search

experiencor / keras-yolo2

Easy training on custom dataset. Various backends (MobileNet and SqueezeNet) supported. A YOLO demo to detect raccoon run entirely in brower is accessible at https://git.io/vF7vI (not on Windows).
MIT License
1.73k stars 787 forks source link

Minor question: why use b_batch? #407

Open abarajithan11 opened 5 years ago

abarajithan11 commented 5 years ago

The input and output to the network are ([image_batch, b_batch], y_batch). However, the b_batch is neither in loss function, the model, prediction nor anywhere else. For prediction we give a dummy array as b_batch. Then, what is purpose of batching it, sending it into the model for training..etc?

Why can't we simply have (image_batch, y_batch) as the input, output of the network?

hokim72 commented 5 years ago

b_batch is fed to true_boxes true_boxes is used for masks for confidence and class in loss function

mumuyanyan commented 5 years ago

b_batch is fed to true_boxes true_boxes is used for masks for confidence and class in loss function

In the training,there is already true_boxes.Sorry,i just don't get it.

speeding-motor commented 4 years ago

have you slove it now, I can not get it too