training model with no bias very slow to converge?

[ndcuong91]

For more suitable model to deploy in FPGA hardware, i tried to training this model with no bias. But the model converged slowly (mAP~35% after 200k iter with GTX 1080). So i have some question:

• Can this model work well with no bias term?
• How can i speed up training process? i use base_lr=0.0005, weight_decay=0.00005, gamma=0.5 Here is my model with no bias