关于训练步骤的请教

[littletomatodonkey]

你好，感谢你的开源！我直接使用CAMO数据训练，得到的指标如下所示，和论文差了比较多，这个可能是什么原因呢？

metric1: 0.4876                                                                                                                                                                                                    
metric2: 0.6093
metric3: 0.2453
metric4: 0.2108

我的训练脚本如下。

python3 -m torch.distributed.launch \
--master_port=12000 --nnodes 1 --nproc_per_node 4 \
train.py \
--config configs/demo.yaml

[tianrun-chen]

Greetings! Could you kindly verify if the correct training protocol mentioned in the manuscript has been employed? For our experiment, we have adopted the approach of prior studies and utilized a COMBINED training set, comprising both the CAMO and COD-10K training sets. Our evaluation entails using both the CAMO and COD-10K test sets, taking into account the CAMO's limited training sample. We will also release the pre-trained weights for verification.

[zhengziqiang]

你好，感谢你的开源！我直接使用CAMO数据训练，得到的指标如下所示，和论文差了比较多，这个可能是什么原因呢？

metric1: 0.4876                                                                                                                                                                                                    
metric2: 0.6093
metric3: 0.2453
metric4: 0.2108

我的训练脚本如下。

python3 -m torch.distributed.launch \
--master_port=12000 --nnodes 1 --nproc_per_node 4 \
train.py \
--config configs/demo.yaml

I want to know your training resources. I used 4x3090 but cannot run the experiments.

[littletomatodonkey]

Greetings! Could you kindly verify if the correct training protocol mentioned in the manuscript has been employed? For our experiment, we have adopted the approach of prior studies and utilized a COMBINED training set, comprising both the CAMO and COD-10K training sets. Our evaluation entails using both the CAMO and COD-10K test sets, taking into account the CAMO's limited training sample. We will also release the pre-trained weights for verification.

好的，我仅仅使用了CAMO数据，训练集有1000张图片，这个可能和你的训练步骤没有对齐，多谢~期待开源ckpt用于评估

[littletomatodonkey]

你好，感谢你的开源！我直接使用CAMO数据训练，得到的指标如下所示，和论文差了比较多，这个可能是什么原因呢？

metric1: 0.4876                                                                                                                                                                                                    
metric2: 0.6093
metric3: 0.2453
metric4: 0.2108

我的训练脚本如下。

python3 -m torch.distributed.launch \
--master_port=12000 --nnodes 1 --nproc_per_node 4 \
train.py \
--config configs/demo.yaml

I want to know your training resources. I used 4x3090 but cannot run the experiments.

4xA100（80G），SAM对于显存要求比较高，你可以把bs修改为1

[zhengziqiang]

你好，感谢你的开源！我直接使用CAMO数据训练，得到的指标如下所示，和论文差了比较多，这个可能是什么原因呢？

metric1: 0.4876                                                                                                                                                                                                    
metric2: 0.6093
metric3: 0.2453
metric4: 0.2108

我的训练脚本如下。

python3 -m torch.distributed.launch \
--master_port=12000 --nnodes 1 --nproc_per_node 4 \
train.py \
--config configs/demo.yaml

I want to know your training resources. I used 4x3090 but cannot run the experiments.

4xA100（80G），SAM对于显存要求比较高，你可以把bs修改为1

Thanks for your sharing. I have tried to change the BS to 1, but still cannot work. Now I have changed to only finetune the mask decoder part.

[tianrun-chen]

We use 4xA100 (80G) for training. We have released the pretrained weight! Thanks for your interests!

[Jack-bo1220]

你好，感谢你的开源！我直接使用CAMO数据训练，得到的指标如下所示，和论文差了比较多，这个可能是什么原因呢？

metric1: 0.4876                                                                                                                                                                                                    
metric2: 0.6093
metric3: 0.2453
metric4: 0.2108

我的训练脚本如下。

python3 -m torch.distributed.launch \
--master_port=12000 --nnodes 1 --nproc_per_node 4 \
train.py \
--config configs/demo.yaml

I want to know your training resources. I used 4x3090 but cannot run the experiments.

4xA100（80G），SAM对于显存要求比较高，你可以把bs修改为1

Thanks for your sharing. I have tried to change the BS to 1, but still cannot work. Now I have changed to only finetune the mask decoder part.

Hi, How can I change to only finetune the mask decoder part? Thanks.