Closed Sylva-Lin closed 3 months ago
We use A100s for all experiments except the inference time benchmark. From your description, the error occurred in the evaluation as we set grid_size to 0.01 and no crop during the evaluation. Maybe you can also add a SphereCrop during the crop, the test process after training will provide a more precise result. (If the default test process is also OOM, you might need to enable a special crop strategy during our precise test, if this occurs, I will tell you how to handle it later.)
We use A100s for all experiments except the inference time benchmark. From your description, the error occurred in the evaluation as we set grid_size to 0.01 and no crop during the evaluation. Maybe you can also add a SphereCrop during the crop, the test process after training will provide a more precise result. (If the default test process is also OOM, you might need to enable a special crop strategy during our precise test, if this occurs, I will tell you how to handle it later.)
Thanks, I added Sphere Crop and it was able to train and evaluate normally.
We use A100s for all experiments except the inference time benchmark. From your description, the error occurred in the evaluation as we set grid_size to 0.01 and no crop during the evaluation. Maybe you can also add a SphereCrop during the crop, the test process after training will provide a more precise result. (If the default test process is also OOM, you might need to enable a special crop strategy during our precise test, if this occurs, I will tell you how to handle it later.)
Hi, as you said, I am now experiencing OOM again in the testing phase, how can I introduce the crop in the testing phase?
Set a SphereCrop in "data.test.test_cfg.crop" and set the mode of "ShpereCrop" is "all" (https://github.com/Pointcept/Pointcept/blob/main/pointcept/datasets/transform.py#L927)
Hi, thank you for your work. When running S3DIS, the GPU memory is insufficient, the training configuration is 4RTX3090, the batch_size is set to 4, and the maximum number of points is set to 25600. In the training stage, the GPU memory is sufficient, but in the validation stage, the GPU memory is insufficient because the number of input point clouds is a normal room. What GPU can satisfy? I used 43090 with 24GB of GPU for my experiment