search

yaoyao-liu / meta-transfer-learning

TensorFlow and PyTorch implementation of "Meta-Transfer Learning for Few-Shot Learning" (CVPR2019)
https://lyy.mpi-inf.mpg.de/mtl/
MIT License
731 stars 147 forks source link

some questions about meta-training and meta-testing phase #25

Open Sword-keeper opened 4 years ago

Sword-keeper commented 4 years ago

It's really a nice code ! However,I still have some questions about the meta training phase. 1.In the main.py, I found that in the meta-training phase, "-num_batch" is 100. It means there are 100 meta-training tasks in the meta-training phase. It was chosen from 64 classes in the 'train' folder. Each task contains 5 classes .In the support set, it is 51 samples. In the query set, it contains 515 samples. Besides,in the meta-eval phase,it is chosen from 20 classes in the 'test' folder. Each task contains 5 classes .In the support set, it is 51 samples. In the query set, it contains 515 samples. Am I right? 2.I can't understand the meaning of "--max_epoch"(line 29). Is it just like normal DL? For example, a: epoch 1: task 1~100, epoch 2: task 101~200,...... epoch 100: task 1001~1100 b: epoch 1: task 1~100, epoch 2: task 1~100.............epoch 100: task 1~100 which one is right? Maybe they are all wrong.

Sword-keeper commented 4 years ago

oh it maybe something wrong about my issue. Let me comment is again.

It's really a nice code ! However,I still have some questions about the meta training phase. 1.In the main.py, I found that in the meta-training phase, "-num_batch" is 100. It means there are 100 meta-training tasks in the meta-training phase. It was chosen from 64 classes in the 'train' folder. Each task contains 5 classes .In the support set, it is 51 samples. In the query set, it contains 515 samples. Besides,in the meta-eval phase,it is chosen from 20 classes in the 'test' folder. Each task contains 5 classes .In the support set, it is 51 samples. In the query set, it contains 515 samples. Am I right? 2.I can't understand the meaning of "--max_epoch"(line 29). Is it just like normal DL? For example, a: epoch 1: task 1--100, epoch 2: task 101--200,...... epoch 100: task 1001--1100 b: epoch 1: task 1--100, epoch 2: task 1--100.............epoch 100: task 1--100 which one is right? Maybe they are all wrong.

yaoyao-liu commented 4 years ago

Hi @Sword-keeper,

Thanks for your interest in our work. The PyTorch implementation is built based on the open-source code for FEAT. So we follow its settings.

num_batch=100 means we randomly sampled 100 tasks from the meta-train set. Each task contains 5 support samples and 75 query samples for 1-shot, 5-class setting. max_epoch=100 means we optimize the meta model on these 100 tasks for 100 times.

If you have any further questions, feel free to add more comments.

Sword-keeper commented 4 years ago 
  oh there still something wrong with my second comment, "51 samples " means 5×1 samples ， “515samples” means 5×15 samples. So , I think we have same opinion about "num_batch".
  As for "num-batch=100", as far as i know, when number of the dataset's sample is low, if our epoch is too high, the model is very easy to overfitting.  As for this "5-way 1-shot" classification task, why we set the "max_epoch=100"?
yaoyao-liu commented 4 years ago

Hi @Sword-keeper,

As I have explained, this setting follows FEAT. I have tried to use 10000 tasks and meta-train the model for 1 epoch. The performance is similar to using 100 tasks for 100 epochs.

Besides, if data augmentation is applied, the overfitting problem won't be serious during meta training.