search

jzhzhang / 3DAwareNav

[CVPR 2023] We propose a framework for the challenging 3D-aware ObjectNav based on two straightforward sub-policies. The two sub-polices, namely corner-guided exploration policy and category-aware identification policy, simultaneously perform by utilizing online fused 3D points as observation.
MIT License
57 stars 3 forks source link

Questions about training frame rate and training time #4

Closed peakfly closed 11 months ago

peakfly commented 1 year ago

This algorithm can only run 2-3 frames when I run the "sh_train.sh".However,the paper says the algorithm can run 15 frames.It takes a long time to train.The code runs on run on a workstation with an Intel(R) Xeon(R) Gold 6226R CPU with 256GB RAM and 2 Nvidia RTX3090 GPUs with 24GB memory.The "sh_train.sh" is as follows:

export GLOG_minloglevel=2
export MAGNUM_LOG="quiet"

python main.py --auto_gpu_config 0  -n 8 --num_training_frames 1000000\
    --sem_gpu_id_list "3"  --policy_gpu_id "cuda:4"  --sim_gpu_id "4" \
    --split train  --backbone_2d "rednet"  \
    --task_config "tasks/challenge_objectnav2021.local.rgbd.yaml"  --dataset "mp3d" \
    --num_sem_categories 22 --deactivate_entropymap \
    --print_images 1  -d ./tmp  --exp_name exp_kl_goal1  --save_periodic 10000

image

jzhzhang commented 1 year ago

During training, there are lots of operations including rollout storage and backward finetuning.

In our paper, we report the reconstruction performance, which only measures the construction time. If you want more higher FPS (like close to real-time) in construction time, you can simply downsample the per-frame point cloud (It's doable because of significant overlap).

peakfly commented 11 months ago

您的邮件已收到！谢谢！--------------------------------------------This is an automatic reply, confirming that your e-mail was received.Thank you