search

MedChaabane / DEFT

Joint detection and tracking model named DEFT, or ``Detection Embeddings for Tracking." Our approach relies on an appearance-based object matching network jointly-learned with an underlying object detection network. An LSTM is also added to capture motion constraints.
MIT License
269 stars 43 forks source link

Some problems when working on reproducing the results #23

Open junjunxia opened 2 years ago

junjunxia commented 2 years ago

Hi, thank you every much for releasing the code ! I’ve been working on reproducing the results on the nuScences dataset and met some problems.

  1. Hyper-parameters differs between code and paper. Which one should be applied ? code: lr_decay_step = 60 batch_size = 1 paper: lr_decay_step = 30,60,70 batch_size = 8
  2. What is the option '--pre_hm' used for ? Why is this layer used only in training but not in testing?
  3. The checkpoint file I saved during training is larger than pretrained model dowaloaded(264M vs. 102M). Where does the difference come from ?
  4. Training for 80 epochs cost more than a week on 8x3090 gpu machine. Is it as expected ?

Could you help me please ?

xjturjc commented 1 year ago

@junjunxia Hi, do you have solved your porblems. I meet the same problem: Training for 80 epochs cost more than a week on 8 gpu machine. Looking forward to your reply. Thanks.