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
746 stars 149 forks source link

How to use a custom dataset with this model? #41

Open Vaibhav-ML opened 4 years ago

Vaibhav-ML commented 4 years ago

Hi, I have my own dataset. Folder Structure is as follows

Data - Folder 1(contains images and csv,json file) Folder 2 (same as above) . . Folder N(contains images and csv,json file)

How do I use my own dataset with this model?

yaoyao-liu commented 4 years ago

Hi @Vaibhav-ML,

Sorry for the late reply. If you need to run our code on your own dataset, you may revise the files in meta-transfer-learning/tensorflow/data_generator folder according to your dataset. If you have any further questions, feel free to add comments on this issue or email me.

Best, Yaoyao

Vaibhav-ML commented 4 years ago

Hi Liu, I have another query. My dataset contains images of traffic. So each image has multiple cars,trucks,cycles etc.

So, is it possible to detect multiple objects in a single image using your model?

If yes, then exactly what code changes would you advise to do for this to work?

Also, I have a json file which contains the bounding box info for the objects. How can I use that?

Regards, Vaibhav Sharma TU Dresden

On Mon, Oct 5, 2020 at 7:25 PM Yaoyao Liu notifications@github.com wrote:

Hi @Vaibhav-ML https://github.com/Vaibhav-ML,

Sorry for the late reply. If you need to run our code on your own dataset, you may revise the files in meta-transfer-learning/tensorflow/data_generator folder according to your dataset. If you have any further questions, feel free to add comments on this issue or email me.

Best, Yaoyao

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/yaoyao-liu/meta-transfer-learning/issues/41#issuecomment-703648559, or unsubscribe https://github.com/notifications/unsubscribe-auth/AN4IS5WMFDQ3AXHW2WSRHSLSJHF5VANCNFSM4SAGBGOA .

yaoyao-liu commented 4 years ago

Hi @Vaibhav-ML,

Our method cannot directly work on object detection tasks. As I am not working on object detection, I am not able to give you advice on that. Maybe you can follow the papers like this one.

Best. Yaoyao

Vaibhav-ML commented 4 years ago

Hi Liu, Thanks for your response. So you mean to say that this model can only be used for image classification but not object detection right?

Regards, Vaibhav Sharma TU Dresden

On Tue, 6 Oct 2020, 15:25 Yaoyao Liu, notifications@github.com wrote:

Hi @Vaibhav-ML https://github.com/Vaibhav-ML,

Our method cannot directly work on object detection tasks. As I am not working on object detection, I am not able to give you advice on that. Maybe you can follow the papers like this one https://openaccess.thecvf.com/content_ICCV_2019/papers/Kang_Few-Shot_Object_Detection_via_Feature_Reweighting_ICCV_2019_paper.pdf .

Best. Yaoyao

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/yaoyao-liu/meta-transfer-learning/issues/41#issuecomment-704161550, or unsubscribe https://github.com/notifications/unsubscribe-auth/AN4IS5WMBBKOHO4CMOL3O4DSJLSPRANCNFSM4SAGBGOA .

yaoyao-liu commented 4 years ago

Exactly. Most of the few-shot learning papers evaluate their method on classification, some also provide results for reinforcement learning, e.g., MAML. However, I seldom see papers titled "few-shot learning" evaluate their method on the object detection task. If they run experiments on object detection, they usually directly specify "few-shot object detection" in the title of the paper.