sungnyun/openssl-simcore

(CVPR 2023) Coreset Sampling from Open-Set for Fine-Grained Self-Supervised Learning

openssl-simcore's Issues

Training time cost

Hi, thanks for sharing this impressive work! I would like to know what is the specific GPU you used in experiments, like 3090 or v100? And how long it takes to finish the pre train process on X and OS+X?

Results problem on aircraft dataset

Hi, I have read your paper repeatedly and that is a good job. At present, I am trying to replicate your results. However, I have encountered some problems in the process of replicating your results, which are as follows:
Following your experimental settings, I first conducted ssl training for Aircraft(train) by using simclr method. Then, test its performance on X(Aircraft test) using run_selfup.sh. However, the test results on X(Aircraft) are only [20.55, 28.32, 90.25](Acc@1, Acc@5, Train Acc).
There is a clear difference between this result and the paper.
Could you possibly give some suggestions as to the cause of this gap?
The specific training parameters I use are as follows:

TAG="aircraft_vanilla"
DATA="aircraft"

CUDA_VISIBLE_DEVICES=0,1,2,3 python train_selfsup.py --tag $TAG \
    --no_sampling \
    --model resnet50 \
    --batch_size 512 \
    --precision \
    --dataset $DATA \
    --data_folder ./data \
    --method simclr \
    --epochs 5000 \
    --cosine \
    --learning_rate 1e-1 \
    --weight_decay 1e-4

About the result of directly pretrained on target fine-grained datasets.

Hi, I have read your paper and that is a great work. But I faced some problems when pretraining on fine-grained datasets. I chose BYOL with resnet50 and Cars datasets. I set batchsize=2048, epoch=1000, cosine lr schedule with lr=1e-1, and weight decay=1e-4. The loss has decayed to 0.1 but the linear evaluation (fixed backbone) result is far away from 50%. I am confused and hope to know what're suitable settings, or have I missed the experiments' details?

About the accuracy of fine-tuning on the target fine-grained datasetes

Hi, I have read your paper which is a great work. But I have a question, how is the performance of fine-tuning the entire ResNet50 (both encoder and classifier) on these target fine-grained datasetes with supervised learning? Do you have the result?

Dataset split in SSL and evaluation

Hello, thanks for the great work!

Just wanted to check something:

1. In the appendix it's said that "we incorporate the train and validation set of the original benchmarks [...] to enlarge the number of training samples".
2. In the paper itself, Table 2 shows the size of the training set.

So when doing SSL and then evaluating just on aircraft for example (the 46.56% value), you use all 10k samples in SSL, and in table 2 you report the accuracy on the training set after training the linear classifier? Thanks in advance!