From b64432404102e22b7d7f9a5be80c861ae5650e67 Mon Sep 17 00:00:00 2001 From: gitouni <30369762+gitouni@users.noreply.github.com> Date: Mon, 25 Jul 2022 18:43:50 +0800 Subject: [PATCH] Update README.md --- README.md | 22 ++-------------------- 1 file changed, 2 insertions(+), 20 deletions(-) diff --git a/README.md b/README.md index a87e7da..325c0ba 100644 --- a/README.md +++ b/README.md @@ -88,30 +88,12 @@ ln -s /PATH/TO/MyData/dataset data ### Train The following command is fit with a 12GB GPU. ```bash -python train.py --batch_size=2 --epoch=100 --inner_iter=5 --pcd_sample=20000 --name=cam2_muliter +python train.py --batch_size=8 --epoch=100 --inner_iter=1 --pcd_sample=4096 --name=cam2_oneiter ``` -
-Recommended training strategy - -A more successful way is to train the `one-iter` model first and then train the `multi-iter` one with the pretrained weigths of `one-iter` model. -```bash -python train.py --inner_iter=1 --name=cam2_oneiter --skip_frame=30 --pcd_sample=20000 -python train.py --inner_iter=5 --pretrained=./checkpoint/cam2_oneiter_best.pth --name=cam2_muliter --skip_frame=30 --pcd_sample=20000 -``` - -Similar to [mmsegmentation](https://github.com/open-mmlab/mmsegmentation/blob/master/docs/en/model_zoo.md), We use ResNetV1c instead of ResNetV1b. - -However, we don't use pretrained weights because it isn't effective enough. If you wan't to use pretrained resnet18, please see [https://download.openmmlab.com/pretrain/third_party/resnet18_v1c-b5776b93.pth](https://download.openmmlab.com/pretrain/third_party/resnet18_v1c-b5776b93.pth). - -Relevant training logs can be found in [log](./log) dir. - -Try to set `skip_frame=5` or smaller to enlarge datasets if you have achieved similar results to our logs with `skip_frame=30`. - -
### Test ```bash -python test.py --batch_size=1 --inner_iter=5 --pretrained=./checkpoint/cam2_muliter_best.pth --skip_frame=1 --pcd_sample=-1 +python test.py --inner_iter=1 --pretrained=./checkpoint/cam2_oneiter_best.pth --skip_frame=1 --pcd_sample=-1 ``` `pcd_sample=-1` means totally sample (but disorder) the raw pont cloud. However, you need to keep `batch_size=1` to avoid batch collect_fn error.