Try using a network pretrained on ImageNet (e.g. VGG16)

[marco-c]

Try using a network pretrained on ImageNet (e.g. VGG16)

[Shashi456]

Although transfer learning might work in this , we'd have to finetune pretty deep into the network to make it work for this usecase . VGG16 was trained on 1000 classes of dogs and cats so finetuning just the top layers for this usecase might not work.

[Shashi456]

@marco-c can we make a list of architectures not tried yet?

[marco-c]

@Shashi456 anything that is not in network.py has not been tried yet.

[marco-c]

If you have an idea for a specific network, you can file a new issue and mention this one.

[Shashi456]

@marco-c has any networked pretrained on image-net(like vgg16 or inception) been run for this, the network.py although includes vgg i dont know if transfer learning has been tested?

[marco-c]

No, not yet.

[vikasmahato]

Hi @marco-c
I am interested in working on this issue.

[marco-c]

@vikasmahato go ahead. Choose a network you want to try, file an issue about it and link the issue to this issue.

[Shashi456]

@marco-c in the create network function in networks.py i see that we can load weights onto the network, have we tried that for vgg16 yet ?

[marco-c]

No, nobody has tried that yet. If you do, please file an issue and link it here.

[marco-c]

Closing in favor of #194.