3D Pose Prediction

Project Link: https://machine-perception.ait.ethz.ch/project4/

How to reproduce the results

cd to the root folder and

Run:

python -u src/train.py --model [model] --n_epochs 2000 --lr 0.0005 --seed 42 --divide_lr_every 400

where [model] can be either:

ConvAttModel for the GCN with the convolutional 1d attention network
GCNAttModel for the GCN model with encoder/decoder structure,
GCNModel for the GCN model without an attention network.

Evaluate:

python src/evaluate.py --model_id [model id]

Development workspace and configuration

Create new models

To create a new model, create a new python script in the models folder. This script has to define a class which inherits from the BaseModel class found in base_model.py. You can then select to run train.py with this newly defined model using the tag --model followed by the name of the class. If no model arguments are given, the DummyModel will be selected by default. The create_model function in the models module will go through all the classes in all the scripts in the models folder, so you could technically define two different model classes in one script if necessary.

Testing

Visualization of the predictions can be done using evaluation.py. If ran with only --model_id provided, it will predict the target sequence of the test data and display 10 randomly picked samples. If --eval_on_val is provided and its value is 1, it will evaluate the model on the validation set and display the prediction alongside the ground truth.

Formatting

The predictions on the test data are in a .csv.gz format which can be directly uploaded on the submission website. After training, this file will be automatically generated and put in the model folder in your experiment folder along with the saved model parameters and configuration.

How to run locally

Using conda

Create a virtual environment in conda
```
conda create --name MP python=3.7.4
```
Activate it
```
conda activate MP
```

Install pytorch and torchvision

conda install pytorch=1.6.0 torchvision=0.7.0

Then, install requirements. You might have to comment out torch and torchvision.
```
conda install --file requirements.txt
```

Add MP_DATA and MP_EXPERIMENTS environment variables. This is how to do it on windows:

conda env config vars set MP_DATA=..\..\project4_data
conda env config vars set MP_EXPERIMENTS=..\experiments

It should now run
```
python train.py --model DummyModel
```

How to run on Leonhard

Setting up the environment

Connect to the Leonhard host (with your terminal or with VS Code) with
```
ssh [ethzusername]@login.leonhard.ethz.ch
```
Clone the project (only for the first time)
```
git clone [ssh project link from Gitlab] 
```
Warning: it could be necessary to copy the ssh key from Leonhard to Gitlab to have access to the Repository
Install the Python GPU Module on Leonhard (see wiki) with:
```
module load python_gpu/3.7.4
```

Run

export MP_DATA="/cluster/project/infk/hilliges/lectures/mp21/project4"

and

export MP_EXPERIMENTS="$HOME/Appliedscience/experiments"

Add these commands to the bashrc file to automatically run them when starting up the cluster:

module load python_gpu/3.7.4
source $HOME/.local/bin/virtualenvwrapper.sh
workon "MP21"
export MP_DATA="/cluster/project/infk/hilliges/lectures/mp21/project4"
export MP_EXPERIMENTS="$HOME/Appliedscience/experiments"

Running the script on leonhard

Go to the folder where your train.py file is located. You can run the following command:

bsub -n 1 -W 4:00 -o [outputname] -J [jobname] -R "rusage[mem=8096, ngpus_excl_p=1]" python -u train.py --model [model_class_name]

You can add other CL arguments like --lr or --n_epochs. Careful: the output file of the job will be saved in the folder you submitted the job from. This means that it will end up in the local git repo on your Leonhard home if you run the above command. Please do not push anything that contains a job output file onto the github as this would make it messy very quickly. Either have some sort of prefix system and add all files that start with said suffix to .gitignore or you can submit the job from outside of your git repo :

bsub -n 1 -W 4:00 -o [outputname] -J [jobname] -R "rusage[mem=8096, ngpus_excl_p=1]" python -u mp_project/src/train.py --model [model_class_name]

geiszla/machine-perception-21S