Training on higher resolution on custom dataset about bts HOT 5 CLOSED

Due to the strided convolutions and the simple upconv layers in our network, the height and width of the input should be multiples of 32.

from bts.

Thanks, I'll try training by making image size to be multiple of 32.

from bts.

Hi @cogaplex-bts,

It works really well on higher resolution images. I also want to know how do I apply Transfer learning in order to train bts on my custom dataset.

I tried training by using the pretrained encoder using NYU pretrained weights, instead of using ImageNet pretrained weights (and training the decoder from scratch). I did the following changes:

In bts.py, to load pretrained NYU V2 weights of densenet121 encoder (here BtsModelOriginal is the original bts model required to load weights):

https://github.com/cogaplex-bts/bts/blob/e63ccc00a1d44cc417c745714e3cabe58559f054/pytorch/bts.py#L273-#L276

     if True:
             model = BtsModelOriginal(params)
             model = torch.nn.DataParallel(model)
             checkpoint = torch.load('/kharshit/DepthEstimation/bts/models/bts_nyu_v2_pytorch_densenet121/model')
             model.load_state_dict(checkpoint['model'])
             self.base_model = model.module.encoder.base_model
             self.feat_names = ['relu0', 'pool0', 'transition1', 'transition2', 'norm5']

And, everything the same. Should I do some other modifications as well (in the decoder)?

from bts.

@kHarshit hi! What dataset are you using?

from bts.

@oljike It's a custom dataset prepared manually.

from bts.