Directory structure of links about chainercv HOT 8 CLOSED

Note that I omitted __init__.py for brevity.

from chainercv.

Also note that links/model and links/loss are good for consistency with chainer.links.

from chainercv.

I think the proposed structure is good. But if we expect losses are reused in other models, it is better to give them names which is not depend on models. For example, ssd_loss.py -> multibox_loss.py (of course, this name should be discussed).

from chainercv.

Yeah, ssd_loss.py and faster_rcnn_loss.py should be discussed.
On faster_rcnn_loss.py, it is very difficult to provide it in a general form because it assumes heavily the structure of Faster RCNN.
I think that if losses assume too much about the underlying models, it would be better to put it under model/MODEL_NAME.
Or, we can just leave it as faster_rcnn_loss.py indicating that it is model specific.
I think making it general is something that should be encouraged, but not necessary.

from chainercv.

I think that if losses assume too much about the underlying models, it would be better to put it under model/MODEL_NAME.

I like this solution. Only if we can make a loss general, we should put it into links/model/loss.

from chainercv.

I will implement the loss of SSD in links/model/ssd/multibox_loss.py first. If it seems enough general, I will move it into links/model/loss/<good_name>.py.

from chainercv.

I think some utils in Faster-RCNN are useful for SSD. After both models are fixed, I will suggest integration.

from chainercv.

The issue can be closed.

from chainercv.