: int, width: int, ) -> BaseModelOutput: if self.config.use_mid_position_embeddings: interpolated_mid_position_embeddings = self.interpolation(self.mid_position_embeddings, (height, width)) for i, layer_module in enumerate(self.layer): hidden_states = layer_module(hidden_states) if self.config.use_mid_position_embeddings: if i < (self.config.num_hidden_layers - 1): hidden_states = hidden_states + interpolated_mid_position_embeddings[i] return BaseModelOutput(last_hidden_state=hidden_states) @auto_docstring class YolosPreTrainedModel(PreTrainedModel): config: YolosConfig base_model_prefix = "vit" main_input_name = "pixel_values" input_modalities = ("image",) supports_gradient_checkpointing = True _no_split_modules = [] _supports_sdpa = True _supports_flash_attn = True _supports_flex_attn = True _supports_attention_backend = True _can_record_outputs = { "hidden_states": YolosLayer, "attentions": YolosSelfAttention, } @auto_docstring class YolosModel(YolosPreTrainedModel): def __init__(self, config: YolosConfig, add_pooling_layer: bool = True): r""" add_pooling_layer (bool, *optional*, defaults to `True`): Whether to add a pooling layer """ super().__init__(config) self.config = config self.embeddings = YolosEmbeddings(config) self.encoder = YolosEncoder(config) self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps) self.pooler = YolosPooler(config) if add_pooling_layer else None # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self) -> YolosPatchEmbeddings: return self.embeddings.patch_embeddings @merge_with_config_defaults @capture_outputs(tie_last_hidden_states=False) @auto_docstring def forward( self, pixel_values: torch.Tensor | None = None, **kwargs: Unpack[TransformersKwargs], ) -> BaseModelOutputWithPooling: if pixel_values is None: raise ValueError("You have to specify pixel_values") embedding_output = self.embeddings(pixel_values) height, width = pixel_values.shape[-2:] encoder_outputs: BaseModelOutput = self.encoder(embedding_output, height=height, width=width) sequence_output = encoder_outputs.last_hidden_state sequence_output = self.layernorm(sequence_output) pooled_output = self.pooler(sequence_output) if self.pooler is not None else None return BaseModelOutputWithPooling(last_hidden_state=sequence_output, pooler_output=pooled_output) class YolosPooler(nn.Module): def __init__(self, config: YolosConfig): super().__init__() self.dense = nn.Linear(config.hidden_size, config.hidden_size) self.activation = nn.Tanh() def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: # We "pool" the model by simply taking the hidden state corresponding # to the first token. first_token_tensor = hidden_states[:, 0] pooled_output = self.dense(first_token_tensor) pooled_output = self.activation(pooled_output) return pooled_output # Copied from transformers.models.detr.modeling_detr.DetrMLPPredictionHead with Detr->Yolos class YolosMLPPredictionHead(nn.Module): """ Very simple multi-layer perceptron (MLP, also called FFN), used to predict the normalized center coordinates, height and width of a bounding box w.r.t. an image. """ def __init__(self, input_dim, hidden_dim, output_dim, num_layers): super().__init__() self.num_layers = num_layers h = [hidden_dim] * (num_layers - 1) self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])) def forward(self, x): for i, layer in enumerate(self.layers): x = nn.functional.relu(layer(x)) if i < self.num_layers - 1 else layer(x) return x @auto_docstring( custom_intro=""" YOLOS Model (consisting of a ViT encoder) with object detection heads on top, for tasks such as COCO detection. """ ) class YolosForObjectDetection(YolosPreTrainedModel): def __init__(self, config: YolosConfig): super().__init__(config) # YOLOS (ViT) encoder model self.vit = YolosModel(config, add_pooling_layer=False) # Object detection heads # We add one for the "no object" class self.class_labels_classifier = YolosMLPPredictionHead( input_dim=config.hidden_size, hidden_dim=config.hidden_size, output_dim=config.num_labels + 1, num_layers=3 ) self.bbox_predictor = YolosMLPPredictionHead( input_dim=config.hidden_size, hidden_dim=config.hidden_size, output_dim=4, num_layers=3 ) # Initialize weights and apply final processing self.post_init() # taken from https://github.com/facebookresearch/detr/blob/master/models/detr.py def _set_aux_loss(self, outputs_class, outputs_coord): return [{"logits": a, "pred_boxes": b} for a, b in zip(outputs_class[:-1], outputs_coord[:-1])] @can_return_tuple @auto_docstring def forward( self, pixel_values: torch.FloatTensor, labels: list[dict] | None = None, **kwargs: Unpack[TransformersKwargs], ) -> YolosObjectDetectionOutput: r""" 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)`. Examples: ```python >>> from transformers import AutoImageProcessor, AutoModelForObjectDetection >>> import torch >>> from PIL import Image >>> import httpx >>> from io import BytesIO >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg" >>> with httpx.stream("GET", url) as response: ... image = Image.open(BytesIO(response.read())) >>> image_processor = AutoImageProcessor.from_pretrained("hustvl/yolos-tiny") >>> model = AutoModelForObjectDetection.from_pretrained("hustvl/yolos-tiny") >>> inputs = image_processor(images=image, return_tensors="pt") >>> outputs = model(**inputs) >>> # convert outputs (bounding boxes and class logits) to Pascal VOC format (xmin, ymin, xmax, ymax) >>> 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 0.991 at location [46.48, 72.78, 178.98, 119.3] Detected remote with confidence 0.908 at location [336.48, 79.27, 368.23, 192.36] Detected cat with confidence 0.934 at location [337.18, 18.06, 638.14, 373.09] Detected cat with confidence 0.979 at location [10.93, 53.74, 313.41, 470.67] Detected remote with confidence 0.974 at location [41.63, 72.23, 178.09, 119.99] ```""" # First, sent images through YOLOS base model to obtain hidden states outputs: BaseModelOutputWithPooling = self.vit(pixel_values, **kwargs) sequence_output = outputs.last_hidden_state # Take the final hidden states of the detection tokens sequence_output = sequence_output[:, -self.config.num_detection_tokens :, :] # Class logits + predicted bounding boxes logits = self.class_labels_classifier(sequence_output) pred_boxes = self.bbox_predictor(sequence_output).sigmoid() loss, loss_dict, auxiliary_outputs = None, None, None if labels is not None: outputs_class, outputs_coord = None, None if self.config.auxiliary_loss: intermediate = outputs.hidden_states outputs_class = self.class_labels_classifier(intermediate) outputs_coord = self.bbox_predictor(intermediate).sigmoid() loss, loss_dict, auxiliary_outputs = self.loss_function( logits, labels, self.device, pred_boxes, self.config, outputs_class, outputs_coord ) return YolosObjectDetectionOutput( loss=loss, loss_dict=loss_dict, logits=logits, pred_boxes=pred_boxes, auxiliary_outputs=auxiliary_outputs, last_hidden_state=outputs.last_hidden_state, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) __all__ = ["YolosForObjectDetection", "YolosModel", "YolosPreTrainedModel"]