Reproduce the results question + linear warmup cosine annealing

[gassercariad]

Hello!

I believe that from other issues as well, and also from my own experience to train your code with new torch versions, the results are really worse. Given that nothing is changed, but these things in the data loader :-
(1) self.interp = Image.ANTIALIAS
(2) transforms.ColorJitter.get_params(
self.brightness, self.contrast, self.saturation, self.hue)
(3) [in trainer.py] next(self.val_iter()) instead of self.val_iter().next()

I really don't think that all of these things have real effect, but I am really curious about the actual reasons why the algorithm (ManyDepth as well) behave way worse in the improved torch versions? Do you have a suggestion or reason for that?

My second question about the effect of linear warmup cosine annealing, and whether you test its effectiviness? Is there a specific reason that you decided to use it? And also the motivation to split the two different optimizers for pose & depth, instead of using just one optimizer (like the predecessors self-supervised?)

I really look forward to your replies and insights, especially about the weird performance in the newer torch versions :)

[noahzn]

Hi @gassercariad thanks for your interest in our work. If the training result is very bad using a newer PyTorch version please check if the default interpolation mode has been changed. This may affect the training.
From our observation the scheduler with learning rate reset used in our training helped to improve the trainings. The training converges fast but sometimes it falls into local minima, and we hope that the reset of learning rate may mitigate this.

[gassercariad]

@noahzn Thanks for your answer! I tried training now without changing anything of your repo with the default env of manydepth [torch 1.7] and newer torch [1.13] I did not change anything [except that using .png for KITTI] but the results are way off, is there any suspect for that performance? I just pulled the repo again, and did not change anything, and the results are as follows:-
at epoch 49 :-
"See the documentation of nn.Upsample for details.".format(mode))
-> Evaluating
Mono evaluation - using median scaling
Scaling ratios | med: 33.294 | std: 0.081

abs_rel | sq_rel | rmse | rmse_log | a1 | a2 | a3 |
& 0.129 & 0.987 & 5.058 & 0.208 & 0.847 & 0.947 & 0.977 \

And as it seems that it still learns somehow..
from tensorboard :-

[noahzn]

@gassercariad Did you load the pre-trained weights (pre-trained on ImageNet)?

[gassercariad]

Thanks for the quick answer!
The default weights_init is "pretrained", so it loads the pre-trained weights of ImageNet, right?
This is the exact command I did, python train.py --data_path xx --model_name xx , so I used also batch_size 16 and epoch 50, as default

[gassercariad]

I followed up on this, so I think the "weights init" only initalize for the posenet, but I have to get the pretrained weights for Imagenet and load them manually, right?

[noahzn]

Please download ImageNet-pretrained weights and add this --mypretrain path/to/your/pretrained/weights