packnet and depthformer training config

[chenjx1005]

There are only inference config for packnet and depthformer. Could u pls provide training configs for them, so we can reproduce the results in your paper?

Thank you.

[Houssem-25]

I have create the following config but I didn't validate it yet. @VitorGuizilini-TRI @RaresAmbrus is this config correct ?

wrapper:
    recipe: wrapper|default
    max_epochs: 1
arch:
    model:
        file: depth/DepthFormerModel
        warp_context: [ -1,1 ]
        match_context: [ -1 ]
        motion_masking: True
        matching_augmentation: False
        freeze_teacher_and_pose: 45
    networks:
        transformer:
            recipe: networks/transformer|depthformer
        mono_depth:
            recipe: networks/mono_depth_res_net|default
            depth_range: [0.1,100.0]
        multi_depth:
            recipe: networks/multi_depth_res_net|depthformer
        pose:
            recipe: networks/pose_net|default
    losses:
        reprojection:
            recipe: losses/reprojection|default
        smoothness:
            recipe: losses/smoothness|default
        supervision:
            recipe: losses/supervised_depth|l1
evaluation:
    depth:
        recipe: evaluation/depth|kitti_resize

optimizers:
    multi_depth:
        recipe: optimizers|adam_20_05
    mono_depth:
        recipe: optimizers|adam_20_05
    pose:
        recipe: optimizers|adam_20_05
    transformer:
        recipe: optimizers|adam_20_05
datasets:
    train:
        recipe: datasets/kitti_tiny|train_selfsup_mr
        labels: [pose]
        context: [ -1,1 ]
    validation:
        recipe: datasets/kitti_tiny|validation_mr
        labels: [ depth,pose ]
        context: [ -1,1 ]

save:
    recipe: save|depth_splitname

[VitorGuizilini-TRI]

Seems correct, but I'll upload a training config file for DepthFormer soon.
One minor comment: you don't need pose in the training split, since it is also learned.

[Houssem-25]

The model fails if there no pose in the batch as in #15

[VitorGuizilini-TRI]

That's not an issue with the config file, it's in the model, you'll have to comment out parts relating to stereo, so it doesn't look for multiple cameras and their relative pose.

[CydKnocking]

Hi @VitorGuizilini-TRI, it will be definitely helpful if the training config file for DepthFormer (on KITTI) is released. I tried @Houssem-25 config but failed to reproduce the result in paper.

Thanks for your great work!

[KOuldamer]

Thank you for this work, @VitorGuizilini-TRI could you upload the self-calibration inference config file?

[jonasitzmann]

Do you have any updates on the release of the depthformer training config @VitorGuizilini-TRI ? Thanks for sharing this repo!

[tcourat]

I am interested by the training config for DepthFormer and PackNet too.