Problem about training on a small trainset

facebookresearch / supervision-by-registration

Supervision-by-Registration: An Unsupervised Approach to Improve the Precision of Facial Landmark Detectors

Other

762 stars 165 forks source link

Problem about training on a small trainset #67

Open JINXvvv opened 4 years ago

JINXvvv commented 4 years ago

I have trained a model on 300W and it works properly and the training loss is about 60.But when I decrease the number of trainset(200 images),the training loss is about 130.I have never modify the code! PS: I decrease the number of trainset for some other use.

D-X-Y commented 4 years ago

Our base detector is CPM. 200 images might be insufficient to train it. Why not combine 300W with your new training set to train? :)