Closed JudasDie closed 3 years ago
PatrickHua
commented
3 years ago What kind of data? cifar/imagenet put target object in the middle and only one instance may appear for contrastive learning. The data augmentation (resized random crop etc.) is tailored to this type of input.
JudasDie
commented
3 years ago What kind of data? cifar/imagenet put target object in the middle and only one instance may appear for contrastive learning. The data augmentation (resized random crop etc.) is tailored to this type of input.
Similar to coco format, yet I crop the target from the original image. The target is centrally present in the image, as in Imagenet.
matthiasware
commented
3 years ago Has anyone met the problem that the loss close to -1 at the beginning of training? BTW, the training data is not sourced from traditional classification data like Imagenet or cifar.
Yes! Especially for parameters that differ from the ones given in the paper! Checkout Fig.2 in the paper and track the channel std!
JudasDie
commented
3 years ago Has anyone met the problem that the loss close to -1 at the beginning of training? BTW, the training data is not sourced from traditional classification data like Imagenet or cifar.
Yes! Especially for parameters that differ from the ones given in the paper! Checkout Fig.2 in the paper and track the channel std!
Hi，do you use Imagenet or custom data?
matthiasware
commented
3 years ago Has anyone met the problem that the loss close to -1 at the beginning of training? BTW, the training data is not sourced from traditional classification data like Imagenet or cifar.
Yes! Especially for parameters that differ from the ones given in the paper! Checkout Fig.2 in the paper and track the channel std!
Hi，do you use Imagenet or custom data?
custom dataset! but i can get good results by choosing the right parameters!
JudasDie
commented
3 years ago Has anyone met the problem that the loss close to -1 at the beginning of training? BTW, the training data is not sourced from traditional classification data like Imagenet or cifar.
Yes! Especially for parameters that differ from the ones given in the paper! Checkout Fig.2 in the paper and track the channel std!
Hi，do you use Imagenet or custom data?
custom dataset! but i can get good results by choosing the right parameters!
Could you share some experience? e.g., how to choose the parameter? or which parameter affects most?
PatrickHua
commented
3 years ago Has anyone met the problem that the loss close to -1 at the beginning of training? BTW, the training data is not sourced from traditional classification data like Imagenet or cifar.
Yes! Especially for parameters that differ from the ones given in the paper!
Interesting! So SimSiam is very sensitive to certain parameters right? What hyperparameter did you change ?
Checkout Fig.2 in the paper and track the channel std!
What's with the channel std?
Has anyone met the problem that the loss close to -1 at the beginning of training? BTW, the training data is not sourced from traditional classification data like Imagenet or cifar.