search

aimagelab / mcmr

PyTorch code for 3DV 2021 paper: "Multi-Category Mesh Reconstruction From Image Collections"
MIT License
18 stars 2 forks source link

Question about inference time on CUB dataset #7

Closed AlbertoRemus closed 2 years ago

AlbertoRemus commented 2 years ago

Hello for each image of the CUB dataset the test script https://github.com/aimagelab/mcmr#cub takes around 10 seconds (even without saving the reconstrucred .obj) on my NVIDIA GTX 1050 + Intel® Core™ i7-7700HQ CPU @ 2.80GHz × 8 architecture compared to few milliseconds on the NVIDIA GTX 1080 Ti + Intel Core i7-7700K stated in your paper and I would like to ask if you have any insights about it

This is the list of parameters

"args": ["--dataset_name", "cub", 
         "--dataset_dir", "path2mcmr/mcmr/datasets/cub/UCMR_CUB_data/cub/",
         "--classes", "all",
         "--single_mean_shape",
         "--subdivide", "4",
         "--sdf_subdivide_steps", "351",
         "--use_learned_class",
         "--num_learned_shapes", "1",
         "--checkpoint_dir", "path2mcmr/mcmr/checkpoint/meanshape01",
         "--log_dir", "path2mcmr/mcmr/log",
         "--pretrained_weights", "path2mcmr/mcmr/checkpoint/meanshape01/net_latest.pth",
         "--cam_loss_wt", "2.0",
         "--cam_reg_wt", "0.1",
         "--mask_loss_wt", "100.0",
         "--deform_reg_wt", "0.005",
         "--laplacian_wt", "6.0",
         "--laplacian_delta_wt", "1.8",
         "--graph_laplacian_wt", "0.0",
         "--tex_percept_loss_wt", "0.8",
         "--tex_color_loss_wt", "0.03",
         "--tex_pixel_loss_wt", "0.005",
         "--save_dir", "path2mcmr/mcmr/output",
         "--save_results",
         "--qualitative_results",
         "--faster"]

Thanks in advance

AlbertoRemus commented 2 years ago

Curiosely the above huge inference time was achieved by running the script from Visual Studio Code

Testing today from standard terminal provided an inference time is a much better ~0.3 seconds that become ~0.2 seconds without saving results

python main.py --dataset_name cub --dataset_dir path2mcmr/mcmr/datasets/cub/UCMR_CUB_data/cub/ \
               --classes all \
               --single_mean_shape \
               --subdivide 4 \
               --sdf_subdivide_steps 351 \
               --use_learned_class \
               --num_learned_shapes 1 \
               --checkpoint_dir path2mcmr/mcmr/checkpoint/meanshape01 \
               --log_dir log \
               --pretrained_weights path2mcmr/mcmr/checkpoint/meanshape01/net_latest.pth \
               --cam_loss_wt 2.0 \
               --cam_reg_wt 0.1 \
               --mask_loss_wt 100.0 \
               --deform_reg_wt 0.005 \
               --laplacian_wt 6.0 \
               --laplacian_delta_wt 1.8 \
               --graph_laplacian_wt 0.0 \
               --tex_percept_loss_wt 0.8 \
               --tex_color_loss_wt 0.03 \
               --tex_pixel_loss_wt 0.005 \
               --save_dir path2mcmr/mcmr/output
               --save_results \
               --qualitative_results \
               --faster
stefanopini commented 2 years ago

That's strange, maybe the python scripts from VS Code are executed attaching a debugger and that slows down a lot the diff. renderer or the model was running entirely on cpu for some reason.