Open
Conversation
long8v
commented
Sep 23, 2022
| } | ||
|
|
||
|
|
||
| class DeformableDetrConfig(PretrainedConfig): |
Comment on lines
+105
to
+108
| encoder_n_points (`int`, *optional*, defaults to 4): | ||
| The number of sampled keys in each feature level for each attention head in the encoder. | ||
| decoder_n_points (`int`, *optional*, defaults to 4): | ||
| The number of sampled keys in each feature level for each attention head in the decoder. |
Owner
Author
There was a problem hiding this comment.
deformable attention에서 필요한 config
Comment on lines
+109
to
+113
| two_stage (`bool`, *optional*, defaults to `False`): | ||
| Whether to apply a two-stage deformable DETR, where the region proposals are also generated by a variant of | ||
| Deformable DETR, which are further fed into the decoder for iterative bounding box refinement. | ||
| two_stage_num_proposals (`int`, *optional*, defaults to 300): | ||
| The number of region proposals to be generated, in case `two_stage` is set to `True`. |
Owner
Author
There was a problem hiding this comment.
two-stage model 관련 config들
Comment on lines
+114
to
+116
| with_box_refine (`bool`, *optional*, defaults to `False`): | ||
| Whether to apply iterative bounding box refinement, where each decoder layer refines the bounding boxes | ||
| based on the predictions from the previous layer. |
Owner
Author
There was a problem hiding this comment.
with_box_refine : 이전 레이어에서 나온 bbox를 초기값으로 사용함
Comment on lines
+103
to
+104
| num_feature_levels (`int`, *optional*, defaults to 4): | ||
| The number of input feature levels. |
Comment on lines
+1718
to
+1742
| if self.config.two_stage: | ||
| object_query_embedding, output_proposals = self.gen_encoder_output_proposals( | ||
| encoder_outputs[0], ~mask_flatten, spatial_shapes | ||
| ) | ||
|
|
||
| # hack implementation for two-stage Deformable DETR | ||
| # apply a detection head to each pixel (A.4 in paper) | ||
| # linear projection for bounding box binary classification (i.e. foreground and background) | ||
| enc_outputs_class = self.decoder.class_embed[-1](object_query_embedding) | ||
| # 3-layer FFN to predict bounding boxes coordinates (bbox regression branch) | ||
| delta_bbox = self.decoder.bbox_embed[-1](object_query_embedding) | ||
| enc_outputs_coord_logits = delta_bbox + output_proposals | ||
|
|
||
| # only keep top scoring `config.two_stage_num_proposals` proposals | ||
| topk = self.config.two_stage_num_proposals | ||
| topk_proposals = torch.topk(enc_outputs_class[..., 0], topk, dim=1)[1] | ||
| topk_coords_logits = torch.gather( | ||
| enc_outputs_coord_logits, 1, topk_proposals.unsqueeze(-1).repeat(1, 1, 4) | ||
| ) | ||
|
|
||
| topk_coords_logits = topk_coords_logits.detach() | ||
| reference_points = topk_coords_logits.sigmoid() | ||
| init_reference_points = reference_points | ||
| pos_trans_out = self.pos_trans_norm(self.pos_trans(self.get_proposal_pos_embed(topk_coords_logits))) | ||
| query_embed, target = torch.split(pos_trans_out, num_channels, dim=2) |
Owner
Author
There was a problem hiding this comment.
two-stage는 일단 모든 픽셀에 대해 bbox들을 뽑고 top k의 bbox coordinate을 positional embedding을 query_embed로 줌
| """, | ||
| DEFORMABLE_DETR_START_DOCSTRING, | ||
| ) | ||
| class DeformableDetrForObjectDetection(DeformableDetrPreTrainedModel): |
|
|
||
| @add_start_docstrings_to_model_forward(DEFORMABLE_DETR_INPUTS_DOCSTRING) | ||
| @replace_return_docstrings(output_type=DeformableDetrObjectDetectionOutput, config_class=_CONFIG_FOR_DOC) | ||
| def forward( |
Comment on lines
+1920
to
+1932
| for level in range(hidden_states.shape[0]): | ||
| if level == 0: | ||
| reference = init_reference | ||
| else: | ||
| reference = inter_references[level - 1] | ||
| reference = inverse_sigmoid(reference) | ||
| outputs_class = self.class_embed[level](hidden_states[level]) | ||
| delta_bbox = self.bbox_embed[level](hidden_states[level]) | ||
| if reference.shape[-1] == 4: | ||
| outputs_coord_logits = delta_bbox + reference | ||
| elif reference.shape[-1] == 2: | ||
| delta_bbox[..., :2] += reference | ||
| outputs_coord_logits = delta_bbox |
Owner
Author
There was a problem hiding this comment.
전 결과 값의 역시그모이드 + dx의 sigmoid 별거 아니고 0~1 값으로 해주려고 하는 짓임
|
|
||
|
|
||
| # Copied from transformers.models.detr.modeling_detr.DetrHungarianMatcher | ||
| class DeformableDetrHungarianMatcher(nn.Module): |
long8v
commented
Sep 27, 2022
| # 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. | ||
| """Feature extractor class for Deformable DETR.""" |
Owner
Author
There was a problem hiding this comment.
원래 DETR이랑 다른건 다 같은데 post_process가 다르다고 함!
the postprocessing of Deformable DETR is actually different compared to regular DETR. Namely, a sigmoid activation function is used rather than softmax, and the no-object class is included, whereas DETR discards this class.
| return color | ||
|
|
||
|
|
||
| class DeformableDetrFeatureExtractor(FeatureExtractionMixin, ImageFeatureExtractionMixin): |
Owner
Author
There was a problem hiding this comment.
겸사겸사해서 FeatureExtractor를 보자! 이 클래스는 전처리하는 class임!
Comment on lines
+134
to
+135
| do_resize (`bool`, *optional*, defaults to `True`): | ||
| Whether to resize the input to a certain `size`. |
Owner
Author
There was a problem hiding this comment.
resize=True! 어떻게 resize해주려나
Comment on lines
+136
to
+140
| size (`int`, *optional*, defaults to 800): | ||
| Resize the input to the given size. Only has an effect if `do_resize` is set to `True`. If size is a | ||
| sequence like `(width, height)`, output size will be matched to this. If size is an int, smaller edge of | ||
| the image will be matched to this number. i.e, if `height > width`, then image will be rescaled to `(size * | ||
| height / width, size)`. |
Owner
Author
There was a problem hiding this comment.
width, height로 들어가면 거기에 맞추고, 하나만 들어가면 단축을 size에 맞게 맞춤
Comment on lines
+141
to
+143
| max_size (`int`, *optional*, defaults to 1333): | ||
| The largest size an image dimension can have (otherwise it's capped). Only has an effect if `do_resize` is | ||
| set to `True`. |
Owner
Author
There was a problem hiding this comment.
들어올 수 있는 최대(장축)의 image size.
| return image, target | ||
|
|
||
| # Copied from transformers.models.detr.feature_extraction_detr.DetrFeatureExtractor._resize | ||
| def _resize(self, image, size, target=None, max_size=None): |
Comment on lines
+327
to
+333
| def get_size_with_aspect_ratio(image_size, size, max_size=None): | ||
| w, h = image_size | ||
| if max_size is not None: | ||
| min_original_size = float(min((w, h))) | ||
| max_original_size = float(max((w, h))) | ||
| if max_original_size / min_original_size * size > max_size: | ||
| size = int(round(max_size * min_original_size / max_original_size)) |
Owner
Author
There was a problem hiding this comment.
max_size가 주어졌을 경우에 size 구함
Comment on lines
+347
to
+353
| def get_size(image_size, size, max_size=None): | ||
| if isinstance(size, (list, tuple)): | ||
| return size | ||
| else: | ||
| # size returned must be (w, h) since we use PIL to resize images | ||
| # so we revert the tuple | ||
| return get_size_with_aspect_ratio(image_size, size, max_size)[::-1] |
Owner
Author
There was a problem hiding this comment.
max_size만 주어지면 비율에 맞춰서 크기 만듦! 즉 비율은 안바뀜! -> 현재 학습 중인 DETR의 셋팅인듯?
|
|
||
| return encoded_inputs | ||
|
|
||
| def post_process(self, outputs, target_sizes): |
| if target_sizes.shape[1] != 2: | ||
| raise ValueError("Each element of target_sizes must contain the size (h, w) of each image of the batch") | ||
|
|
||
| prob = out_logits.sigmoid() |
long8v
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
원본 코드는 https://github.com/fundamentalvision/Deformable-DETR/blob/main/models/deformable_transformer.py 인데 어차피 transformers 쓸거니 transformers 코드 보자 https://github.com/huggingface/transformers/tree/main/src/transformers/models/deformable_detr