riaz.somc/transformers
1
0
Fork 0
mirror of https://github.com/huggingface/transformers.git synced 2025-07-03 12:50:06 +06:00
Code Issues Actions 28 Packages Projects Releases Wiki Activity

Fix incorrect bbox_embed initialization when decoder_bbox_embed_share=False in GroundingDINO (#38238)

Browse Source 
* A shallow copy in groundingdino
Fixes #37333

* Supprimer une ligne vide dans la classe GroundingDinoForObjectDetection

* Translate comments in the GroundingDinoForObjectDetection class from French to English
This commit is contained in:
islemyakoubi 2025-05-30 14:02:18 +01:00 committed by GitHub
parent d0fccbf7ef
commit 98568d1e25
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 16 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
src/transformers/models/grounding_dino/modeling_grounding_dino.py
View File

@ -2454,16 +2454,25 @@ class GroundingDinoForObjectDetection(GroundingDinoPreTrainedModel):
_class_embed = GroundingDinoContrastiveEmbedding(config)
if config.decoder_bbox_embed_share:
_bbox_embed = GroundingDinoMLPPredictionHead(
# a single shared instance
shared_head = GroundingDinoMLPPredictionHead(
input_dim=config.d_model, hidden_dim=config.d_model, output_dim=4, num_layers=3
)
self.bbox_embed = nn.ModuleList([_bbox_embed for _ in range(config.decoder_layers)])
self.bbox_embed = nn.ModuleList([shared_head] * config.decoder_layers)
else:
for _ in range(config.decoder_layers):
_bbox_embed = GroundingDinoMLPPredictionHead(
input_dim=config.d_model, hidden_dim=config.d_model, output_dim=4, num_layers=3
# each layer has its own head (implicit deep copy through a new instance)
self.bbox_embed = nn.ModuleList(
[
GroundingDinoMLPPredictionHead(
input_dim=config.d_model,
hidden_dim=config.d_model,
output_dim=4,
num_layers=3,
)
self.bbox_embed = nn.ModuleList([_bbox_embed for _ in range(config.decoder_layers)])
for _ in range(config.decoder_layers)
]
)
self.class_embed = nn.ModuleList([_class_embed for _ in range(config.decoder_layers)])
# hack for box-refinement
self.model.decoder.bbox_embed = self.bbox_embed