Trained model

[WeedsInML]

Hello.
This is very nice work.
Could you please provide your trained model?

[zerchen]

Hi,

Yes, I will upload my trained model soon. You could first give a try to use my preprocessed data (https://drive.google.com/drive/folders/1GoaA6vB6TwAAHmaobVo5GjRoCq2wT21R) to do the training.

Best,
Zerui

[EAST-J]

Hi,

Yes, I will upload my trained model soon. You could first give a try to use my preprocessed data (https://drive.google.com/drive/folders/1GoaA6vB6TwAAHmaobVo5GjRoCq2wT21R) to do the training.

Best, Zerui

Hi,
Have you uploaded the trained model?

[zerchen]

Hi,

Thanks for your reminder. I have put the trained model on ObMan on here. Please give a note if it does not work or not compatible with this codebase. The checkpoints trained on dexycb will be released with a new codebase, which is more user-friendly.

Best,
Zerui

[RNGrunshen]

Hi,

Thanks for your reminder. I have put the trained model on ObMan on here. Please give a note if it does not work or not compatible with this codebase. The checkpoints trained on dexycb will be released with a new codebase, which is more user-friendly.

Best, Zerui

hello how can I use the trained model to evaluate

[happyiminjay1]

Hi,

Thanks for your reminder. I have put the trained model on ObMan on here. Please give a note if it does not work or not compatible with this codebase. The checkpoints trained on dexycb will be released with a new codebase, which is more user-friendly.

Best, Zerui

Hello,

I think the specs.json file which is uploaded on the experiments/obman/*.json file is not compatible to the trained model.

Can you upload the right specs.json file?

Thanks.

[zerchen]

Hi,

Thanks for your interests. Actually, I have not touched this codebase for a while. Some configurations might have changed during iterations of different versions.
I suggest using the latest codebase, which provides a cleaned implementation of AlignSDF. To test the model, you could download the directory and place it under $gSDF_ROOT/outputs. Then, please follow the instructions written here to test and evaluate the model. Please note that the current checkpoint is only trained with 750 epochs, instead of 1600 epochs used in the original paper. More checkpoints will come later.

Best,
Zerui

[happyiminjay1]

Thank you for the clear explanation! I will try on the latest codebase.

Someone who needs the specs.json file can use the following text.

{
"Description": "3D hand reconstruction on the mini obman dataset.",
"DataSource": "data",
"ImageSource": "rgb",
"TrainSplit": "experiments/splits/obman_80k.json",
"Dataset": "obman",
"ModelType": "1encoder2decoder",
"LMDB": true,
"ImageSize": [256, 256],
"SdfScaleFactor": 7.018621123357809,
"LatentSize": 256,
"PointFeatSize": 6,
"EncodeStyle": "nerf",
"ScaleAug": false,
"PoseFeatSize": 15,
"SnapshotFrequency": 500,
"LogFrequency": 5,
"LogFrequencyStep": 10,
"NumEpochs": 1600,
"Backbone": "resnet18",
"Resume": "latest.pth",
"Freeze": "none",
"PixelAlign": false,
"AdditionalSnapshots": [
100,
500
],
"AdditionalLossStart": 1201,
"SamplesPerScene": 2000,
"ScenesPerBatch": 64,
"DataLoaderThreads": 5,
"ClampingDistance": 0.05,
"HandBranch": true,
"ObjectBranch": true,
"ObjectPoseBranch": true,
"ManoBranch": true,
"DepthBranch": false,
"Render": false,
"ClassifierBranch": false,
"ClassifierWeight": 0.005,
"PenetrationLoss": false,
"ContactLoss": false,
"IndependentObjScale": false,
"IgnorePointFromOtherMesh": false,
"HandSdfWeight":0.5,
"ObjSdfWeight":0.5,
"JointWeight":0.5,
"VertWeight":0,
"ShapeRegWeight":0.0000005,
"PoseRegWeight":0.00005,
"SegWeight":0.004,
"ObjCenterWeight":1,
"ObjCornerWeight":0,
"PenetrationLossWeight":15.0,
"ContactLossWeight":0.005,
"DisableAug":false,
"BackgroundAug":false,
"NetworkSpecs": {
"dims": [
512,
512,
512,
512
],
"dropout": [
0,
1,
2,
3
],
"dropout_prob": 0.2,
"norm_layers": [
0,
1,
2,
3
],
"latent_in": [
2
],
"num_class": 6,
"xyz_in_all": false,
"use_tanh": false,
"latent_dropout": false,
"weight_norm": true
},
"LearningRateSchedule": [
{
"Type": "Step",
"Initial": 0.0001,
"Interval": 600,
"Factor": 0.5
}
]
}