Add classification examples about chainercv HOT 9 CLOSED

I think it makes ChainerCV more convenient to provide examples that reproduce fb.resnet.torch (https://github.com/facebook/fb.resnet.torch).
This code also parses the directory tree to provide a labelled dataset.

from chainercv.

To make this more concrete, I added a classification link that I would like to add to ChainerCV.
#265

Please join this discussion @Hakuyume , @mitmul @rezoo

1. What do you think about adding classification links to ChainerCV?
2. What do you think about adding a ImageDataset to ChainerCV?

from chainercv.

@yuyu2172 I agree with both 1 and 2. Classification links and datasets in Chainer don't follow the convention of ChainerCV (CHW, RGB). We need our own version (Of course, if we can change the models in Chainer, we don't need ChainerCV version). I think the ChainerCV implementation of fb.resnet.torch is very convenient.

Add links. Possibly, it can be located at chainercv.links.model.classifcation.

This namespace is not consistent with other existing models. We don't have chainercv.links.model.detection or chainercv.links.model.semantic_segmentation.

from chainercv.

This namespace is not consistent with other existing models. We don't have
chainercv.links.model.detection or chainercv.links.model.semantic_segmentation.

OK.
How about chainercv.links.model.vgg, chainercv.links.model.resnet etc.
We can have different variants of VGGs under chainercv.links.model.vgg. For example, VGG16 and VGG19.
This seems better. Thanks for a suggestion.

from chainercv.

Of course, if we can change the models in Chainer, we don't need ChainerCV version

The change I would like to introduce (BGR->RGB and a lot of change in API) is too large. I think we can not hope to change the main Chainer code in the way I want.

from chainercv.

How about chainercv.links.model.vgg, chainercv.links.model.resnet etc.

Sounds good.

The change I would like to introduce (BGR->RGB and a lot of change in API) is too large. I think we can not hope to change the main Chainer code in the way I want.

I agree with you. Some of our modifications will not be accepted to Chainer. For example, we need our own VGG-16. I think it is OK.

from chainercv.

I agree with both, too.
As for 2., the Keras's flow_from_directory-like utility that enables to obtain image classification dataset by just parsing the directory tree, would be nice to be included as one of features of it. See here for the details: https://keras.io/preprocessing/image/

from chainercv.

As for 2., the Keras's flow_from_directory-like utility that enables to obtain image classification dataset by just parsing the directory tree, would be nice to have as one of features of it.

I agree with you. This will be really handy, and we should implement it.

from chainercv.

Done

from chainercv.