This is a new readme; please check the git commits to see what this used to contain. Also please refer to the original readme
I found it difficult to run the official repository, so this is my documentation to run the code in a relatively reliable manner. Ideally, the steps in this readme should be reproducable, though I must confess I have only been able to run this in my present xUbuntu 24 system; and not in linux mint 22.
The only "change" I believe I have made to the workflow is using miniforge(mamba) instead of conda.
- Mamba (Conda can be used, just change the commands appropriately)
- After you git clone the repo, change the name to
soma
- I had downloaded all the required files into a local directory
./downloads/
and I know the relative path from this directory to the miniforge folder. Make appropriate changes to theinit.sh
file. NOTE THAT YOU CANNOT ACTIVATE ENVIRONMENTS EASILY WITH A BASH SCRIPT, SO YOU HAVE TO COPY PASTE THESE COMMANDS INTO THE SHELL - Additionally, install tbb in the environment; this is not present in ubuntu24, so it will have to be installed via mamba (There is one more error which ill rememeber how to fix when I see it again)
- Once the environment is made, you need to modify the following files
- welp i need to write these as I do these again
Main cell 1: I cannot:
'trainer.fast_dev_run': True, # if true then only one iteration of training and validation is done.
I cannot make that true without it giving a version error; and I cannot find which version is supposed to work? (Because of this, i needed to use the default trained SOMA)
├── data
│ ├── V48_01_SOMA [1]
│ └── V48_02_SOMA [0] First Cell
├── support_files
│ ├── evaluation_mocaps
│ │ └── original
│ │ └── SOMA_unlabeled_mpc [2]
│ ├── marker_layouts
│ │ └── SOMA [3]
│ ├── smplx [4]
│ │ ├── amass_neutral
│ │ │ ├── ACCAD [5]
│ │ │ ├── CMU [6]
│ │ │ ├── HumanEva [7]
│ │ │ ├── PosePrior [8]
│ │ │ ├── TotalCapture [9]
│ │ │ └── Transitions [10]
│ └── ssm_head_marker_corr.npz [11]
├── training_experiments [0] Cell 1
│ └── V48_02_SOMA
│ └── OC_05_G_03_real_000_synt_100 [12]
│ ├── evaluations
│ │ ├── mosh_results_tracklet [0] Cell Mosh
│ │ │ └── SOMA_unlabeled_mpc
│ │ │ ├── ['null']
│ │ │ └── SOMA_unlabeled_mpc_smplx.json
│ │ └── soma_labeled_mocap_tracklet [0] Cell 2
│ │ └── SOMA_unlabeled_mpc
│ │ ├── ['null']
│ │ └── soma_subject1 [0] Copy from brother folder
│ └── V48_02_SOMA_OC_05_G_03_real_000_synt_100.yaml [0] Generated, dont change
└── wierd.sh [mine]
- 0 - Generated by code, followed by cell and other info
- 1 - V48_01_SOMA
- 2 - SOMA_unlabeled_mpc
- 3 - SOMA
- 4 - smplx and this
- 5 - ACCAD
- 6 - CMU
- 7 - HumanEva
- 8 - PosePrior
- 9 - TotalCapture
- 10 - Transitions
- 11 - ssm_head_marker_corr.npz
- 12 - OC_05_G_03_real_000_synt_100 (almost all)