Closed becauseofAI closed 3 years ago
You may modify the LT_Dataset
class here to use a custom dataset.
You may modify the
LT_Dataset
class here to use a custom dataset.
@jiawei-ren I mean what is the meaning of the pipeline that you described as follow:
python main.py --cfg ./config/CIFAR10_LT/softmax_imba200.yaml
Alternatively, you may download a pretrained model here and put it in the corresponding log folder.
python main.py --cfg ./config/CIFAR10_LT/balanced_softmax_imba200.yaml
After obataining the base model, train the classifier with the following command:
python main.py --cfg ./config/CIFAR10_LT/decouple_balanced_softmax_imba200.yaml
python main.py --cfg ./config/CIFAR10_LT/balms_imba200.yaml
Method 1: End-to-end Training step 1:
python main.py --cfg ./config/CIFAR10_LT/softmax_imba200.yaml
Alternatively, you may download a pretrained model here and put it in the corresponding log folder. step 2:
python main.py --cfg ./config/CIFAR10_LT/balanced_softmax_imba200.yaml
Method 2: Decoupled Training step 1:
python main.py --cfg ./config/CIFAR10_LT/softmax_imba200.yaml
Alternatively, you may download a pretrained model here and put it in the corresponding log folder.
step 2:
python main.py --cfg ./config/CIFAR10_LT/decouple_balanced_softmax_imba200.yaml
Method 3: Decoupled Training step 1:
python main.py --cfg ./config/CIFAR10_LT/softmax_imba200.yaml
Alternatively, you may download a pretrained model here and put it in the corresponding log folder.
step 2:
python main.py --cfg ./config/CIFAR10_LT/balms_imba200.yaml
If this is true, which method works better or is more recommended? If not, what is the correct one?
Does that mean there are three kinds of training method?
It is mostly true, except that there is no step 1 in Method 1. End-to-end training doesn't require a pre-trained backbone.
If this is true, which method works better or is more recommended?
Method 3 is the most recommended. You may refer to the results reported in the README or the paper.
Does that mean there are three kinds of training method?
It is mostly true, except that there is no step 1 in Method 1. End-to-end training doesn't require a pre-trained backbone.
If this is true, which method works better or is more recommended?
Method 3 is the most recommended. You may refer to the results reported in the README or the paper.
Got it. Thx u very much! @jiawei-ren
I will close the issue for now. Feel free to reopen it if you have any further questions.
Thx for your jobs.
Can you describe the training pipline for custom dataset clearly? @jiawei-ren