search

Sha-Lab / FEAT

The code repository for "Few-Shot Learning via Embedding Adaptation with Set-to-Set Functions"
MIT License
418 stars 84 forks source link

hyper-parameters to reproduce protonet results on resnet-12? #43

Closed Tsingularity closed 3 years ago

Tsingularity commented 3 years ago

Hi thanks for the great work!

I went through the github issue and found someone asked the same question. However, seems like at different times, your response is also slightly different, sometimes lr=1e-3 and sometimes be 1e-4. So could you finalize the answer for the setting of init lr, schedule, gamma, max epoch and temperature? thanks

i have tried some combination of these hyper-parameters with the pre-trained model you provided. But for 5way 5shot i cannot reach 80. Also when i look at the curve of training acc/error and validation acc, i found they actually change very little through the whole training process and the best validation epoch is basically the first 5 epochs. Just wondering is this behaviour normal or similar to your end?

Thanks in advance!

Han-Jia commented 3 years ago

Please use the following command to get the results. In my experiments, loss value will be small at first with the help of the pre-trained weights, but the model with the best meta-val accuracy will be after 50 epochs.

python train_fsl.py --max_epoch 200 --model_class FEAT --backbone_class Res12 --dataset MiniImageNet --way 5 --eval_way 5 --shot 5 --eval_shot 5 --query 15 --eval_query 15 --balance 0.1 --temperature 64 --temperature2 32 --lr 0.0002 --lr_mul 10 --lr_scheduler step --step_size 40 --gamma 0.5 --gpu 0 --init_weights ./saves/initialization/miniimagenet/Res12-pre.pth --eval_interval 1 --use_euclidean

Tsingularity commented 3 years ago

thanks! let me take a try lol

JimZAI commented 3 years ago

thanks! let me take a try lol

Hi thanks for the great work!

I went through the github issue and found someone asked the same question. However, seems like at different times, your response is also slightly different, sometimes lr=1e-3 and sometimes be 1e-4. So could you finalize the answer for the setting of init lr, schedule, gamma, max epoch and temperature? thanks

i have tried some combination of these hyper-parameters with the pre-trained model you provided. But for 5way 5shot i cannot reach 80. Also when i look at the curve of training acc/error and validation acc, i found they actually change very little through the whole training process and the best validation epoch is basically the first 5 epochs. Just wondering is this behaviour normal or similar to your end?

Thanks in advance!

May I ask if the experimental phenomenon you mentioned above is related to ProtoNet method? I also I found this problem when i reproduced the ProtoNet. Or are you just talking about FEAT?

Tsingularity commented 3 years ago

thanks! let me take a try lol

Hi thanks for the great work! I went through the github issue and found someone asked the same question. However, seems like at different times, your response is also slightly different, sometimes lr=1e-3 and sometimes be 1e-4. So could you finalize the answer for the setting of init lr, schedule, gamma, max epoch and temperature? thanks i have tried some combination of these hyper-parameters with the pre-trained model you provided. But for 5way 5shot i cannot reach 80. Also when i look at the curve of training acc/error and validation acc, i found they actually change very little through the whole training process and the best validation epoch is basically the first 5 epochs. Just wondering is this behaviour normal or similar to your end? Thanks in advance!

May I ask if the experimental phenomenon you mentioned above is related to ProtoNet method? I also I found this problem when i reproduced the ProtoNet. Or are you just talking about FEAT?

I was talking about using their pre-trained weights to train protonet. but i also found similar phenomenon when i train the FEAT.

JimZAI commented 3 years ago

Thank you for your response!

Actually, my real concern is Issue #45. Please excuse my carelessness!

At 2020-11-17 04:59:28, "Luming Tang" notifications@github.com wrote:

thanks! let me take a try lol

Hi thanks for the great work! I went through the github issue and found someone asked the same question. However, seems like at different times, your response is also slightly different, sometimes lr=1e-3 and sometimes be 1e-4. So could you finalize the answer for the setting of init lr, schedule, gamma, max epoch and temperature? thanks i have tried some combination of these hyper-parameters with the pre-trained model you provided. But for 5way 5shot i cannot reach 80. Also when i look at the curve of training acc/error and validation acc, i found they actually change very little through the whole training process and the best validation epoch is basically the first 5 epochs. Just wondering is this behaviour normal or similar to your end? Thanks in advance!

May I ask if the experimental phenomenon you mentioned above is related to ProtoNet method? I also I found this problem when i reproduced the ProtoNet. Or are you just talking about FEAT?

I was talking about using their pre-trained weights to train protonet. but i also found similar phenomenon when i train the FEAT.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.